Sa2VA-1B/sam2.py

import os
from collections import OrderedDict
from tqdm import tqdm
import torch.distributed

from torch.nn.init import trunc_normal_

import copy

from typing import List, Any, Optional, Tuple, Type, Union

import numpy as np

import math
import warnings
from functools import partial

import torch
import torch.nn.functional as F
from torch import nn, Tensor

# a large negative value as a placeholder score for missing objects
NO_OBJ_SCORE = -1024.0

warnings.simplefilter(action="ignore", category=FutureWarning)
# OLD_GPU, USE_FLASH_ATTN, MATH_KERNEL_ON = get_sdpa_settings()
OLD_GPU, USE_FLASH_ATTN, MATH_KERNEL_ON = True, True, True

def load_checkpoint_with_prefix(filename, prefix=None, map_location='cpu', logger='current'):
    """Load partial pretrained model with specific prefix.

    Args:
        prefix (str): The prefix of sub-module.
        filename (str): Accept local filepath, URL, ``torchvision://xxx``,
            ``open-mmlab://xxx``. Please refer to ``docs/model_zoo.md`` for
            details.
        map_location (str | None): Same as :func:`torch.load`.
            Defaults to None.
        logger: logger

    Returns:
        dict or OrderedDict: The loaded checkpoint.
    """
    checkpoint = torch.load(filename, map_location=map_location)

    if 'state_dict' in checkpoint:
        state_dict = checkpoint['state_dict']
    elif 'model' in checkpoint:
        state_dict = checkpoint['model']
    else:
        state_dict = checkpoint
    if not prefix:
        return state_dict
    if not prefix.endswith('.'):
        prefix += '.'
    prefix_len = len(prefix)

    state_dict = {
        k[prefix_len:]: v
        for k, v in state_dict.items() if k.startswith(prefix)
    }

    assert state_dict, f'{prefix} is not in the pretrained model'
    return state_dict

def load_state_dict_to_model(model, state_dict,  logger='current'):
    missing_keys, unexpected_keys = model.load_state_dict(state_dict)
    if missing_keys:
        print(missing_keys)
        raise RuntimeError()
    if unexpected_keys:
        print(unexpected_keys)
        raise RuntimeError()
    print("Loaded checkpoint successfully")

class SAM2(nn.Module):
    def __init__(
            self,
            ckpt_path: str = None,
    ):
        super().__init__()

        image_encoder = self.build_image_encoder()
        memory_attention = self.build_memory_attention()
        memory_encoder = self.build_memory_encoder()
        sam2_model = SAM2VideoPredictor(
            image_encoder=image_encoder,
            memory_attention=memory_attention,
            memory_encoder=memory_encoder,
            num_maskmem = 7,
            image_size = 1024,
            # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
            sigmoid_scale_for_mem_enc = 20.0,
            sigmoid_bias_for_mem_enc = -10.0,
            use_mask_input_as_output_without_sam = True,
            # Memory
            directly_add_no_mem_embed = True,
            # use high-resolution feature map in the SAM mask decoder
            use_high_res_features_in_sam = True,
            # output 3 masks on the first click on initial conditioning frames
            multimask_output_in_sam = True,
            # SAM heads
            iou_prediction_use_sigmoid = True,
            # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
            use_obj_ptrs_in_encoder = True,
            add_tpos_enc_to_obj_ptrs = False,
            only_obj_ptrs_in_the_past_for_eval = True,
            # object occlusion prediction
            pred_obj_scores = True,
            pred_obj_scores_mlp = True,
            fixed_no_obj_ptr = True,
            # multimask tracking settings
            multimask_output_for_tracking = True,
            use_multimask_token_for_obj_ptr = True,
            multimask_min_pt_num = 0,
            multimask_max_pt_num = 1,
            use_mlp_for_obj_ptr_proj = True,
            # Compilation flag
            compile_image_encoder = False,
            sam_mask_decoder_extra_args={
                'dynamic_multimask_via_stability':True,
                'dynamic_multimask_stability_delta': 0.05,
                'dynamic_multimask_stability_thresh': 0.98,
            }
        )
        if ckpt_path is not None:
            state_dict = load_checkpoint_with_prefix(ckpt_path)
            load_state_dict_to_model(sam2_model, state_dict)

        self.sam2_model = sam2_model

        self.hidden_dim = self.sam2_model.hidden_dim

        self.img_mean = (0.485, 0.456, 0.406)
        self.img_std = (0.229, 0.224, 0.225)

    def build_image_encoder(self):
        def build_trunk():
            embed_dim = 144
            num_heads = 2
            stages = [2, 6, 36, 4]
            global_att_blocks = [23, 33, 43]
            window_pos_embed_bkg_spatial_size = [7, 7]
            window_spec = [8, 4, 16, 8]
            ret = Hiera(
                embed_dim=embed_dim,
                num_heads=num_heads,
                stages=stages,
                global_att_blocks=global_att_blocks,
                window_pos_embed_bkg_spatial_size=window_pos_embed_bkg_spatial_size,
                window_spec=window_spec,
            )
            return ret
        def build_neck():
            def build_position_encoding():
                num_pos_feats = 256
                normalize = True
                scale = None
                temperature = 10000
                ret = PositionEmbeddingSine(
                    num_pos_feats=num_pos_feats,
                    normalize=normalize,
                    scale=scale,
                    temperature=temperature,
                )
                return ret
            d_model = 256
            backbone_channel_list = [1152, 576, 288, 144]
            fpn_top_down_levels = [2, 3]  # output level 0 and 1 directly use the backbone features
            fpn_interp_model = 'nearest'
            position_encoding = build_position_encoding()
            ret = FpnNeck(
                d_model=d_model,
                position_encoding=position_encoding,
                backbone_channel_list=backbone_channel_list,
                fpn_top_down_levels=fpn_top_down_levels,
                fpn_interp_model=fpn_interp_model,
            )
            return ret
        scalp = 1
        trunk = build_trunk()
        neck = build_neck()
        ret = ImageEncoder(scalp=scalp, trunk=trunk, neck=neck)
        return ret

    def build_memory_attention(self):
        def build_layer():
            def build_self_attention():
                rope_theta = 10000.0
                feat_sizes = [32, 32]
                embedding_dim = 256
                num_heads = 1
                downsample_rate = 1
                dropout = 0.1
                ret = RoPEAttention(
                    rope_theta=rope_theta,
                    feat_sizes=feat_sizes,
                    embedding_dim=embedding_dim,
                    num_heads=num_heads,
                    downsample_rate=downsample_rate,
                    dropout=dropout
                )
                return ret
            def build_cross_attention():
                rope_theta = 10000.0
                feat_sizes = [32, 32]
                rope_k_repeat = True
                embedding_dim = 256
                num_heads = 1
                downsample_rate = 1
                dropout = 0.1
                kv_in_dim = 64
                ret = RoPEAttention(
                    rope_theta=rope_theta,
                    feat_sizes=feat_sizes,
                    rope_k_repeat=rope_k_repeat,
                    embedding_dim=embedding_dim,
                    num_heads=num_heads,
                    downsample_rate=downsample_rate,
                    dropout=dropout,
                    kv_in_dim=kv_in_dim
                )
                return ret
            activation = 'relu'
            dim_feedforward = 2048
            dropout = 0.1
            pos_enc_at_attn = False
            d_model = 256
            pos_enc_at_cross_attn_keys = True
            pos_enc_at_cross_attn_queries = False
            self_attention = build_self_attention()
            cross_attention = build_cross_attention()
            ret = MemoryAttentionLayer(
                activation=activation,
                dim_feedforward=dim_feedforward,
                dropout=dropout,
                pos_enc_at_attn=pos_enc_at_attn,
                d_model=d_model,
                pos_enc_at_cross_attn_queries=pos_enc_at_cross_attn_queries,
                pos_enc_at_cross_attn_keys=pos_enc_at_cross_attn_keys,
                self_attention=self_attention,
                cross_attention=cross_attention,
            )
            return ret
        d_model = 256
        pos_enc_at_input = True
        num_layers = 4
        layer = build_layer()
        ret = MemoryAttention(
            d_model=d_model,
            pos_enc_at_input=pos_enc_at_input,
            num_layers=num_layers,
            layer=layer,
        )
        return ret

    def build_memory_encoder(self):
        def build_position_encoding():
            num_pos_feats = 64
            normalize = True
            scale = None
            temperature = 10000
            ret = PositionEmbeddingSine(
                num_pos_feats=num_pos_feats,
                normalize=normalize,
                scale=scale,
                temperature=temperature,
            )
            return ret

        def build_mask_downsampler():
            kernel_size = 3
            stride = 2
            padding = 1
            ret = MaskDownSampler(
                kernel_size=kernel_size,
                stride=stride,
                padding=padding,
            )
            return ret

        def build_fuser():
            def build_layer():
                dim = 256
                kernel_size = 7
                padding = 3
                layer_scale_init_value = 1e-6
                use_dwconv = True  # depth-wise convs
                ret = CXBlock(
                    dim=dim, kernel_size=kernel_size,
                    padding=padding, layer_scale_init_value=layer_scale_init_value,
                    use_dwconv=use_dwconv,
                )
                return ret

            num_layers = 2
            layer = build_layer()
            ret = Fuser(
                layer=layer,
                num_layers=num_layers
            )
            return ret

        out_dim = 64
        position_encoding = build_position_encoding()
        mask_downsampler = build_mask_downsampler()
        fuser = build_fuser()
        ret = MemoryEncoder(
            out_dim=out_dim,
            position_encoding=position_encoding,
            mask_downsampler=mask_downsampler,
            fuser=fuser,
        )
        return ret

    def inject_language_embd(self, inference_state, language_embd):
        num_frame = len(language_embd)
        num_obj = len(language_embd[0])
        mask_out = []
        for frame_idx in range(num_frame):
            frame_mask_out = []
            for obj_idx in range(num_obj):
                _language_embd = language_embd[frame_idx][obj_idx][None][None]
                _, _, out_mask_logits = self.sam2_model.add_language_embd(inference_state, frame_idx, obj_idx + 100, _language_embd)
                frame_mask_out.append(out_mask_logits)
            frame_mask_out = torch.cat(frame_mask_out, dim=1)
            mask_out.append(frame_mask_out)
        mask_out = torch.cat(mask_out, dim=0)
        return mask_out


    def language_embd_inference(self, inference_state, language_embd):
        num_frame = len(language_embd)
        num_obj = len(language_embd[0])
        mask_out = []
        with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
            for frame_idx in range(num_frame):
                frame_mask_out = []

                for obj_idx in range(num_obj):
                    _language_embd = language_embd[frame_idx][obj_idx][None][None]
                    _, _, out_mask_logits = self.sam2_model.add_language_embd(
                        inference_state,
                        frame_idx,
                        obj_idx + 100,
                        _language_embd,
                        inference=True,
                    )
                    frame_mask_out.append(out_mask_logits)
                frame_mask_out = torch.cat(frame_mask_out, dim=1)
                mask_out.append(frame_mask_out)


            mask_out = []
            for out_frame_idx, out_obj_ids, out_mask_logits in self.sam2_model.propagate_in_video(inference_state):
                mask_out.append(out_mask_logits)
            mask_out = torch.cat(mask_out, dim=0)
        return mask_out

    def get_sam2_embeddings(self, images):
        return self.sam2_model.init_state(images)

    def forward(self, batch):
        raise NotImplementedError

    def preprocess_image(self, image: torch.Tensor, dtype=torch.bfloat16) -> torch.Tensor:
        image = image / 255.

        img_mean = torch.tensor(self.img_mean, dtype=dtype, device=image.device)[:, None, None]
        img_std = torch.tensor(self.img_std, dtype=dtype, device=image.device)[:, None, None]
        image -= img_mean
        image /= img_std

        return image

class MemoryAttentionLayer(nn.Module):

    def __init__(
        self,
        activation: str,
        cross_attention: nn.Module,
        d_model: int,
        dim_feedforward: int,
        dropout: float,
        pos_enc_at_attn: bool,
        pos_enc_at_cross_attn_keys: bool,
        pos_enc_at_cross_attn_queries: bool,
        self_attention: nn.Module,
    ):
        super().__init__()
        self.d_model = d_model
        self.dim_feedforward = dim_feedforward
        self.dropout_value = dropout
        self.self_attn = self_attention
        self.cross_attn_image = cross_attention

        # Implementation of Feedforward model
        self.linear1 = nn.Linear(d_model, dim_feedforward)
        self.dropout = nn.Dropout(dropout)
        self.linear2 = nn.Linear(dim_feedforward, d_model)

        self.norm1 = nn.LayerNorm(d_model)
        self.norm2 = nn.LayerNorm(d_model)
        self.norm3 = nn.LayerNorm(d_model)
        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)
        self.dropout3 = nn.Dropout(dropout)

        self.activation_str = activation
        self.activation = get_activation_fn(activation)

        # Where to add pos enc
        self.pos_enc_at_attn = pos_enc_at_attn
        self.pos_enc_at_cross_attn_queries = pos_enc_at_cross_attn_queries
        self.pos_enc_at_cross_attn_keys = pos_enc_at_cross_attn_keys

    def _forward_sa(self, tgt, query_pos):
        # Self-Attention
        tgt2 = self.norm1(tgt)
        q = k = tgt2 + query_pos if self.pos_enc_at_attn else tgt2
        tgt2 = self.self_attn(q, k, v=tgt2)
        tgt = tgt + self.dropout1(tgt2)
        return tgt

    def _forward_ca(self, tgt, memory, query_pos, pos, num_k_exclude_rope=0):
        kwds = {}
        if num_k_exclude_rope > 0:
            assert isinstance(self.cross_attn_image, RoPEAttention)
            kwds = {"num_k_exclude_rope": num_k_exclude_rope}

        # Cross-Attention
        tgt2 = self.norm2(tgt)
        tgt2 = self.cross_attn_image(
            q=tgt2 + query_pos if self.pos_enc_at_cross_attn_queries else tgt2,
            k=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
            v=memory,
            **kwds,
        )
        tgt = tgt + self.dropout2(tgt2)
        return tgt

    def forward(
        self,
        tgt,
        memory,
        pos: Optional[Tensor] = None,
        query_pos: Optional[Tensor] = None,
        num_k_exclude_rope: int = 0,
    ) -> torch.Tensor:

        # Self-Attn, Cross-Attn
        tgt = self._forward_sa(tgt, query_pos)
        tgt = self._forward_ca(tgt, memory, query_pos, pos, num_k_exclude_rope)
        # MLP
        tgt2 = self.norm3(tgt)
        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
        tgt = tgt + self.dropout3(tgt2)
        return tgt


class MemoryAttention(nn.Module):
    def __init__(
        self,
        d_model: int,
        pos_enc_at_input: bool,
        layer: nn.Module,
        num_layers: int,
        batch_first: bool = True,  # Do layers expect batch first input?
    ):
        super().__init__()
        self.d_model = d_model
        self.layers = get_clones(layer, num_layers)
        self.num_layers = num_layers
        self.norm = nn.LayerNorm(d_model)
        self.pos_enc_at_input = pos_enc_at_input
        self.batch_first = batch_first

    def forward(
        self,
        curr: torch.Tensor,  # self-attention inputs
        memory: torch.Tensor,  # cross-attention inputs
        curr_pos: Optional[Tensor] = None,  # pos_enc for self-attention inputs
        memory_pos: Optional[Tensor] = None,  # pos_enc for cross-attention inputs
        num_obj_ptr_tokens: int = 0,  # number of object pointer *tokens*
    ):
        if isinstance(curr, list):
            assert isinstance(curr_pos, list)
            assert len(curr) == len(curr_pos) == 1
            curr, curr_pos = (
                curr[0],
                curr_pos[0],
            )

        assert (
            curr.shape[1] == memory.shape[1]
        ), "Batch size must be the same for curr and memory"

        output = curr
        if self.pos_enc_at_input and curr_pos is not None:
            output = output + 0.1 * curr_pos

        if self.batch_first:
            # Convert to batch first
            output = output.transpose(0, 1)
            curr_pos = curr_pos.transpose(0, 1)
            memory = memory.transpose(0, 1)
            memory_pos = memory_pos.transpose(0, 1)

        for layer in self.layers:
            kwds = {}
            if isinstance(layer.cross_attn_image, RoPEAttention):
                kwds = {"num_k_exclude_rope": num_obj_ptr_tokens}

            output = layer(
                tgt=output,
                memory=memory,
                pos=memory_pos,
                query_pos=curr_pos,
                **kwds,
            )
        normed_output = self.norm(output)

        if self.batch_first:
            # Convert back to seq first
            normed_output = normed_output.transpose(0, 1)
            curr_pos = curr_pos.transpose(0, 1)

        return normed_output

class MaskDownSampler(nn.Module):
    """
    Progressively downsample a mask by total_stride, each time by stride.
    Note that LayerNorm is applied per *token*, like in ViT.

    With each downsample (by a factor stride**2), channel capacity increases by the same factor.
    In the end, we linearly project to embed_dim channels.
    """

    def __init__(
        self,
        embed_dim=256,
        kernel_size=4,
        stride=4,
        padding=0,
        total_stride=16,
        activation=nn.GELU,
    ):
        super().__init__()
        num_layers = int(math.log2(total_stride) // math.log2(stride))
        assert stride**num_layers == total_stride
        self.encoder = nn.Sequential()
        mask_in_chans, mask_out_chans = 1, 1
        for _ in range(num_layers):
            mask_out_chans = mask_in_chans * (stride**2)
            self.encoder.append(
                nn.Conv2d(
                    mask_in_chans,
                    mask_out_chans,
                    kernel_size=kernel_size,
                    stride=stride,
                    padding=padding,
                )
            )
            self.encoder.append(LayerNorm2d(mask_out_chans))
            self.encoder.append(activation())
            mask_in_chans = mask_out_chans

        self.encoder.append(nn.Conv2d(mask_out_chans, embed_dim, kernel_size=1))

    def forward(self, x):
        return self.encoder(x)


# Lightly adapted from ConvNext (https://github.com/facebookresearch/ConvNeXt)
class CXBlock(nn.Module):
    r"""ConvNeXt Block. There are two equivalent implementations:
    (1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)
    (2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back
    We use (2) as we find it slightly faster in PyTorch

    Args:
        dim (int): Number of input channels.
        drop_path (float): Stochastic depth rate. Default: 0.0
        layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
    """

    def __init__(
        self,
        dim,
        kernel_size=7,
        padding=3,
        drop_path=0.0,
        layer_scale_init_value=1e-6,
        use_dwconv=True,
    ):
        super().__init__()
        self.dwconv = nn.Conv2d(
            dim,
            dim,
            kernel_size=kernel_size,
            padding=padding,
            groups=dim if use_dwconv else 1,
        )  # depthwise conv
        self.norm = LayerNorm2d(dim, eps=1e-6)
        self.pwconv1 = nn.Linear(
            dim, 4 * dim
        )  # pointwise/1x1 convs, implemented with linear layers
        self.act = nn.GELU()
        self.pwconv2 = nn.Linear(4 * dim, dim)
        # self.gamma = (
        self.g_weight = (
            nn.Parameter(layer_scale_init_value * torch.ones((dim)), requires_grad=True)
            if layer_scale_init_value > 0
            else None
        )
        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()

    def forward(self, x):
        input = x
        x = self.dwconv(x)
        x = self.norm(x)
        x = x.permute(0, 2, 3, 1)  # (N, C, H, W) -> (N, H, W, C)
        x = self.pwconv1(x)
        x = self.act(x)
        x = self.pwconv2(x)
        if self.g_weight is not None:
            x = self.g_weight * x
        x = x.permute(0, 3, 1, 2)  # (N, H, W, C) -> (N, C, H, W)

        x = input + self.drop_path(x)
        return x


class Fuser(nn.Module):
    def __init__(self, layer, num_layers, dim=None, input_projection=False):
        super().__init__()
        self.proj = nn.Identity()
        self.layers = get_clones(layer, num_layers)

        if input_projection:
            assert dim is not None
            self.proj = nn.Conv2d(dim, dim, kernel_size=1)

    def forward(self, x):
        # normally x: (N, C, H, W)
        x = self.proj(x)
        for layer in self.layers:
            x = layer(x)
        return x


class MemoryEncoder(nn.Module):
    def __init__(
        self,
        out_dim,
        mask_downsampler,
        fuser,
        position_encoding,
        in_dim=256,  # in_dim of pix_feats
    ):
        super().__init__()

        self.mask_downsampler = mask_downsampler

        self.pix_feat_proj = nn.Conv2d(in_dim, in_dim, kernel_size=1)
        self.fuser = fuser
        self.position_encoding = position_encoding
        self.out_proj = nn.Identity()
        if out_dim != in_dim:
            self.out_proj = nn.Conv2d(in_dim, out_dim, kernel_size=1)

    def forward(
        self,
        pix_feat: torch.Tensor,
        masks: torch.Tensor,
        skip_mask_sigmoid: bool = False,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        ## Process masks
        # sigmoid, so that less domain shift from gt masks which are bool
        if not skip_mask_sigmoid:
            masks = F.sigmoid(masks)
        masks = self.mask_downsampler(masks)

        ## Fuse pix_feats and downsampled masks
        # in case the visual features are on CPU, cast them to CUDA
        pix_feat = pix_feat.to(masks.device)

        x = self.pix_feat_proj(pix_feat)
        x = x + masks
        x = self.fuser(x)
        x = self.out_proj(x)

        pos = self.position_encoding(x).to(x.dtype)

        return {"vision_features": x, "vision_pos_enc": [pos]}


class ImageEncoder(nn.Module):
    def __init__(
        self,
        trunk: nn.Module,
        neck: nn.Module,
        scalp: int = 0,
    ):
        super().__init__()
        self.trunk = trunk
        self.neck = neck
        self.scalp = scalp
        assert (
            self.trunk.channel_list == self.neck.backbone_channel_list
        ), f"Channel dims of trunk and neck do not match. Trunk: {self.trunk.channel_list}, neck: {self.neck.backbone_channel_list}"

    def forward(self, sample: torch.Tensor):
        # Forward through backbone
        features, pos = self.neck(self.trunk(sample))
        if self.scalp > 0:
            # Discard the lowest resolution features
            features, pos = features[: -self.scalp], pos[: -self.scalp]

        src = features[-1]
        output = {
            "vision_features": src,
            "vision_pos_enc": pos,
            "backbone_fpn": features,
        }
        return output


class FpnNeck(nn.Module):
    """
    A modified variant of Feature Pyramid Network (FPN) neck
    (we remove output conv and also do bicubic interpolation similar to ViT
    pos embed interpolation)
    """

    def __init__(
        self,
        position_encoding: nn.Module,
        d_model: int,
        backbone_channel_list: List[int],
        kernel_size: int = 1,
        stride: int = 1,
        padding: int = 0,
        fpn_interp_model: str = "bilinear",
        fuse_type: str = "sum",
        fpn_top_down_levels: Optional[List[int]] = None,
    ):
        """Initialize the neck
        :param trunk: the backbone
        :param position_encoding: the positional encoding to use
        :param d_model: the dimension of the model
        :param neck_norm: the normalization to use
        """
        super().__init__()
        self.position_encoding = position_encoding
        self.convs = nn.ModuleList()
        self.backbone_channel_list = backbone_channel_list
        for dim in backbone_channel_list:
            current = nn.Sequential()
            current.add_module(
                "conv",
                nn.Conv2d(
                    in_channels=dim,
                    out_channels=d_model,
                    kernel_size=kernel_size,
                    stride=stride,
                    padding=padding,
                ),
            )

            self.convs.append(current)
        self.fpn_interp_model = fpn_interp_model
        assert fuse_type in ["sum", "avg"]
        self.fuse_type = fuse_type

        # levels to have top-down features in its outputs
        # e.g. if fpn_top_down_levels is [2, 3], then only outputs of level 2 and 3
        # have top-down propagation, while outputs of level 0 and level 1 have only
        # lateral features from the same backbone level.
        if fpn_top_down_levels is None:
            # default is to have top-down features on all levels
            fpn_top_down_levels = range(len(self.convs))
        self.fpn_top_down_levels = list(fpn_top_down_levels)

    def forward(self, xs: List[torch.Tensor]):

        out = [None] * len(self.convs)
        pos = [None] * len(self.convs)
        assert len(xs) == len(self.convs)
        # fpn forward pass
        # see https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/fpn.py
        prev_features = None
        # forward in top-down order (from low to high resolution)
        n = len(self.convs) - 1
        for i in range(n, -1, -1):
            x = xs[i]
            lateral_features = self.convs[n - i](x)
            if i in self.fpn_top_down_levels and prev_features is not None:
                top_down_features = F.interpolate(
                    prev_features.to(dtype=torch.float32),
                    scale_factor=2.0,
                    mode=self.fpn_interp_model,
                    align_corners=(
                        None if self.fpn_interp_model == "nearest" else False
                    ),
                    antialias=False,
                )
                prev_features = lateral_features + top_down_features
                if self.fuse_type == "avg":
                    prev_features /= 2
            else:
                prev_features = lateral_features
            x_out = prev_features
            out[i] = x_out
            pos[i] = self.position_encoding(x_out).to(x_out.dtype)

        return out, pos

def window_partition(x, window_size):
    """
    Partition into non-overlapping windows with padding if needed.
    Args:
        x (tensor): input tokens with [B, H, W, C].
        window_size (int): window size.
    Returns:
        windows: windows after partition with [B * num_windows, window_size, window_size, C].
        (Hp, Wp): padded height and width before partition
    """
    B, H, W, C = x.shape

    pad_h = (window_size - H % window_size) % window_size
    pad_w = (window_size - W % window_size) % window_size
    if pad_h > 0 or pad_w > 0:
        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
    Hp, Wp = H + pad_h, W + pad_w

    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
    windows = (
        x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
    )
    return windows, (Hp, Wp)


def window_unpartition(windows, window_size, pad_hw, hw):
    """
    Window unpartition into original sequences and removing padding.
    Args:
        x (tensor): input tokens with [B * num_windows, window_size, window_size, C].
        window_size (int): window size.
        pad_hw (Tuple): padded height and width (Hp, Wp).
        hw (Tuple): original height and width (H, W) before padding.
    Returns:
        x: unpartitioned sequences with [B, H, W, C].
    """
    Hp, Wp = pad_hw
    H, W = hw
    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
    x = windows.view(
        B, Hp // window_size, Wp // window_size, window_size, window_size, -1
    )
    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)

    if Hp > H or Wp > W:
        x = x[:, :H, :W, :].contiguous()
    return x


class PatchEmbed(nn.Module):
    """
    Image to Patch Embedding.
    """

    def __init__(
        self,
        kernel_size: Tuple[int, ...] = (7, 7),
        stride: Tuple[int, ...] = (4, 4),
        padding: Tuple[int, ...] = (3, 3),
        in_chans: int = 3,
        embed_dim: int = 768,
    ):
        """
        Args:
            kernel_size (Tuple): kernel size of the projection layer.
            stride (Tuple): stride of the projection layer.
            padding (Tuple): padding size of the projection layer.
            in_chans (int): Number of input image channels.
            embed_dim (int):  embed_dim (int): Patch embedding dimension.
        """
        super().__init__()
        self.proj = nn.Conv2d(
            in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.proj(x)
        # B C H W -> B H W C
        x = x.permute(0, 2, 3, 1)
        return x

def do_pool(x: torch.Tensor, pool: nn.Module, norm: nn.Module = None) -> torch.Tensor:
    if pool is None:
        return x
    # (B, H, W, C) -> (B, C, H, W)
    x = x.permute(0, 3, 1, 2)
    x = pool(x)
    # (B, C, H', W') -> (B, H', W', C)
    x = x.permute(0, 2, 3, 1)
    if norm:
        x = norm(x)

    return x


class MultiScaleAttention(nn.Module):
    def __init__(
        self,
        dim: int,
        dim_out: int,
        num_heads: int,
        q_pool: nn.Module = None,
    ):
        super().__init__()

        self.dim = dim
        self.dim_out = dim_out

        self.num_heads = num_heads
        head_dim = dim_out // num_heads
        self.scale = head_dim**-0.5

        self.q_pool = q_pool
        self.qkv = nn.Linear(dim, dim_out * 3)
        self.proj = nn.Linear(dim_out, dim_out)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        B, H, W, _ = x.shape
        # qkv with shape (B, H * W, 3, nHead, C)
        qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1)
        # q, k, v with shape (B, H * W, nheads, C)
        q, k, v = torch.unbind(qkv, 2)

        # Q pooling (for downsample at stage changes)
        if self.q_pool:
            q = do_pool(q.reshape(B, H, W, -1), self.q_pool)
            H, W = q.shape[1:3]  # downsampled shape
            q = q.reshape(B, H * W, self.num_heads, -1)

        # Torch's SDPA expects [B, nheads, H*W, C] so we transpose
        x = F.scaled_dot_product_attention(
            q.transpose(1, 2),
            k.transpose(1, 2),
            v.transpose(1, 2),
        )
        # Transpose back
        x = x.transpose(1, 2)
        x = x.reshape(B, H, W, -1)

        x = self.proj(x)

        return x


class MultiScaleBlock(nn.Module):
    def __init__(
        self,
        dim: int,
        dim_out: int,
        num_heads: int,
        mlp_ratio: float = 4.0,
        drop_path: float = 0.0,
        norm_layer: Union[nn.Module, str] = "LayerNorm",
        q_stride: Tuple[int, int] = None,
        act_layer: nn.Module = nn.GELU,
        window_size: int = 0,
    ):
        super().__init__()

        if isinstance(norm_layer, str):
            norm_layer = partial(getattr(nn, norm_layer), eps=1e-6)

        self.dim = dim
        self.dim_out = dim_out
        self.norm1 = norm_layer(dim)

        self.window_size = window_size

        self.pool, self.q_stride = None, q_stride
        if self.q_stride:
            self.pool = nn.MaxPool2d(
                kernel_size=q_stride, stride=q_stride, ceil_mode=False
            )

        self.attn = MultiScaleAttention(
            dim,
            dim_out,
            num_heads=num_heads,
            q_pool=self.pool,
        )
        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()

        self.norm2 = norm_layer(dim_out)
        self.mlp = MLP(
            dim_out,
            int(dim_out * mlp_ratio),
            dim_out,
            num_layers=2,
            activation=act_layer,
        )

        if dim != dim_out:
            self.proj = nn.Linear(dim, dim_out)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        shortcut = x  # B, H, W, C
        x = self.norm1(x)

        # Skip connection
        if self.dim != self.dim_out:
            shortcut = do_pool(self.proj(x), self.pool)

        # Window partition
        window_size = self.window_size
        if window_size > 0:
            H, W = x.shape[1], x.shape[2]
            x, pad_hw = window_partition(x, window_size)

        # Window Attention + Q Pooling (if stage change)
        x = self.attn(x)
        if self.q_stride:
            # Shapes have changed due to Q pooling
            window_size = self.window_size // self.q_stride[0]
            H, W = shortcut.shape[1:3]

            pad_h = (window_size - H % window_size) % window_size
            pad_w = (window_size - W % window_size) % window_size
            pad_hw = (H + pad_h, W + pad_w)

        # Reverse window partition
        if self.window_size > 0:
            x = window_unpartition(x, window_size, pad_hw, (H, W))

        x = shortcut + self.drop_path(x)
        # MLP
        x = x + self.drop_path(self.mlp(self.norm2(x)))
        return x


class Hiera(nn.Module):
    """
    Reference: https://arxiv.org/abs/2306.00989
    """

    def __init__(
        self,
        embed_dim: int = 96,  # initial embed dim
        num_heads: int = 1,  # initial number of heads
        drop_path_rate: float = 0.0,  # stochastic depth
        q_pool: int = 3,  # number of q_pool stages
        q_stride: Tuple[int, int] = (2, 2),  # downsample stride bet. stages
        stages: Tuple[int, ...] = (2, 3, 16, 3),  # blocks per stage
        dim_mul: float = 2.0,  # dim_mul factor at stage shift
        head_mul: float = 2.0,  # head_mul factor at stage shift
        window_pos_embed_bkg_spatial_size: Tuple[int, int] = (14, 14),
        # window size per stage, when not using global att.
        window_spec: Tuple[int, ...] = (
            8,
            4,
            14,
            7,
        ),
        # global attn in these blocks
        global_att_blocks: Tuple[int, ...] = (
            12,
            16,
            20,
        ),
        return_interm_layers=True,  # return feats from every stage
    ):
        super().__init__()

        assert len(stages) == len(window_spec)
        self.window_spec = window_spec

        depth = sum(stages)
        self.q_stride = q_stride
        self.stage_ends = [sum(stages[:i]) - 1 for i in range(1, len(stages) + 1)]
        assert 0 <= q_pool <= len(self.stage_ends[:-1])
        self.q_pool_blocks = [x + 1 for x in self.stage_ends[:-1]][:q_pool]
        self.return_interm_layers = return_interm_layers

        self.patch_embed = PatchEmbed(
            embed_dim=embed_dim,
        )
        # Which blocks have global att?
        self.global_att_blocks = global_att_blocks

        # Windowed positional embedding (https://arxiv.org/abs/2311.05613)
        self.window_pos_embed_bkg_spatial_size = window_pos_embed_bkg_spatial_size
        self.pos_embed = nn.Parameter(
            torch.zeros(1, embed_dim, *self.window_pos_embed_bkg_spatial_size)
        )
        self.pos_embed_window = nn.Parameter(
            torch.zeros(1, embed_dim, self.window_spec[0], self.window_spec[0])
        )

        dpr = [
            x.item() for x in torch.linspace(0, drop_path_rate, depth)
        ]  # stochastic depth decay rule

        cur_stage = 1
        self.blocks = nn.ModuleList()

        for i in range(depth):
            dim_out = embed_dim
            # lags by a block, so first block of
            # next stage uses an initial window size
            # of previous stage and final window size of current stage
            window_size = self.window_spec[cur_stage - 1]

            if self.global_att_blocks is not None:
                window_size = 0 if i in self.global_att_blocks else window_size

            if i - 1 in self.stage_ends:
                dim_out = int(embed_dim * dim_mul)
                num_heads = int(num_heads * head_mul)
                cur_stage += 1

            block = MultiScaleBlock(
                dim=embed_dim,
                dim_out=dim_out,
                num_heads=num_heads,
                drop_path=dpr[i],
                q_stride=self.q_stride if i in self.q_pool_blocks else None,
                window_size=window_size,
            )

            embed_dim = dim_out
            self.blocks.append(block)

        self.channel_list = (
            [self.blocks[i].dim_out for i in self.stage_ends[::-1]]
            if return_interm_layers
            else [self.blocks[-1].dim_out]
        )

    def _get_pos_embed(self, hw: Tuple[int, int]) -> torch.Tensor:
        h, w = hw
        window_embed = self.pos_embed_window
        pos_embed = F.interpolate(self.pos_embed, size=(h, w), mode="bicubic")
        pos_embed = pos_embed + window_embed.tile(
            [x // y for x, y in zip(pos_embed.shape, window_embed.shape)]
        )
        pos_embed = pos_embed.permute(0, 2, 3, 1)
        return pos_embed

    def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
        x = self.patch_embed(x)
        # x: (B, H, W, C)

        # Add pos embed
        x = x + self._get_pos_embed(x.shape[1:3])

        outputs = []
        for i, blk in enumerate(self.blocks):
            x = blk(x)
            if (i == self.stage_ends[-1]) or (
                i in self.stage_ends and self.return_interm_layers
            ):
                feats = x.permute(0, 3, 1, 2)
                outputs.append(feats)

        return outputs

class TwoWayTransformer(nn.Module):
    def __init__(
        self,
        depth: int,
        embedding_dim: int,
        num_heads: int,
        mlp_dim: int,
        activation: Type[nn.Module] = nn.ReLU,
        attention_downsample_rate: int = 2,
    ) -> None:
        """
        A transformer decoder that attends to an input image using
        queries whose positional embedding is supplied.

        Args:
          depth (int): number of layers in the transformer
          embedding_dim (int): the channel dimension for the input embeddings
          num_heads (int): the number of heads for multihead attention. Must
            divide embedding_dim
          mlp_dim (int): the channel dimension internal to the MLP block
          activation (nn.Module): the activation to use in the MLP block
        """
        super().__init__()
        self.depth = depth
        self.embedding_dim = embedding_dim
        self.num_heads = num_heads
        self.mlp_dim = mlp_dim
        self.layers = nn.ModuleList()

        for i in range(depth):
            self.layers.append(
                TwoWayAttentionBlock(
                    embedding_dim=embedding_dim,
                    num_heads=num_heads,
                    mlp_dim=mlp_dim,
                    activation=activation,
                    attention_downsample_rate=attention_downsample_rate,
                    skip_first_layer_pe=(i == 0),
                )
            )

        self.final_attn_token_to_image = Attention(
            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
        )
        self.norm_final_attn = nn.LayerNorm(embedding_dim)

    def forward(
        self,
        image_embedding: Tensor,
        image_pe: Tensor,
        point_embedding: Tensor,
    ) -> Tuple[Tensor, Tensor]:
        """
        Args:
          image_embedding (torch.Tensor): image to attend to. Should be shape
            B x embedding_dim x h x w for any h and w.
          image_pe (torch.Tensor): the positional encoding to add to the image. Must
            have the same shape as image_embedding.
          point_embedding (torch.Tensor): the embedding to add to the query points.
            Must have shape B x N_points x embedding_dim for any N_points.

        Returns:
          torch.Tensor: the processed point_embedding
          torch.Tensor: the processed image_embedding
        """
        # BxCxHxW -> BxHWxC == B x N_image_tokens x C
        bs, c, h, w = image_embedding.shape
        image_embedding = image_embedding.flatten(2).permute(0, 2, 1)
        image_pe = image_pe.flatten(2).permute(0, 2, 1)

        # Prepare queries
        queries = point_embedding
        keys = image_embedding

        # Apply transformer blocks and final layernorm
        for layer in self.layers:
            queries, keys = layer(
                queries=queries,
                keys=keys,
                query_pe=point_embedding,
                key_pe=image_pe,
            )

        # Apply the final attention layer from the points to the image
        q = queries + point_embedding
        k = keys + image_pe
        attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)
        queries = queries + attn_out
        queries = self.norm_final_attn(queries)

        return queries, keys


class TwoWayAttentionBlock(nn.Module):
    def __init__(
        self,
        embedding_dim: int,
        num_heads: int,
        mlp_dim: int = 2048,
        activation: Type[nn.Module] = nn.ReLU,
        attention_downsample_rate: int = 2,
        skip_first_layer_pe: bool = False,
    ) -> None:
        """
        A transformer block with four layers: (1) self-attention of sparse
        inputs, (2) cross attention of sparse inputs to dense inputs, (3) mlp
        block on sparse inputs, and (4) cross attention of dense inputs to sparse
        inputs.

        Arguments:
          embedding_dim (int): the channel dimension of the embeddings
          num_heads (int): the number of heads in the attention layers
          mlp_dim (int): the hidden dimension of the mlp block
          activation (nn.Module): the activation of the mlp block
          skip_first_layer_pe (bool): skip the PE on the first layer
        """
        super().__init__()
        self.self_attn = Attention(embedding_dim, num_heads)
        self.norm1 = nn.LayerNorm(embedding_dim)

        self.cross_attn_token_to_image = Attention(
            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
        )
        self.norm2 = nn.LayerNorm(embedding_dim)

        self.mlp = MLP(
            embedding_dim, mlp_dim, embedding_dim, num_layers=2, activation=activation
        )
        self.norm3 = nn.LayerNorm(embedding_dim)

        self.norm4 = nn.LayerNorm(embedding_dim)
        self.cross_attn_image_to_token = Attention(
            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
        )

        self.skip_first_layer_pe = skip_first_layer_pe

    def forward(
        self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor
    ) -> Tuple[Tensor, Tensor]:
        # Self attention block
        if self.skip_first_layer_pe:
            queries = self.self_attn(q=queries, k=queries, v=queries)
        else:
            q = queries + query_pe
            attn_out = self.self_attn(q=q, k=q, v=queries)
            queries = queries + attn_out
        queries = self.norm1(queries)

        # Cross attention block, tokens attending to image embedding
        q = queries + query_pe
        k = keys + key_pe
        attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)
        queries = queries + attn_out
        queries = self.norm2(queries)

        # MLP block
        mlp_out = self.mlp(queries)
        queries = queries + mlp_out
        queries = self.norm3(queries)

        # Cross attention block, image embedding attending to tokens
        q = queries + query_pe
        k = keys + key_pe
        attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)
        keys = keys + attn_out
        keys = self.norm4(keys)

        return queries, keys


class Attention(nn.Module):
    """
    An attention layer that allows for downscaling the size of the embedding
    after projection to queries, keys, and values.
    """

    def __init__(
        self,
        embedding_dim: int,
        num_heads: int,
        downsample_rate: int = 1,
        dropout: float = 0.0,
        kv_in_dim: int = None,
    ) -> None:
        super().__init__()
        self.embedding_dim = embedding_dim
        self.kv_in_dim = kv_in_dim if kv_in_dim is not None else embedding_dim
        self.internal_dim = embedding_dim // downsample_rate
        self.num_heads = num_heads
        assert (
            self.internal_dim % num_heads == 0
        ), "num_heads must divide embedding_dim."

        self.q_proj = nn.Linear(embedding_dim, self.internal_dim)
        self.k_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
        self.v_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
        self.out_proj = nn.Linear(self.internal_dim, embedding_dim)

        self.dropout_p = dropout

    def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:
        b, n, c = x.shape
        x = x.reshape(b, n, num_heads, c // num_heads)
        return x.transpose(1, 2)  # B x N_heads x N_tokens x C_per_head

    def _recombine_heads(self, x: Tensor) -> Tensor:
        b, n_heads, n_tokens, c_per_head = x.shape
        x = x.transpose(1, 2)
        return x.reshape(b, n_tokens, n_heads * c_per_head)  # B x N_tokens x C

    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:
        # Input projections
        q = self.q_proj(q)
        k = self.k_proj(k)
        v = self.v_proj(v)

        # Separate into heads
        q = self._separate_heads(q, self.num_heads)
        k = self._separate_heads(k, self.num_heads)
        v = self._separate_heads(v, self.num_heads)

        dropout_p = self.dropout_p if self.training else 0.0
        # Attention
        with torch.backends.cuda.sdp_kernel(
            enable_flash=USE_FLASH_ATTN,
            # if Flash attention kernel is off, then math kernel needs to be enabled
            enable_math=(OLD_GPU and dropout_p > 0.0) or MATH_KERNEL_ON,
            enable_mem_efficient=OLD_GPU,
        ):
            out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)

        out = self._recombine_heads(out)
        out = self.out_proj(out)

        return out


class RoPEAttention(Attention):
    """Attention with rotary position encoding."""

    def __init__(
        self,
        *args,
        rope_theta=10000.0,
        # whether to repeat q rope to match k length
        # this is needed for cross-attention to memories
        rope_k_repeat=False,
        feat_sizes=(32, 32),  # [w, h] for stride 16 feats at 512 resolution
        **kwargs,
    ):
        super().__init__(*args, **kwargs)

        self.compute_cis = partial(
            compute_axial_cis, dim=self.internal_dim // self.num_heads, theta=rope_theta
        )
        freqs_cis = self.compute_cis(end_x=feat_sizes[0], end_y=feat_sizes[1])
        self.freqs_cis = freqs_cis
        self.rope_k_repeat = rope_k_repeat

    def forward(
        self, q: Tensor, k: Tensor, v: Tensor, num_k_exclude_rope: int = 0
    ) -> Tensor:
        # Input projections
        q = self.q_proj(q)
        k = self.k_proj(k)
        v = self.v_proj(v)

        # Separate into heads
        q = self._separate_heads(q, self.num_heads)
        k = self._separate_heads(k, self.num_heads)
        v = self._separate_heads(v, self.num_heads)

        # Apply rotary position encoding
        w = h = math.sqrt(q.shape[-2])
        self.freqs_cis = self.freqs_cis.to(q.device)
        if self.freqs_cis.shape[0] != q.shape[-2]:
            self.freqs_cis = self.compute_cis(end_x=w, end_y=h).to(q.device)
        if q.shape[-2] != k.shape[-2]:
            assert self.rope_k_repeat

        num_k_rope = k.size(-2) - num_k_exclude_rope
        q, k[:, :, :num_k_rope] = apply_rotary_enc(
            q,
            k[:, :, :num_k_rope],
            freqs_cis=self.freqs_cis,
            repeat_freqs_k=self.rope_k_repeat,
        )

        dropout_p = self.dropout_p if self.training else 0.0
        # Attention
        with torch.backends.cuda.sdp_kernel(
            enable_flash=USE_FLASH_ATTN,
            # if Flash attention kernel is off, then math kernel needs to be enabled
            enable_math=(OLD_GPU and dropout_p > 0.0) or MATH_KERNEL_ON,
            enable_mem_efficient=OLD_GPU,
        ):
            out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)

        out = self._recombine_heads(out)
        out = self.out_proj(out)

        return out


class PromptEncoder(nn.Module):
    def __init__(
        self,
        embed_dim: int,
        image_embedding_size: Tuple[int, int],
        input_image_size: Tuple[int, int],
        mask_in_chans: int,
        activation: Type[nn.Module] = nn.GELU,
    ) -> None:
        """
        Encodes prompts for input to SAM's mask decoder.

        Arguments:
          embed_dim (int): The prompts' embedding dimension
          image_embedding_size (tuple(int, int)): The spatial size of the
            image embedding, as (H, W).
          input_image_size (int): The padded size of the image as input
            to the image encoder, as (H, W).
          mask_in_chans (int): The number of hidden channels used for
            encoding input masks.
          activation (nn.Module): The activation to use when encoding
            input masks.
        """
        super().__init__()
        self.embed_dim = embed_dim
        self.input_image_size = input_image_size
        self.image_embedding_size = image_embedding_size
        self.pe_layer = PositionEmbeddingRandom(embed_dim // 2)

        self.num_point_embeddings: int = 4  # pos/neg point + 2 box corners
        point_embeddings = [
            nn.Embedding(1, embed_dim) for i in range(self.num_point_embeddings)
        ]
        self.point_embeddings = nn.ModuleList(point_embeddings)
        self.not_a_point_embed = nn.Embedding(1, embed_dim)

        self.mask_input_size = (
            4 * image_embedding_size[0],
            4 * image_embedding_size[1],
        )
        self.mask_downscaling = nn.Sequential(
            nn.Conv2d(1, mask_in_chans // 4, kernel_size=2, stride=2),
            LayerNorm2d(mask_in_chans // 4),
            activation(),
            nn.Conv2d(mask_in_chans // 4, mask_in_chans, kernel_size=2, stride=2),
            LayerNorm2d(mask_in_chans),
            activation(),
            nn.Conv2d(mask_in_chans, embed_dim, kernel_size=1),
        )
        self.no_mask_embed = nn.Embedding(1, embed_dim)

    def get_dense_pe(self) -> torch.Tensor:
        """
        Returns the positional encoding used to encode point prompts,
        applied to a dense set of points the shape of the image encoding.

        Returns:
          torch.Tensor: Positional encoding with shape
            1x(embed_dim)x(embedding_h)x(embedding_w)
        """
        return self.pe_layer(self.image_embedding_size).unsqueeze(0)

    def _embed_points(
        self,
        points: torch.Tensor,
        labels: torch.Tensor,
        pad: bool,
    ) -> torch.Tensor:
        """Embeds point prompts."""
        points = points + 0.5  # Shift to center of pixel
        if pad:
            padding_point = torch.zeros((points.shape[0], 1, 2), device=points.device)
            padding_label = -torch.ones((labels.shape[0], 1), device=labels.device)
            points = torch.cat([points, padding_point], dim=1)
            labels = torch.cat([labels, padding_label], dim=1)
        point_embedding = self.pe_layer.forward_with_coords(
            points, self.input_image_size
        )
        point_embedding[labels == -1] = 0.0
        point_embedding[labels == -1] += self.not_a_point_embed.weight
        point_embedding[labels == 0] += self.point_embeddings[0].weight
        point_embedding[labels == 1] += self.point_embeddings[1].weight
        point_embedding[labels == 2] += self.point_embeddings[2].weight
        point_embedding[labels == 3] += self.point_embeddings[3].weight
        return point_embedding

    def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
        """Embeds box prompts."""
        boxes = boxes + 0.5  # Shift to center of pixel
        coords = boxes.reshape(-1, 2, 2)
        corner_embedding = self.pe_layer.forward_with_coords(
            coords, self.input_image_size
        )
        corner_embedding[:, 0, :] += self.point_embeddings[2].weight
        corner_embedding[:, 1, :] += self.point_embeddings[3].weight
        return corner_embedding

    def _embed_masks(self, masks: torch.Tensor) -> torch.Tensor:
        """Embeds mask inputs."""
        mask_embedding = self.mask_downscaling(masks)
        return mask_embedding

    def _get_batch_size(
        self,
        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
        boxes: Optional[torch.Tensor],
        masks: Optional[torch.Tensor],
    ) -> int:
        """
        Gets the batch size of the output given the batch size of the input prompts.
        """
        if points is not None:
            return points[0].shape[0]
        elif boxes is not None:
            return boxes.shape[0]
        elif masks is not None:
            return masks.shape[0]
        else:
            return 1

    def _get_device(self) -> torch.device:
        return self.point_embeddings[0].weight.device

    def forward(
        self,
        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
        boxes: Optional[torch.Tensor],
        masks: Optional[torch.Tensor],
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Embeds different types of prompts, returning both sparse and dense
        embeddings.

        Arguments:
          points (tuple(torch.Tensor, torch.Tensor) or none): point coordinates
            and labels to embed.
          boxes (torch.Tensor or none): boxes to embed
          masks (torch.Tensor or none): masks to embed

        Returns:
          torch.Tensor: sparse embeddings for the points and boxes, with shape
            BxNx(embed_dim), where N is determined by the number of input points
            and boxes.
          torch.Tensor: dense embeddings for the masks, in the shape
            Bx(embed_dim)x(embed_H)x(embed_W)
        """
        bs = self._get_batch_size(points, boxes, masks)
        sparse_embeddings = torch.empty(
            (bs, 0, self.embed_dim), device=self._get_device()
        )
        if points is not None:
            coords, labels = points
            point_embeddings = self._embed_points(coords, labels, pad=(boxes is None))
            sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=1)
        if boxes is not None:
            box_embeddings = self._embed_boxes(boxes)
            sparse_embeddings = torch.cat([sparse_embeddings, box_embeddings], dim=1)

        if masks is not None:
            dense_embeddings = self._embed_masks(masks)
        else:
            dense_embeddings = self.no_mask_embed.weight.reshape(1, -1, 1, 1).expand(
                bs, -1, self.image_embedding_size[0], self.image_embedding_size[1]
            )

        return sparse_embeddings, dense_embeddings

class PositionEmbeddingSine(nn.Module):
    """
    This is a more standard version of the position embedding, very similar to the one
    used by the Attention is all you need paper, generalized to work on images.
    """

    def __init__(
        self,
        num_pos_feats,
        temperature: int = 10000,
        normalize: bool = True,
        scale: Optional[float] = None,
    ):
        super().__init__()
        assert num_pos_feats % 2 == 0, "Expecting even model width"
        self.num_pos_feats = num_pos_feats // 2
        self.temperature = temperature
        self.normalize = normalize
        if scale is not None and normalize is False:
            raise ValueError("normalize should be True if scale is passed")
        if scale is None:
            scale = 2 * math.pi
        self.scale = scale

        self.cache = {}

    def _encode_xy(self, x, y):
        # The positions are expected to be normalized
        assert len(x) == len(y) and x.ndim == y.ndim == 1
        x_embed = x * self.scale
        y_embed = y * self.scale

        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)

        pos_x = x_embed[:, None] / dim_t
        pos_y = y_embed[:, None] / dim_t
        pos_x = torch.stack(
            (pos_x[:, 0::2].sin(), pos_x[:, 1::2].cos()), dim=2
        ).flatten(1)
        pos_y = torch.stack(
            (pos_y[:, 0::2].sin(), pos_y[:, 1::2].cos()), dim=2
        ).flatten(1)
        return pos_x, pos_y

    @torch.no_grad()
    def encode_boxes(self, x, y, w, h):
        pos_x, pos_y = self._encode_xy(x, y)
        pos = torch.cat((pos_y, pos_x, h[:, None], w[:, None]), dim=1)
        return pos

    encode = encode_boxes  # Backwards compatibility

    @torch.no_grad()
    def encode_points(self, x, y, labels):
        (bx, nx), (by, ny), (bl, nl) = x.shape, y.shape, labels.shape
        assert bx == by and nx == ny and bx == bl and nx == nl
        pos_x, pos_y = self._encode_xy(x.flatten(), y.flatten())
        pos_x, pos_y = pos_x.reshape(bx, nx, -1), pos_y.reshape(by, ny, -1)
        pos = torch.cat((pos_y, pos_x, labels[:, :, None]), dim=2)
        return pos

    @torch.no_grad()
    def forward(self, x: torch.Tensor):
        cache_key = (x.shape[-2], x.shape[-1])
        if cache_key in self.cache:
            return self.cache[cache_key][None].repeat(x.shape[0], 1, 1, 1)
        y_embed = (
            torch.arange(1, x.shape[-2] + 1, dtype=torch.float32, device=x.device)
            .view(1, -1, 1)
            .repeat(x.shape[0], 1, x.shape[-1])
        )
        x_embed = (
            torch.arange(1, x.shape[-1] + 1, dtype=torch.float32, device=x.device)
            .view(1, 1, -1)
            .repeat(x.shape[0], x.shape[-2], 1)
        )

        if self.normalize:
            eps = 1e-6
            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale

        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)

        pos_x = x_embed[:, :, :, None] / dim_t
        pos_y = y_embed[:, :, :, None] / dim_t
        pos_x = torch.stack(
            (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4
        ).flatten(3)
        pos_y = torch.stack(
            (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4
        ).flatten(3)
        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
        self.cache[cache_key] = pos[0]
        return pos


class PositionEmbeddingRandom(nn.Module):
    """
    Positional encoding using random spatial frequencies.
    """

    def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:
        super().__init__()
        if scale is None or scale <= 0.0:
            scale = 1.0
        self.register_buffer(
            "positional_encoding_gaussian_matrix",
            scale * torch.randn((2, num_pos_feats)),
        )
        self.first = True

    def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
        """Positionally encode points that are normalized to [0,1]."""
        # assuming coords are in [0, 1]^2 square and have d_1 x ... x d_n x 2 shape
        coords = 2 * coords - 1
        coords = coords.to(self.positional_encoding_gaussian_matrix.dtype)
        if self.first:
            self.positional_encoding_gaussian_matrix = self.positional_encoding_gaussian_matrix.to(coords.device)
            self.first = False
        coords = coords @ self.positional_encoding_gaussian_matrix
        coords = 2 * np.pi * coords
        # outputs d_1 x ... x d_n x C shape
        return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)

    def forward(self, size: Tuple[int, int]) -> torch.Tensor:
        """Generate positional encoding for a grid of the specified size."""
        h, w = size
        device: Any = self.positional_encoding_gaussian_matrix.device
        grid = torch.ones((h, w), device=device, dtype=torch.float32)
        y_embed = grid.cumsum(dim=0) - 0.5
        x_embed = grid.cumsum(dim=1) - 0.5
        y_embed = y_embed / h
        x_embed = x_embed / w

        pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))
        return pe.permute(2, 0, 1)  # C x H x W

    def forward_with_coords(
        self, coords_input: torch.Tensor, image_size: Tuple[int, int]
    ) -> torch.Tensor:
        """Positionally encode points that are not normalized to [0,1]."""
        coords = coords_input.clone()
        coords[:, :, 0] = coords[:, :, 0] / image_size[1]
        coords[:, :, 1] = coords[:, :, 1] / image_size[0]
        return self._pe_encoding(coords.to(torch.float))  # B x N x C


# Rotary Positional Encoding, adapted from:
# 1. https://github.com/meta-llama/codellama/blob/main/llama/model.py
# 2. https://github.com/naver-ai/rope-vit
# 3. https://github.com/lucidrains/rotary-embedding-torch


def init_t_xy(end_x: int, end_y: int):
    t = torch.arange(end_x * end_y, dtype=torch.float32)
    t_x = (t % end_x).float()
    t_y = torch.div(t, end_x, rounding_mode="floor").float()
    return t_x, t_y


def compute_axial_cis(dim: int, end_x: int, end_y: int, theta: float = 10000.0):
    freqs_x = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
    freqs_y = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))

    t_x, t_y = init_t_xy(end_x, end_y)
    freqs_x = torch.outer(t_x, freqs_x)
    freqs_y = torch.outer(t_y, freqs_y)
    freqs_cis_x = torch.polar(torch.ones_like(freqs_x), freqs_x)
    freqs_cis_y = torch.polar(torch.ones_like(freqs_y), freqs_y)
    return torch.cat([freqs_cis_x, freqs_cis_y], dim=-1)


def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor):
    ndim = x.ndim
    assert 0 <= 1 < ndim
    assert freqs_cis.shape == (x.shape[-2], x.shape[-1])
    shape = [d if i >= ndim - 2 else 1 for i, d in enumerate(x.shape)]
    return freqs_cis.view(*shape)


def apply_rotary_enc(
    xq: torch.Tensor,
    xk: torch.Tensor,
    freqs_cis: torch.Tensor,
    repeat_freqs_k: bool = False,
):
    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
    xk_ = (
        torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
        if xk.shape[-2] != 0
        else None
    )
    freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
    if xk_ is None:
        # no keys to rotate, due to dropout
        return xq_out.type_as(xq).to(xq.device), xk
    # repeat freqs along seq_len dim to match k seq_len
    if repeat_freqs_k:
        r = xk_.shape[-2] // xq_.shape[-2]
        freqs_cis = freqs_cis.repeat(*([1] * (freqs_cis.ndim - 2)), r, 1)
    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
    return xq_out.type_as(xq).to(xq.device), xk_out.type_as(xk).to(xk.device)


class MaskDecoder(nn.Module):
    def __init__(
        self,
        *,
        transformer_dim: int,
        transformer: nn.Module,
        num_multimask_outputs: int = 3,
        activation: Type[nn.Module] = nn.GELU,
        iou_head_depth: int = 3,
        iou_head_hidden_dim: int = 256,
        use_high_res_features: bool = False,
        iou_prediction_use_sigmoid=False,
        dynamic_multimask_via_stability=False,
        dynamic_multimask_stability_delta=0.05,
        dynamic_multimask_stability_thresh=0.98,
        pred_obj_scores: bool = False,
        pred_obj_scores_mlp: bool = False,
        use_multimask_token_for_obj_ptr: bool = False,
    ) -> None:
        """
        Predicts masks given an image and prompt embeddings, using a
        transformer architecture.

        Arguments:
          transformer_dim (int): the channel dimension of the transformer
          transformer (nn.Module): the transformer used to predict masks
          num_multimask_outputs (int): the number of masks to predict
            when disambiguating masks
          activation (nn.Module): the type of activation to use when
            upscaling masks
          iou_head_depth (int): the depth of the MLP used to predict
            mask quality
          iou_head_hidden_dim (int): the hidden dimension of the MLP
            used to predict mask quality
        """
        super().__init__()
        self.transformer_dim = transformer_dim
        self.transformer = transformer

        self.num_multimask_outputs = num_multimask_outputs

        self.iou_token = nn.Embedding(1, transformer_dim)
        self.num_mask_tokens = num_multimask_outputs + 1
        self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)

        self.pred_obj_scores = pred_obj_scores
        if self.pred_obj_scores:
            self.obj_score_token = nn.Embedding(1, transformer_dim)
        self.use_multimask_token_for_obj_ptr = use_multimask_token_for_obj_ptr

        self.output_upscaling = nn.Sequential(
            nn.ConvTranspose2d(
                transformer_dim, transformer_dim // 4, kernel_size=2, stride=2
            ),
            LayerNorm2d(transformer_dim // 4),
            activation(),
            nn.ConvTranspose2d(
                transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2
            ),
            activation(),
        )
        self.use_high_res_features = use_high_res_features
        if use_high_res_features:
            self.conv_s0 = nn.Conv2d(
                transformer_dim, transformer_dim // 8, kernel_size=1, stride=1
            )
            self.conv_s1 = nn.Conv2d(
                transformer_dim, transformer_dim // 4, kernel_size=1, stride=1
            )

        self.output_hypernetworks_mlps = nn.ModuleList(
            [
                MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3)
                for i in range(self.num_mask_tokens)
            ]
        )

        self.iou_prediction_head = MLP(
            transformer_dim,
            iou_head_hidden_dim,
            self.num_mask_tokens,
            iou_head_depth,
            sigmoid_output=iou_prediction_use_sigmoid,
        )
        if self.pred_obj_scores:
            self.pred_obj_score_head = nn.Linear(transformer_dim, 1)
            if pred_obj_scores_mlp:
                self.pred_obj_score_head = MLP(transformer_dim, transformer_dim, 1, 3)

        # When outputting a single mask, optionally we can dynamically fall back to the best
        # multimask output token if the single mask output token gives low stability scores.
        self.dynamic_multimask_via_stability = dynamic_multimask_via_stability
        self.dynamic_multimask_stability_delta = dynamic_multimask_stability_delta
        self.dynamic_multimask_stability_thresh = dynamic_multimask_stability_thresh

    def forward(
        self,
        image_embeddings: torch.Tensor,
        image_pe: torch.Tensor,
        sparse_prompt_embeddings: torch.Tensor,
        dense_prompt_embeddings: torch.Tensor,
        multimask_output: bool,
        repeat_image: bool,
        high_res_features: Optional[List[torch.Tensor]] = None,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Predict masks given image and prompt embeddings.

        Arguments:
          image_embeddings (torch.Tensor): the embeddings from the image encoder
          image_pe (torch.Tensor): positional encoding with the shape of image_embeddings
          sparse_prompt_embeddings (torch.Tensor): the embeddings of the points and boxes
          dense_prompt_embeddings (torch.Tensor): the embeddings of the mask inputs
          multimask_output (bool): Whether to return multiple masks or a single
            mask.

        Returns:
          torch.Tensor: batched predicted masks
          torch.Tensor: batched predictions of mask quality
          torch.Tensor: batched SAM token for mask output
        """
        masks, iou_pred, mask_tokens_out, object_score_logits = self.predict_masks(
            image_embeddings=image_embeddings,
            image_pe=image_pe,
            sparse_prompt_embeddings=sparse_prompt_embeddings,
            dense_prompt_embeddings=dense_prompt_embeddings,
            repeat_image=repeat_image,
            high_res_features=high_res_features,
        )

        # Select the correct mask or masks for output
        if multimask_output:
            masks = masks[:, 1:, :, :]
            iou_pred = iou_pred[:, 1:]
        elif self.dynamic_multimask_via_stability and not self.training:
            masks, iou_pred = self._dynamic_multimask_via_stability(masks, iou_pred)
        else:
            masks = masks[:, 0:1, :, :]
            iou_pred = iou_pred[:, 0:1]

        if multimask_output and self.use_multimask_token_for_obj_ptr:
            sam_tokens_out = mask_tokens_out[:, 1:]  # [b, 3, c] shape
        else:
            # Take the mask output token. Here we *always* use the token for single mask output.
            # At test time, even if we track after 1-click (and using multimask_output=True),
            # we still take the single mask token here. The rationale is that we always track
            # after multiple clicks during training, so the past tokens seen during training
            # are always the single mask token (and we'll let it be the object-memory token).
            sam_tokens_out = mask_tokens_out[:, 0:1]  # [b, 1, c] shape

        # Prepare output
        return masks, iou_pred, sam_tokens_out, object_score_logits

    def predict_masks(
        self,
        image_embeddings: torch.Tensor,
        image_pe: torch.Tensor,
        sparse_prompt_embeddings: torch.Tensor,
        dense_prompt_embeddings: torch.Tensor,
        repeat_image: bool,
        high_res_features: Optional[List[torch.Tensor]] = None,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Predicts masks. See 'forward' for more details."""
        # Concatenate output tokens
        s = 0
        if self.pred_obj_scores:
            output_tokens = torch.cat(
                [
                    self.obj_score_token.weight,
                    self.iou_token.weight,
                    self.mask_tokens.weight,
                ],
                dim=0,
            )
            s = 1
        else:
            output_tokens = torch.cat(
                [self.iou_token.weight, self.mask_tokens.weight], dim=0
            )
        output_tokens = output_tokens.unsqueeze(0).expand(
            sparse_prompt_embeddings.size(0), -1, -1
        )
        tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)

        # Expand per-image data in batch direction to be per-mask
        if repeat_image:
            src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
        else:
            assert image_embeddings.shape[0] == tokens.shape[0]
            src = image_embeddings
        src = src + dense_prompt_embeddings
        assert (
            image_pe.size(0) == 1
        ), "image_pe should have size 1 in batch dim (from `get_dense_pe()`)"
        pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
        b, c, h, w = src.shape

        # Run the transformer
        # print('src: ', src.dtype, 'pos_src:', pos_src.dtype, 'tokens:', tokens.dtype)
        _dtype = pos_src.dtype
        src = src.to(_dtype)
        tokens = tokens.to(_dtype)
        hs, src = self.transformer(src, pos_src, tokens)
        iou_token_out = hs[:, s, :]
        mask_tokens_out = hs[:, s + 1 : (s + 1 + self.num_mask_tokens), :]

        # Upscale mask embeddings and predict masks using the mask tokens
        src = src.transpose(1, 2).view(b, c, h, w)
        if not self.use_high_res_features:
            upscaled_embedding = self.output_upscaling(src)
        else:
            dc1, ln1, act1, dc2, act2 = self.output_upscaling
            feat_s0, feat_s1 = high_res_features
            upscaled_embedding = act1(ln1(dc1(src) + feat_s1))
            upscaled_embedding = act2(dc2(upscaled_embedding) + feat_s0)

        hyper_in_list: List[torch.Tensor] = []
        for i in range(self.num_mask_tokens):
            hyper_in_list.append(
                self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :])
            )
        hyper_in = torch.stack(hyper_in_list, dim=1)
        b, c, h, w = upscaled_embedding.shape
        masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)

        # Generate mask quality predictions
        iou_pred = self.iou_prediction_head(iou_token_out)
        if self.pred_obj_scores:
            assert s == 1
            object_score_logits = self.pred_obj_score_head(hs[:, 0, :])
        else:
            # Obj scores logits - default to 10.0, i.e. assuming the object is present, sigmoid(10)=1
            object_score_logits = 10.0 * iou_pred.new_ones(iou_pred.shape[0], 1)

        return masks, iou_pred, mask_tokens_out, object_score_logits

    def _get_stability_scores(self, mask_logits):
        """
        Compute stability scores of the mask logits based on the IoU between upper and
        lower thresholds, similar to https://github.com/fairinternal/onevision/pull/568.
        """
        mask_logits = mask_logits.flatten(-2)
        stability_delta = self.dynamic_multimask_stability_delta
        area_i = torch.sum(mask_logits > stability_delta, dim=-1).float()
        area_u = torch.sum(mask_logits > -stability_delta, dim=-1).float()
        stability_scores = torch.where(area_u > 0, area_i / area_u, 1.0)
        return stability_scores

    def _dynamic_multimask_via_stability(self, all_mask_logits, all_iou_scores):
        """
        When outputting a single mask, if the stability score from the current single-mask
        output (based on output token 0) falls below a threshold, we instead select from
        multi-mask outputs (based on output token 1~3) the mask with the highest predicted
        IoU score. This is intended to ensure a valid mask for both clicking and tracking.
        """
        # The best mask from multimask output tokens (1~3)
        multimask_logits = all_mask_logits[:, 1:, :, :]
        multimask_iou_scores = all_iou_scores[:, 1:]
        best_scores_inds = torch.argmax(multimask_iou_scores, dim=-1)
        batch_inds = torch.arange(
            multimask_iou_scores.size(0), device=all_iou_scores.device
        )
        best_multimask_logits = multimask_logits[batch_inds, best_scores_inds]
        best_multimask_logits = best_multimask_logits.unsqueeze(1)
        best_multimask_iou_scores = multimask_iou_scores[batch_inds, best_scores_inds]
        best_multimask_iou_scores = best_multimask_iou_scores.unsqueeze(1)

        # The mask from singlemask output token 0 and its stability score
        singlemask_logits = all_mask_logits[:, 0:1, :, :]
        singlemask_iou_scores = all_iou_scores[:, 0:1]
        stability_scores = self._get_stability_scores(singlemask_logits)
        is_stable = stability_scores >= self.dynamic_multimask_stability_thresh

        # Dynamically fall back to best multimask output upon low stability scores.
        mask_logits_out = torch.where(
            is_stable[..., None, None].expand_as(singlemask_logits),
            singlemask_logits,
            best_multimask_logits,
        )
        iou_scores_out = torch.where(
            is_stable.expand_as(singlemask_iou_scores),
            singlemask_iou_scores,
            best_multimask_iou_scores,
        )
        return mask_logits_out, iou_scores_out

def select_closest_cond_frames(frame_idx, cond_frame_outputs, max_cond_frame_num):
    """
    Select up to `max_cond_frame_num` conditioning frames from `cond_frame_outputs`
    that are temporally closest to the current frame at `frame_idx`. Here, we take
    - a) the closest conditioning frame before `frame_idx` (if any);
    - b) the closest conditioning frame after `frame_idx` (if any);
    - c) any other temporally closest conditioning frames until reaching a total
         of `max_cond_frame_num` conditioning frames.

    Outputs:
    - selected_outputs: selected items (keys & values) from `cond_frame_outputs`.
    - unselected_outputs: items (keys & values) not selected in `cond_frame_outputs`.
    """
    if max_cond_frame_num == -1 or len(cond_frame_outputs) <= max_cond_frame_num:
        selected_outputs = cond_frame_outputs
        unselected_outputs = {}
    else:
        assert max_cond_frame_num >= 2, "we should allow using 2+ conditioning frames"
        selected_outputs = {}

        # the closest conditioning frame before `frame_idx` (if any)
        idx_before = max((t for t in cond_frame_outputs if t < frame_idx), default=None)
        if idx_before is not None:
            selected_outputs[idx_before] = cond_frame_outputs[idx_before]

        # the closest conditioning frame after `frame_idx` (if any)
        idx_after = min((t for t in cond_frame_outputs if t >= frame_idx), default=None)
        if idx_after is not None:
            selected_outputs[idx_after] = cond_frame_outputs[idx_after]

        # add other temporally closest conditioning frames until reaching a total
        # of `max_cond_frame_num` conditioning frames.
        num_remain = max_cond_frame_num - len(selected_outputs)
        inds_remain = sorted(
            (t for t in cond_frame_outputs if t not in selected_outputs),
            key=lambda x: abs(x - frame_idx),
        )[:num_remain]
        selected_outputs.update((t, cond_frame_outputs[t]) for t in inds_remain)
        unselected_outputs = {
            t: v for t, v in cond_frame_outputs.items() if t not in selected_outputs
        }

    return selected_outputs, unselected_outputs


def get_1d_sine_pe(pos_inds, dim, temperature=10000):
    """
    Get 1D sine positional embedding as in the original Transformer paper.
    """
    pe_dim = dim // 2
    dim_t = torch.arange(pe_dim, dtype=torch.float32, device=pos_inds.device)
    dim_t = temperature ** (2 * (dim_t // 2) / pe_dim)

    pos_embed = pos_inds.unsqueeze(-1) / dim_t
    pos_embed = torch.cat([pos_embed.sin(), pos_embed.cos()], dim=-1)
    return pos_embed


def get_activation_fn(activation):
    """Return an activation function given a string"""
    if activation == "relu":
        return F.relu
    if activation == "gelu":
        return F.gelu
    if activation == "glu":
        return F.glu
    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")


def get_clones(module, N):
    return nn.ModuleList([copy.deepcopy(module) for i in range(N)])


class DropPath(nn.Module):
    # adapted from https://github.com/huggingface/pytorch-image-models/blob/main/timm/layers/drop.py
    def __init__(self, drop_prob=0.0, scale_by_keep=True):
        super(DropPath, self).__init__()
        self.drop_prob = drop_prob
        self.scale_by_keep = scale_by_keep

    def forward(self, x):
        if self.drop_prob == 0.0 or not self.training:
            return x
        keep_prob = 1 - self.drop_prob
        shape = (x.shape[0],) + (1,) * (x.ndim - 1)
        random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
        if keep_prob > 0.0 and self.scale_by_keep:
            random_tensor.div_(keep_prob)
        return x * random_tensor


# Lightly adapted from
# https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py # noqa
class MLP(nn.Module):
    def __init__(
        self,
        input_dim: int,
        hidden_dim: int,
        output_dim: int,
        num_layers: int,
        activation: nn.Module = nn.ReLU,
        sigmoid_output: bool = False,
    ) -> None:
        super().__init__()
        self.num_layers = num_layers
        h = [hidden_dim] * (num_layers - 1)
        self.layers = nn.ModuleList(
            nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])
        )
        self.sigmoid_output = sigmoid_output
        self.act = activation()

    def forward(self, x):
        for i, layer in enumerate(self.layers):
            x = self.act(layer(x)) if i < self.num_layers - 1 else layer(x)
        if self.sigmoid_output:
            x = F.sigmoid(x)
        return x


# From https://github.com/facebookresearch/detectron2/blob/main/detectron2/layers/batch_norm.py # noqa
# Itself from https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119  # noqa
class LayerNorm2d(nn.Module):
    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
        super().__init__()
        self.weight = nn.Parameter(torch.ones(num_channels))
        self.bias = nn.Parameter(torch.zeros(num_channels))
        self.eps = eps

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        u = x.mean(1, keepdim=True)
        s = (x - u).pow(2).mean(1, keepdim=True)
        x = (x - u) / torch.sqrt(s + self.eps)
        x = self.weight[:, None, None] * x + self.bias[:, None, None]
        return x

class SAM2Base_(torch.nn.Module):
    def __init__(
        self,
        image_encoder,
        memory_attention,
        memory_encoder,
        num_maskmem=7,  # default 1 input frame + 6 previous frames
        image_size=512,
        backbone_stride=16,  # stride of the image backbone output
        sigmoid_scale_for_mem_enc=1.0,  # scale factor for mask sigmoid prob
        sigmoid_bias_for_mem_enc=0.0,  # bias factor for mask sigmoid prob
        # During evaluation, whether to binarize the sigmoid mask logits on interacted frames with clicks
        binarize_mask_from_pts_for_mem_enc=False,
        use_mask_input_as_output_without_sam=False,  # on frames with mask input, whether to directly output the input mask without using a SAM prompt encoder + mask decoder
        # The maximum number of conditioning frames to participate in the memory attention (-1 means no limit; if there are more conditioning frames than this limit,
        # we only cross-attend to the temporally closest `max_cond_frames_in_attn` conditioning frames in the encoder when tracking each frame). This gives the model
        # a temporal locality when handling a large number of annotated frames (since closer frames should be more important) and also avoids GPU OOM.
        max_cond_frames_in_attn=-1,
        # on the first frame, whether to directly add the no-memory embedding to the image feature
        # (instead of using the transformer encoder)
        directly_add_no_mem_embed=False,
        # whether to use high-resolution feature maps in the SAM mask decoder
        use_high_res_features_in_sam=False,
        # whether to output multiple (3) masks for the first click on initial conditioning frames
        multimask_output_in_sam=False,
        # the minimum and maximum number of clicks to use multimask_output_in_sam (only relevant when `multimask_output_in_sam=True`;
        # default is 1 for both, meaning that only the first click gives multimask output; also note that a box counts as two points)
        multimask_min_pt_num=1,
        multimask_max_pt_num=1,
        # whether to also use multimask output for tracking (not just for the first click on initial conditioning frames; only relevant when `multimask_output_in_sam=True`)
        multimask_output_for_tracking=False,
        # Whether to use multimask tokens for obj ptr; Only relevant when both
        # use_obj_ptrs_in_encoder=True and multimask_output_for_tracking=True
        use_multimask_token_for_obj_ptr: bool = False,
        # whether to use sigmoid to restrict ious prediction to [0-1]
        iou_prediction_use_sigmoid=False,
        # The memory bank's temporal stride during evaluation (i.e. the `r` parameter in XMem and Cutie; XMem and Cutie use r=5).
        # For r>1, the (self.num_maskmem - 1) non-conditioning memory frames consist of
        # (self.num_maskmem - 2) nearest frames from every r-th frames, plus the last frame.
        memory_temporal_stride_for_eval=1,
        # if `add_all_frames_to_correct_as_cond` is True, we also append to the conditioning frame list any frame that receives a later correction click
        # if `add_all_frames_to_correct_as_cond` is False, we conditioning frame list to only use those initial conditioning frames
        add_all_frames_to_correct_as_cond=False,
        # whether to apply non-overlapping constraints on the object masks in the memory encoder during evaluation (to avoid/alleviate superposing masks)
        non_overlap_masks_for_mem_enc=False,
        # whether to cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
        use_obj_ptrs_in_encoder=False,
        # the maximum number of object pointers from other frames in encoder cross attention (only relevant when `use_obj_ptrs_in_encoder=True`)
        max_obj_ptrs_in_encoder=16,
        # whether to add temporal positional encoding to the object pointers in the encoder (only relevant when `use_obj_ptrs_in_encoder=True`)
        add_tpos_enc_to_obj_ptrs=True,
        # whether to add an extra linear projection layer for the temporal positional encoding in the object pointers to avoid potential interference
        # with spatial positional encoding (only relevant when both `use_obj_ptrs_in_encoder=True` and `add_tpos_enc_to_obj_ptrs=True`)
        proj_tpos_enc_in_obj_ptrs=False,
        # whether to only attend to object pointers in the past (before the current frame) in the encoder during evaluation
        # (only relevant when `use_obj_ptrs_in_encoder=True`; this might avoid pointer information too far in the future to distract the initial tracking)
        only_obj_ptrs_in_the_past_for_eval=False,
        # Whether to predict if there is an object in the frame
        pred_obj_scores: bool = False,
        # Whether to use an MLP to predict object scores
        pred_obj_scores_mlp: bool = False,
        # Only relevant if pred_obj_scores=True and use_obj_ptrs_in_encoder=True;
        # Whether to have a fixed no obj pointer when there is no object present
        # or to use it as an additive embedding with obj_ptr produced by decoder
        fixed_no_obj_ptr: bool = False,
        # Soft no object, i.e. mix in no_obj_ptr softly,
        # hope to make recovery easier if there is a mistake and mitigate accumulation of errors
        soft_no_obj_ptr: bool = False,
        use_mlp_for_obj_ptr_proj: bool = False,
        # extra arguments used to construct the SAM mask decoder; if not None, it should be a dict of kwargs to be passed into `MaskDecoder` class.
        sam_mask_decoder_extra_args=None,
        compile_image_encoder: bool = False,
    ):
        super().__init__()

        # Part 1: the image backbone
        self.image_encoder = image_encoder
        # Use level 0, 1, 2 for high-res setting, or just level 2 for the default setting
        self.use_high_res_features_in_sam = use_high_res_features_in_sam
        self.num_feature_levels = 3 if use_high_res_features_in_sam else 1
        self.use_obj_ptrs_in_encoder = use_obj_ptrs_in_encoder
        self.max_obj_ptrs_in_encoder = max_obj_ptrs_in_encoder
        if use_obj_ptrs_in_encoder:
            # A conv layer to downsample the mask prompt to stride 4 (the same stride as
            # low-res SAM mask logits) and to change its scales from 0~1 to SAM logit scale,
            # so that it can be fed into the SAM mask decoder to generate a pointer.
            self.mask_downsample = torch.nn.Conv2d(1, 1, kernel_size=4, stride=4)
        self.add_tpos_enc_to_obj_ptrs = add_tpos_enc_to_obj_ptrs
        if proj_tpos_enc_in_obj_ptrs:
            assert add_tpos_enc_to_obj_ptrs  # these options need to be used together
        self.proj_tpos_enc_in_obj_ptrs = proj_tpos_enc_in_obj_ptrs
        self.only_obj_ptrs_in_the_past_for_eval = only_obj_ptrs_in_the_past_for_eval

        # Part 2: memory attention to condition current frame's visual features
        # with memories (and obj ptrs) from past frames
        self.memory_attention = memory_attention
        self.hidden_dim = memory_attention.d_model

        # Part 3: memory encoder for the previous frame's outputs
        self.memory_encoder = memory_encoder
        self.mem_dim = self.hidden_dim
        if hasattr(self.memory_encoder, "out_proj") and hasattr(
            self.memory_encoder.out_proj, "weight"
        ):
            # if there is compression of memories along channel dim
            self.mem_dim = self.memory_encoder.out_proj.weight.shape[0]
        self.num_maskmem = num_maskmem  # Number of memories accessible
        # Temporal encoding of the memories
        self.maskmem_tpos_enc = torch.nn.Parameter(
            torch.zeros(num_maskmem, 1, 1, self.mem_dim)
        )
        trunc_normal_(self.maskmem_tpos_enc, std=0.02)
        # a single token to indicate no memory embedding from previous frames
        self.no_mem_embed = torch.nn.Parameter(torch.zeros(1, 1, self.hidden_dim))
        self.no_mem_pos_enc = torch.nn.Parameter(torch.zeros(1, 1, self.hidden_dim))
        trunc_normal_(self.no_mem_embed, std=0.02)
        trunc_normal_(self.no_mem_pos_enc, std=0.02)
        self.directly_add_no_mem_embed = directly_add_no_mem_embed
        # Apply sigmoid to the output raw mask logits (to turn them from
        # range (-inf, +inf) to range (0, 1)) before feeding them into the memory encoder
        self.sigmoid_scale_for_mem_enc = sigmoid_scale_for_mem_enc
        self.sigmoid_bias_for_mem_enc = sigmoid_bias_for_mem_enc
        self.binarize_mask_from_pts_for_mem_enc = binarize_mask_from_pts_for_mem_enc
        self.non_overlap_masks_for_mem_enc = non_overlap_masks_for_mem_enc
        self.memory_temporal_stride_for_eval = memory_temporal_stride_for_eval
        # On frames with mask input, whether to directly output the input mask without
        # using a SAM prompt encoder + mask decoder
        self.use_mask_input_as_output_without_sam = use_mask_input_as_output_without_sam
        self.multimask_output_in_sam = multimask_output_in_sam
        self.multimask_min_pt_num = multimask_min_pt_num
        self.multimask_max_pt_num = multimask_max_pt_num
        self.multimask_output_for_tracking = multimask_output_for_tracking
        self.use_multimask_token_for_obj_ptr = use_multimask_token_for_obj_ptr
        self.iou_prediction_use_sigmoid = iou_prediction_use_sigmoid

        # Part 4: SAM-style prompt encoder (for both mask and point inputs)
        # and SAM-style mask decoder for the final mask output
        self.image_size = image_size
        self.backbone_stride = backbone_stride
        self.sam_mask_decoder_extra_args = sam_mask_decoder_extra_args
        self.pred_obj_scores = pred_obj_scores
        self.pred_obj_scores_mlp = pred_obj_scores_mlp
        self.fixed_no_obj_ptr = fixed_no_obj_ptr
        self.soft_no_obj_ptr = soft_no_obj_ptr
        if self.fixed_no_obj_ptr:
            assert self.pred_obj_scores
            assert self.use_obj_ptrs_in_encoder
        if self.pred_obj_scores and self.use_obj_ptrs_in_encoder:
            self.no_obj_ptr = torch.nn.Parameter(torch.zeros(1, self.hidden_dim))
            trunc_normal_(self.no_obj_ptr, std=0.02)
        self.use_mlp_for_obj_ptr_proj = use_mlp_for_obj_ptr_proj

        self._build_sam_heads()
        self.add_all_frames_to_correct_as_cond = add_all_frames_to_correct_as_cond
        self.max_cond_frames_in_attn = max_cond_frames_in_attn

        # Model compilation
        if compile_image_encoder:
            # Compile the forward function (not the full module) to allow loading checkpoints.
            print(
                "Image encoder compilation is enabled. First forward pass will be slow."
            )
            self.image_encoder.forward = torch.compile(
                self.image_encoder.forward,
                mode="max-autotune",
                fullgraph=True,
                dynamic=False,
            )

    @property
    def device(self):
        return next(self.parameters()).device

    def forward(self, *args, **kwargs):
        raise NotImplementedError(
            "Please use the corresponding methods in SAM2VideoPredictor for inference."
            "See notebooks/video_predictor_example.ipynb for an example."
        )

    def _build_sam_heads(self):
        """Build SAM-style prompt encoder and mask decoder."""
        self.sam_prompt_embed_dim = self.hidden_dim
        self.sam_image_embedding_size = self.image_size // self.backbone_stride

        # build PromptEncoder and MaskDecoder from SAM
        # (their hyperparameters like `mask_in_chans=16` are from SAM code)
        self.sam_prompt_encoder = PromptEncoder(
            embed_dim=self.sam_prompt_embed_dim,
            image_embedding_size=(
                self.sam_image_embedding_size,
                self.sam_image_embedding_size,
            ),
            input_image_size=(self.image_size, self.image_size),
            mask_in_chans=16,
        )
        self.sam_mask_decoder = MaskDecoder(
            num_multimask_outputs=3,
            transformer=TwoWayTransformer(
                depth=2,
                embedding_dim=self.sam_prompt_embed_dim,
                mlp_dim=2048,
                num_heads=8,
            ),
            transformer_dim=self.sam_prompt_embed_dim,
            iou_head_depth=3,
            iou_head_hidden_dim=256,
            use_high_res_features=self.use_high_res_features_in_sam,
            iou_prediction_use_sigmoid=self.iou_prediction_use_sigmoid,
            pred_obj_scores=self.pred_obj_scores,
            pred_obj_scores_mlp=self.pred_obj_scores_mlp,
            use_multimask_token_for_obj_ptr=self.use_multimask_token_for_obj_ptr,
            **(self.sam_mask_decoder_extra_args or {}),
        )
        if self.use_obj_ptrs_in_encoder:
            # a linear projection on SAM output tokens to turn them into object pointers
            self.obj_ptr_proj = torch.nn.Linear(self.hidden_dim, self.hidden_dim)
            if self.use_mlp_for_obj_ptr_proj:
                self.obj_ptr_proj = MLP(
                    self.hidden_dim, self.hidden_dim, self.hidden_dim, 3
                )
        else:
            self.obj_ptr_proj = torch.nn.Identity()
        if self.proj_tpos_enc_in_obj_ptrs:
            # a linear projection on temporal positional encoding in object pointers to
            # avoid potential interference with spatial positional encoding
            self.obj_ptr_tpos_proj = torch.nn.Linear(self.hidden_dim, self.mem_dim)
        else:
            self.obj_ptr_tpos_proj = torch.nn.Identity()

    def _forward_sam_heads(
        self,
        backbone_features,
        point_inputs=None,
        mask_inputs=None,
        high_res_features=None,
        multimask_output=False,
    ):
        """
        Forward SAM prompt encoders and mask heads.

        Inputs:
        - backbone_features: image features of [B, C, H, W] shape
        - point_inputs: a dictionary with "point_coords" and "point_labels", where
          1) "point_coords" has [B, P, 2] shape and float32 dtype and contains the
             absolute pixel-unit coordinate in (x, y) format of the P input points
          2) "point_labels" has shape [B, P] and int32 dtype, where 1 means
             positive clicks, 0 means negative clicks, and -1 means padding
        - mask_inputs: a mask of [B, 1, H*16, W*16] shape, float or bool, with the
          same spatial size as the image.
        - high_res_features: either 1) None or 2) or a list of length 2 containing
          two feature maps of [B, C, 4*H, 4*W] and [B, C, 2*H, 2*W] shapes respectively,
          which will be used as high-resolution feature maps for SAM decoder.
        - multimask_output: if it's True, we output 3 candidate masks and their 3
          corresponding IoU estimates, and if it's False, we output only 1 mask and
          its corresponding IoU estimate.

        Outputs:
        - low_res_multimasks: [B, M, H*4, W*4] shape (where M = 3 if
          `multimask_output=True` and M = 1 if `multimask_output=False`), the SAM
          output mask logits (before sigmoid) for the low-resolution masks, with 4x
          the resolution (1/4 stride) of the input backbone_features.
        - high_res_multimasks: [B, M, H*16, W*16] shape (where M = 3
          if `multimask_output=True` and M = 1 if `multimask_output=False`),
          upsampled from the low-resolution masks, with shape size as the image
          (stride is 1 pixel).
        - ious, [B, M] shape, where (where M = 3 if `multimask_output=True` and M = 1
          if `multimask_output=False`), the estimated IoU of each output mask.
        - low_res_masks: [B, 1, H*4, W*4] shape, the best mask in `low_res_multimasks`.
          If `multimask_output=True`, it's the mask with the highest IoU estimate.
          If `multimask_output=False`, it's the same as `low_res_multimasks`.
        - high_res_masks: [B, 1, H*16, W*16] shape, the best mask in `high_res_multimasks`.
          If `multimask_output=True`, it's the mask with the highest IoU estimate.
          If `multimask_output=False`, it's the same as `high_res_multimasks`.
        - obj_ptr: [B, C] shape, the object pointer vector for the output mask, extracted
          based on the output token from the SAM mask decoder.
        """
        B = backbone_features.size(0)
        device = backbone_features.device
        assert backbone_features.size(1) == self.sam_prompt_embed_dim
        assert backbone_features.size(2) == self.sam_image_embedding_size
        assert backbone_features.size(3) == self.sam_image_embedding_size

        # a) Handle point prompts
        if point_inputs is not None:
            sam_point_coords = point_inputs["point_coords"]
            sam_point_labels = point_inputs["point_labels"]
            assert sam_point_coords.size(0) == B and sam_point_labels.size(0) == B
        else:
            # If no points are provide, pad with an empty point (with label -1)
            sam_point_coords = torch.zeros(B, 1, 2, device=device)
            sam_point_labels = -torch.ones(B, 1, dtype=torch.int32, device=device)

        # b) Handle mask prompts
        if mask_inputs is not None:
            # If mask_inputs is provided, downsize it into low-res mask input if needed
            # and feed it as a dense mask prompt into the SAM mask encoder
            assert len(mask_inputs.shape) == 4 and mask_inputs.shape[:2] == (B, 1)
            if mask_inputs.shape[-2:] != self.sam_prompt_encoder.mask_input_size:
                sam_mask_prompt = F.interpolate(
                    mask_inputs.float(),
                    size=self.sam_prompt_encoder.mask_input_size,
                    align_corners=False,
                    mode="bilinear",
                    antialias=True,  # use antialias for downsampling
                )
            else:
                sam_mask_prompt = mask_inputs
        else:
            # Otherwise, simply feed None (and SAM's prompt encoder will add
            # a learned `no_mask_embed` to indicate no mask input in this case).
            sam_mask_prompt = None

        sparse_embeddings, dense_embeddings = self.sam_prompt_encoder(
            points=(sam_point_coords, sam_point_labels),
            boxes=None,
            masks=sam_mask_prompt,
        )
        (
            low_res_multimasks,
            ious,
            sam_output_tokens,
            object_score_logits,
        ) = self.sam_mask_decoder(
            image_embeddings=backbone_features,
            image_pe=self.sam_prompt_encoder.get_dense_pe(),
            sparse_prompt_embeddings=sparse_embeddings,
            dense_prompt_embeddings=dense_embeddings,
            multimask_output=multimask_output,
            repeat_image=False,  # the image is already batched
            high_res_features=high_res_features,
        )
        if self.pred_obj_scores:
            is_obj_appearing = object_score_logits > 0

            # Mask used for spatial memories is always a *hard* choice between obj and no obj,
            # consistent with the actual mask prediction
            low_res_multimasks = torch.where(
                is_obj_appearing[:, None, None],
                low_res_multimasks,
                NO_OBJ_SCORE,
            )

        # convert masks from possibly bfloat16 (or float16) to float32
        # (older PyTorch versions before 2.1 don't support `interpolate` on bf16)
        _dtype = low_res_multimasks.dtype
        # low_res_multimasks = low_res_multimasks.float()
        high_res_multimasks = F.interpolate(
            low_res_multimasks.float(),
            size=(self.image_size, self.image_size),
            mode="bilinear",
            align_corners=False,
        ).to(_dtype)

        sam_output_token = sam_output_tokens[:, 0]
        if multimask_output:
            # take the best mask prediction (with the highest IoU estimation)
            best_iou_inds = torch.argmax(ious, dim=-1)
            batch_inds = torch.arange(B, device=device)
            low_res_masks = low_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
            high_res_masks = high_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
            if sam_output_tokens.size(1) > 1:
                sam_output_token = sam_output_tokens[batch_inds, best_iou_inds]
        else:
            low_res_masks, high_res_masks = low_res_multimasks, high_res_multimasks

        # Extract object pointer from the SAM output token (with occlusion handling)
        obj_ptr = self.obj_ptr_proj(sam_output_token)
        if self.pred_obj_scores:
            # Allow *soft* no obj ptr, unlike for masks
            if self.soft_no_obj_ptr:
                # Only hard possible with gt
                assert not self.teacher_force_obj_scores_for_mem
                lambda_is_obj_appearing = object_score_logits.sigmoid()
            else:
                lambda_is_obj_appearing = is_obj_appearing.float()

            if self.fixed_no_obj_ptr:
                obj_ptr = lambda_is_obj_appearing * obj_ptr
            obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr

        return (
            low_res_multimasks,
            high_res_multimasks,
            ious,
            low_res_masks,
            high_res_masks,
            obj_ptr,
            object_score_logits,
        )

    def _use_mask_as_output(self, backbone_features, high_res_features, mask_inputs):
        """
        Directly turn binary `mask_inputs` into a output mask logits without using SAM.
        (same input and output shapes as in _forward_sam_heads above).
        """
        # Use -10/+10 as logits for neg/pos pixels (very close to 0/1 in prob after sigmoid).
        out_scale, out_bias = 20.0, -10.0  # sigmoid(-10.0)=4.5398e-05
        mask_inputs_float = mask_inputs.float()
        high_res_masks = mask_inputs_float * out_scale + out_bias
        low_res_masks = F.interpolate(
            high_res_masks,
            size=(high_res_masks.size(-2) // 4, high_res_masks.size(-1) // 4),
            align_corners=False,
            mode="bilinear",
            antialias=True,  # use antialias for downsampling
        )
        # a dummy IoU prediction of all 1's under mask input
        ious = mask_inputs.new_ones(mask_inputs.size(0), 1).float()
        if not self.use_obj_ptrs_in_encoder:
            # all zeros as a dummy object pointer (of shape [B, C])
            obj_ptr = torch.zeros(
                mask_inputs.size(0), self.hidden_dim, device=mask_inputs.device
            )
        else:
            # produce an object pointer using the SAM decoder from the mask input
            _, _, _, _, _, obj_ptr, _ = self._forward_sam_heads(
                backbone_features=backbone_features,
                mask_inputs=self.mask_downsample(mask_inputs_float),
                high_res_features=high_res_features,
            )
        # In this method, we are treating mask_input as output, e.g. using it directly to create spatial mem;
        # Below, we follow the same design axiom to use mask_input to decide if obj appears or not instead of relying
        # on the object_scores from the SAM decoder.
        is_obj_appearing = torch.any(mask_inputs.flatten(1).float() > 0.0, dim=1)
        is_obj_appearing = is_obj_appearing[..., None]
        lambda_is_obj_appearing = is_obj_appearing.float()
        object_score_logits = out_scale * lambda_is_obj_appearing + out_bias
        if self.pred_obj_scores:
            if self.fixed_no_obj_ptr:
                obj_ptr = lambda_is_obj_appearing * obj_ptr
            obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr

        return (
            low_res_masks,
            high_res_masks,
            ious,
            low_res_masks,
            high_res_masks,
            obj_ptr,
            object_score_logits,
        )

    def forward_image(self, img_batch: torch.Tensor):
        """Get the image feature on the input batch."""
        backbone_out = self.image_encoder(img_batch)
        if self.use_high_res_features_in_sam:
            # precompute projected level 0 and level 1 features in SAM decoder
            # to avoid running it again on every SAM click
            backbone_out["backbone_fpn"][0] = self.sam_mask_decoder.conv_s0(
                backbone_out["backbone_fpn"][0]
            )
            backbone_out["backbone_fpn"][1] = self.sam_mask_decoder.conv_s1(
                backbone_out["backbone_fpn"][1]
            )
        return backbone_out

    def _prepare_backbone_features(self, backbone_out):
        """Prepare and flatten visual features."""
        backbone_out = backbone_out.copy()
        assert len(backbone_out["backbone_fpn"]) == len(backbone_out["vision_pos_enc"])
        assert len(backbone_out["backbone_fpn"]) >= self.num_feature_levels

        feature_maps = backbone_out["backbone_fpn"][-self.num_feature_levels :]
        vision_pos_embeds = backbone_out["vision_pos_enc"][-self.num_feature_levels :]

        feat_sizes = [(x.shape[-2], x.shape[-1]) for x in vision_pos_embeds]
        # flatten NxCxHxW to HWxNxC
        vision_feats = [x.flatten(2).permute(2, 0, 1) for x in feature_maps]
        vision_pos_embeds = [x.flatten(2).permute(2, 0, 1) for x in vision_pos_embeds]

        return backbone_out, vision_feats, vision_pos_embeds, feat_sizes

    def _prepare_memory_conditioned_features(
        self,
        frame_idx,
        is_init_cond_frame,
        current_vision_feats,
        current_vision_pos_embeds,
        feat_sizes,
        output_dict,
        num_frames,
        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
    ):
        """Fuse the current frame's visual feature map with previous memory."""
        B = current_vision_feats[-1].size(1)  # batch size on this frame
        C = self.hidden_dim
        H, W = feat_sizes[-1]  # top-level (lowest-resolution) feature size
        device = current_vision_feats[-1].device
        # The case of `self.num_maskmem == 0` below is primarily used for reproducing SAM on images.
        # In this case, we skip the fusion with any memory.
        if self.num_maskmem == 0:  # Disable memory and skip fusion
            pix_feat = current_vision_feats[-1].permute(1, 2, 0).view(B, C, H, W)
            return pix_feat

        num_obj_ptr_tokens = 0
        # Step 1: condition the visual features of the current frame on previous memories
        if not is_init_cond_frame:
            # Retrieve the memories encoded with the maskmem backbone
            to_cat_memory, to_cat_memory_pos_embed = [], []
            # Add conditioning frames's output first (all cond frames have t_pos=0 for
            # when getting temporal positional embedding below)
            assert len(output_dict["cond_frame_outputs"]) > 0
            # Select a maximum number of temporally closest cond frames for cross attention
            cond_outputs = output_dict["cond_frame_outputs"]
            selected_cond_outputs, unselected_cond_outputs = select_closest_cond_frames(
                frame_idx, cond_outputs, self.max_cond_frames_in_attn
            )
            t_pos_and_prevs = [(0, out) for out in selected_cond_outputs.values()]
            # Add last (self.num_maskmem - 1) frames before current frame for non-conditioning memory
            # the earliest one has t_pos=1 and the latest one has t_pos=self.num_maskmem-1
            # We also allow taking the memory frame non-consecutively (with r>1), in which case
            # we take (self.num_maskmem - 2) frames among every r-th frames plus the last frame.
            r = self.memory_temporal_stride_for_eval
            for t_pos in range(1, self.num_maskmem):
                t_rel = self.num_maskmem - t_pos  # how many frames before current frame
                if t_rel == 1:
                    # for t_rel == 1, we take the last frame (regardless of r)
                    if not track_in_reverse:
                        # the frame immediately before this frame (i.e. frame_idx - 1)
                        prev_frame_idx = frame_idx - t_rel
                    else:
                        # the frame immediately after this frame (i.e. frame_idx + 1)
                        prev_frame_idx = frame_idx + t_rel
                else:
                    # for t_rel >= 2, we take the memory frame from every r-th frames
                    if not track_in_reverse:
                        # first find the nearest frame among every r-th frames before this frame
                        # for r=1, this would be (frame_idx - 2)
                        prev_frame_idx = ((frame_idx - 2) // r) * r
                        # then seek further among every r-th frames
                        prev_frame_idx = prev_frame_idx - (t_rel - 2) * r
                    else:
                        # first find the nearest frame among every r-th frames after this frame
                        # for r=1, this would be (frame_idx + 2)
                        prev_frame_idx = -(-(frame_idx + 2) // r) * r
                        # then seek further among every r-th frames
                        prev_frame_idx = prev_frame_idx + (t_rel - 2) * r
                out = output_dict["non_cond_frame_outputs"].get(prev_frame_idx, None)
                if out is None:
                    # If an unselected conditioning frame is among the last (self.num_maskmem - 1)
                    # frames, we still attend to it as if it's a non-conditioning frame.
                    out = unselected_cond_outputs.get(prev_frame_idx, None)
                t_pos_and_prevs.append((t_pos, out))

            for t_pos, prev in t_pos_and_prevs:
                if prev is None:
                    continue  # skip padding frames
                # "maskmem_features" might have been offloaded to CPU in demo use cases,
                # so we load it back to GPU (it's a no-op if it's already on GPU).
                feats = prev["maskmem_features"].cuda(non_blocking=True)
                to_cat_memory.append(feats.flatten(2).permute(2, 0, 1))
                # Spatial positional encoding (it might have been offloaded to CPU in eval)
                maskmem_enc = prev["maskmem_pos_enc"][-1].cuda()
                maskmem_enc = maskmem_enc.flatten(2).permute(2, 0, 1)
                # Temporal positional encoding
                maskmem_enc = (
                    maskmem_enc + self.maskmem_tpos_enc[self.num_maskmem - t_pos - 1]
                )
                to_cat_memory_pos_embed.append(maskmem_enc)

            # Construct the list of past object pointers
            if self.use_obj_ptrs_in_encoder:
                max_obj_ptrs_in_encoder = min(num_frames, self.max_obj_ptrs_in_encoder)
                # First add those object pointers from selected conditioning frames
                # (optionally, only include object pointers in the past during evaluation)
                if not self.training and self.only_obj_ptrs_in_the_past_for_eval:
                    ptr_cond_outputs = {
                        t: out
                        for t, out in selected_cond_outputs.items()
                        if (t >= frame_idx if track_in_reverse else t <= frame_idx)
                    }
                else:
                    ptr_cond_outputs = selected_cond_outputs
                pos_and_ptrs = [
                    # Temporal pos encoding contains how far away each pointer is from current frame
                    (abs(frame_idx - t), out["obj_ptr"])
                    for t, out in ptr_cond_outputs.items()
                ]
                # Add up to (max_obj_ptrs_in_encoder - 1) non-conditioning frames before current frame
                for t_diff in range(1, max_obj_ptrs_in_encoder):
                    t = frame_idx + t_diff if track_in_reverse else frame_idx - t_diff
                    if t < 0 or (num_frames is not None and t >= num_frames):
                        break
                    out = output_dict["non_cond_frame_outputs"].get(
                        t, unselected_cond_outputs.get(t, None)
                    )
                    if out is not None:
                        pos_and_ptrs.append((t_diff, out["obj_ptr"]))
                # If we have at least one object pointer, add them to the across attention
                if len(pos_and_ptrs) > 0:
                    pos_list, ptrs_list = zip(*pos_and_ptrs)
                    # stack object pointers along dim=0 into [ptr_seq_len, B, C] shape
                    obj_ptrs = torch.stack(ptrs_list, dim=0)
                    # a temporal positional embedding based on how far each object pointer is from
                    # the current frame (sine embedding normalized by the max pointer num).
                    if self.add_tpos_enc_to_obj_ptrs:
                        t_diff_max = max_obj_ptrs_in_encoder - 1
                        tpos_dim = C if self.proj_tpos_enc_in_obj_ptrs else self.mem_dim
                        obj_pos = torch.tensor(pos_list, device=device)
                        obj_pos = get_1d_sine_pe(obj_pos / t_diff_max, dim=tpos_dim)
                        obj_pos = self.obj_ptr_tpos_proj(obj_pos)
                        obj_pos = obj_pos.unsqueeze(1).expand(-1, B, self.mem_dim)
                    else:
                        obj_pos = obj_ptrs.new_zeros(len(pos_list), B, self.mem_dim)
                    if self.mem_dim < C:
                        # split a pointer into (C // self.mem_dim) tokens for self.mem_dim < C
                        obj_ptrs = obj_ptrs.reshape(
                            -1, B, C // self.mem_dim, self.mem_dim
                        )
                        obj_ptrs = obj_ptrs.permute(0, 2, 1, 3).flatten(0, 1)
                        obj_pos = obj_pos.repeat_interleave(C // self.mem_dim, dim=0)
                    to_cat_memory.append(obj_ptrs)
                    to_cat_memory_pos_embed.append(obj_pos)
                    num_obj_ptr_tokens = obj_ptrs.shape[0]
                else:
                    num_obj_ptr_tokens = 0
        else:
            # for initial conditioning frames, encode them without using any previous memory
            if self.directly_add_no_mem_embed:
                # directly add no-mem embedding (instead of using the transformer encoder)
                pix_feat_with_mem = current_vision_feats[-1] + self.no_mem_embed
                pix_feat_with_mem = pix_feat_with_mem.permute(1, 2, 0).view(B, C, H, W)
                return pix_feat_with_mem

            # Use a dummy token on the first frame (to avoid emtpy memory input to tranformer encoder)
            to_cat_memory = [self.no_mem_embed.expand(1, B, self.mem_dim)]
            to_cat_memory_pos_embed = [self.no_mem_pos_enc.expand(1, B, self.mem_dim)]

        # Step 2: Concatenate the memories and forward through the transformer encoder
        memory = torch.cat(to_cat_memory, dim=0)
        memory_pos_embed = torch.cat(to_cat_memory_pos_embed, dim=0)

        pix_feat_with_mem = self.memory_attention(
            curr=current_vision_feats,
            curr_pos=current_vision_pos_embeds,
            memory=memory,
            memory_pos=memory_pos_embed,
            num_obj_ptr_tokens=num_obj_ptr_tokens,
        )
        # reshape the output (HW)BC => BCHW
        pix_feat_with_mem = pix_feat_with_mem.permute(1, 2, 0).view(B, C, H, W)
        return pix_feat_with_mem

    def _encode_new_memory(
        self,
        current_vision_feats,
        feat_sizes,
        pred_masks_high_res,
        is_mask_from_pts,
    ):
        """Encode the current image and its prediction into a memory feature."""
        B = current_vision_feats[-1].size(1)  # batch size on this frame
        C = self.hidden_dim
        H, W = feat_sizes[-1]  # top-level (lowest-resolution) feature size
        # top-level feature, (HW)BC => BCHW
        pix_feat = current_vision_feats[-1].permute(1, 2, 0).view(B, C, H, W)
        if self.non_overlap_masks_for_mem_enc and not self.training:
            # optionally, apply non-overlapping constraints to the masks (it's applied
            # in the batch dimension and should only be used during eval, where all
            # the objects come from the same video under batch size 1).
            pred_masks_high_res = self._apply_non_overlapping_constraints(
                pred_masks_high_res
            )
        # scale the raw mask logits with a temperature before applying sigmoid
        binarize = self.binarize_mask_from_pts_for_mem_enc and is_mask_from_pts
        if binarize and not self.training:
            mask_for_mem = (pred_masks_high_res > 0).float()
        else:
            # apply sigmoid on the raw mask logits to turn them into range (0, 1)
            mask_for_mem = torch.sigmoid(pred_masks_high_res)
        # apply scale and bias terms to the sigmoid probabilities
        if self.sigmoid_scale_for_mem_enc != 1.0:
            mask_for_mem = mask_for_mem * self.sigmoid_scale_for_mem_enc
        if self.sigmoid_bias_for_mem_enc != 0.0:
            mask_for_mem = mask_for_mem + self.sigmoid_bias_for_mem_enc
        maskmem_out = self.memory_encoder(
            pix_feat, mask_for_mem, skip_mask_sigmoid=True  # sigmoid already applied
        )
        maskmem_features = maskmem_out["vision_features"]
        maskmem_pos_enc = maskmem_out["vision_pos_enc"]

        return maskmem_features, maskmem_pos_enc

    def track_step(
        self,
        frame_idx,
        is_init_cond_frame,
        current_vision_feats,
        current_vision_pos_embeds,
        feat_sizes,
        point_inputs,
        mask_inputs,
        output_dict,
        num_frames,
        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
        # Whether to run the memory encoder on the predicted masks. Sometimes we might want
        # to skip the memory encoder with `run_mem_encoder=False`. For example,
        # in demo we might call `track_step` multiple times for each user click,
        # and only encode the memory when the user finalizes their clicks. And in ablation
        # settings like SAM training on static images, we don't need the memory encoder.
        run_mem_encoder=True,
        # The previously predicted SAM mask logits (which can be fed together with new clicks in demo).
        prev_sam_mask_logits=None,
    ):
        current_out = {"point_inputs": point_inputs, "mask_inputs": mask_inputs}
        # High-resolution feature maps for the SAM head, reshape (HW)BC => BCHW
        if len(current_vision_feats) > 1:
            high_res_features = [
                x.permute(1, 2, 0).view(x.size(1), x.size(2), *s)
                for x, s in zip(current_vision_feats[:-1], feat_sizes[:-1])
            ]
        else:
            high_res_features = None
        if mask_inputs is not None and self.use_mask_input_as_output_without_sam:
            # When use_mask_input_as_output_without_sam=True, we directly output the mask input
            # (see it as a GT mask) without using a SAM prompt encoder + mask decoder.
            pix_feat = current_vision_feats[-1].permute(1, 2, 0)
            pix_feat = pix_feat.view(-1, self.hidden_dim, *feat_sizes[-1])
            sam_outputs = self._use_mask_as_output(
                pix_feat, high_res_features, mask_inputs
            )
        else:
            # fused the visual feature with previous memory features in the memory bank
            pix_feat_with_mem = self._prepare_memory_conditioned_features(
                frame_idx=frame_idx,
                is_init_cond_frame=is_init_cond_frame,
                current_vision_feats=current_vision_feats[-1:],
                current_vision_pos_embeds=current_vision_pos_embeds[-1:],
                feat_sizes=feat_sizes[-1:],
                output_dict=output_dict,
                num_frames=num_frames,
                track_in_reverse=track_in_reverse,
            )
            # apply SAM-style segmentation head
            # here we might feed previously predicted low-res SAM mask logits into the SAM mask decoder,
            # e.g. in demo where such logits come from earlier interaction instead of correction sampling
            # (in this case, any `mask_inputs` shouldn't reach here as they are sent to _use_mask_as_output instead)
            if prev_sam_mask_logits is not None:
                assert point_inputs is not None and mask_inputs is None
                mask_inputs = prev_sam_mask_logits
            multimask_output = self._use_multimask(is_init_cond_frame, point_inputs)
            sam_outputs = self._forward_sam_heads(
                backbone_features=pix_feat_with_mem,
                point_inputs=point_inputs,
                mask_inputs=mask_inputs,
                high_res_features=high_res_features,
                multimask_output=multimask_output,
            )
        (
            _,
            _,
            _,
            low_res_masks,
            high_res_masks,
            obj_ptr,
            _,
        ) = sam_outputs

        current_out["pred_masks"] = low_res_masks
        current_out["pred_masks_high_res"] = high_res_masks
        current_out["obj_ptr"] = obj_ptr

        # Finally run the memory encoder on the predicted mask to encode
        # it into a new memory feature (that can be used in future frames)
        if run_mem_encoder and self.num_maskmem > 0:
            high_res_masks_for_mem_enc = high_res_masks
            maskmem_features, maskmem_pos_enc = self._encode_new_memory(
                current_vision_feats=current_vision_feats,
                feat_sizes=feat_sizes,
                pred_masks_high_res=high_res_masks_for_mem_enc,
                is_mask_from_pts=(point_inputs is not None),
            )
            current_out["maskmem_features"] = maskmem_features
            current_out["maskmem_pos_enc"] = maskmem_pos_enc
        else:
            current_out["maskmem_features"] = None
            current_out["maskmem_pos_enc"] = None

        return current_out

    def _use_multimask(self, is_init_cond_frame, point_inputs):
        """Whether to use multimask output in the SAM head."""
        num_pts = 0 if point_inputs is None else point_inputs["point_labels"].size(1)
        multimask_output = (
            self.multimask_output_in_sam
            and (is_init_cond_frame or self.multimask_output_for_tracking)
            and (self.multimask_min_pt_num <= num_pts <= self.multimask_max_pt_num)
        )
        return multimask_output

    def _apply_non_overlapping_constraints(self, pred_masks):
        """
        Apply non-overlapping constraints to the object scores in pred_masks. Here we
        keep only the highest scoring object at each spatial location in pred_masks.
        """
        batch_size = pred_masks.size(0)
        if batch_size == 1:
            return pred_masks

        device = pred_masks.device
        # "max_obj_inds": object index of the object with the highest score at each location
        max_obj_inds = torch.argmax(pred_masks, dim=0, keepdim=True)
        # "batch_obj_inds": object index of each object slice (along dim 0) in `pred_masks`
        batch_obj_inds = torch.arange(batch_size, device=device)[:, None, None, None]
        keep = max_obj_inds == batch_obj_inds
        # suppress overlapping regions' scores below -10.0 so that the foreground regions
        # don't overlap (here sigmoid(-10.0)=4.5398e-05)
        pred_masks = torch.where(keep, pred_masks, torch.clamp(pred_masks, max=-10.0))
        return pred_masks

class SAM2Base(SAM2Base_):

    def track_step(
        self,
        frame_idx,
        is_init_cond_frame,
        current_vision_feats,
        current_vision_pos_embeds,
        feat_sizes,
        point_inputs,
        mask_inputs,
        output_dict,
        num_frames,
        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
        # Whether to run the memory encoder on the predicted masks. Sometimes we might want
        # to skip the memory encoder with `run_mem_encoder=False`. For example,
        # in demo we might call `track_step` multiple times for each user click,
        # and only encode the memory when the user finalizes their clicks. And in ablation
        # settings like SAM training on static images, we don't need the memory encoder.
        run_mem_encoder=True,
        # The previously predicted SAM mask logits (which can be fed together with new clicks in demo).
        prev_sam_mask_logits=None,
        ## Extension: LLM prompt
        language_embd=None,
    ):
        current_out = {"point_inputs": point_inputs, "mask_inputs": mask_inputs}
        # High-resolution feature maps for the SAM head, reshape (HW)BC => BCHW
        if len(current_vision_feats) > 1:
            high_res_features = [
                x.permute(1, 2, 0).view(x.size(1), x.size(2), *s)
                for x, s in zip(current_vision_feats[:-1], feat_sizes[:-1])
            ]
        else:
            high_res_features = None
        if mask_inputs is not None and self.use_mask_input_as_output_without_sam:
            # When use_mask_input_as_output_without_sam=True, we directly output the mask input
            # (see it as a GT mask) without using a SAM prompt encoder + mask decoder.
            pix_feat = current_vision_feats[-1].permute(1, 2, 0)
            pix_feat = pix_feat.view(-1, self.hidden_dim, *feat_sizes[-1])
            sam_outputs = self._use_mask_as_output(
                pix_feat, high_res_features, mask_inputs
            )
        else:
            # fused the visual feature with previous memory features in the memory bank
            pix_feat_with_mem = self._prepare_memory_conditioned_features(
                frame_idx=frame_idx,
                is_init_cond_frame=is_init_cond_frame,
                current_vision_feats=current_vision_feats[-1:],
                current_vision_pos_embeds=current_vision_pos_embeds[-1:],
                feat_sizes=feat_sizes[-1:],
                output_dict=output_dict,
                num_frames=num_frames,
                track_in_reverse=track_in_reverse,
            )
            # apply SAM-style segmentation head
            # here we might feed previously predicted low-res SAM mask logits into the SAM mask decoder,
            # e.g. in demo where such logits come from earlier interaction instead of correction sampling
            # (in this case, any `mask_inputs` shouldn't reach here as they are sent to _use_mask_as_output instead)
            if prev_sam_mask_logits is not None:
                assert point_inputs is not None and mask_inputs is None
                mask_inputs = prev_sam_mask_logits
            multimask_output = self._use_multimask(is_init_cond_frame, point_inputs)
            sam_outputs = self._forward_sam_heads(
                backbone_features=pix_feat_with_mem,
                point_inputs=point_inputs,
                mask_inputs=mask_inputs,
                high_res_features=high_res_features,
                multimask_output=multimask_output,
                # Inject language Embed if possible
                language_embd=language_embd,
            )
        (
            _,
            _,
            _,
            low_res_masks,
            high_res_masks,
            obj_ptr,
            _,
        ) = sam_outputs

        current_out["pred_masks"] = low_res_masks
        current_out["pred_masks_high_res"] = high_res_masks
        current_out["obj_ptr"] = obj_ptr

        # Finally run the memory encoder on the predicted mask to encode
        # it into a new memory feature (that can be used in future frames)
        if run_mem_encoder and self.num_maskmem > 0:
            high_res_masks_for_mem_enc = high_res_masks
            maskmem_features, maskmem_pos_enc = self._encode_new_memory(
                current_vision_feats=current_vision_feats,
                feat_sizes=feat_sizes,
                pred_masks_high_res=high_res_masks_for_mem_enc,
                is_mask_from_pts=(point_inputs is not None),
            )
            current_out["maskmem_features"] = maskmem_features
            current_out["maskmem_pos_enc"] = maskmem_pos_enc
        else:
            current_out["maskmem_features"] = None
            current_out["maskmem_pos_enc"] = None

        return current_out


    def _forward_sam_heads(
        self,
        backbone_features,
        point_inputs=None,
        mask_inputs=None,
        high_res_features=None,
        multimask_output=False,
        ## Extension: LLM prompt
        language_embd=None,
    ):
        """
        Forward SAM prompt encoders and mask heads.

        Inputs:
        - backbone_features: image features of [B, C, H, W] shape
        - point_inputs: a dictionary with "point_coords" and "point_labels", where
          1) "point_coords" has [B, P, 2] shape and float32 dtype and contains the
             absolute pixel-unit coordinate in (x, y) format of the P input points
          2) "point_labels" has shape [B, P] and int32 dtype, where 1 means
             positive clicks, 0 means negative clicks, and -1 means padding
        - mask_inputs: a mask of [B, 1, H*16, W*16] shape, float or bool, with the
          same spatial size as the image.
        - high_res_features: either 1) None or 2) or a list of length 2 containing
          two feature maps of [B, C, 4*H, 4*W] and [B, C, 2*H, 2*W] shapes respectively,
          which will be used as high-resolution feature maps for SAM decoder.
        - multimask_output: if it's True, we output 3 candidate masks and their 3
          corresponding IoU estimates, and if it's False, we output only 1 mask and
          its corresponding IoU estimate.

        Outputs:
        - low_res_multimasks: [B, M, H*4, W*4] shape (where M = 3 if
          `multimask_output=True` and M = 1 if `multimask_output=False`), the SAM
          output mask logits (before sigmoid) for the low-resolution masks, with 4x
          the resolution (1/4 stride) of the input backbone_features.
        - high_res_multimasks: [B, M, H*16, W*16] shape (where M = 3
          if `multimask_output=True` and M = 1 if `multimask_output=False`),
          upsampled from the low-resolution masks, with shape size as the image
          (stride is 1 pixel).
        - ious, [B, M] shape, where (where M = 3 if `multimask_output=True` and M = 1
          if `multimask_output=False`), the estimated IoU of each output mask.
        - low_res_masks: [B, 1, H*4, W*4] shape, the best mask in `low_res_multimasks`.
          If `multimask_output=True`, it's the mask with the highest IoU estimate.
          If `multimask_output=False`, it's the same as `low_res_multimasks`.
        - high_res_masks: [B, 1, H*16, W*16] shape, the best mask in `high_res_multimasks`.
          If `multimask_output=True`, it's the mask with the highest IoU estimate.
          If `multimask_output=False`, it's the same as `high_res_multimasks`.
        - obj_ptr: [B, C] shape, the object pointer vector for the output mask, extracted
          based on the output token from the SAM mask decoder.
        """
        B = backbone_features.size(0)
        device = backbone_features.device
        assert backbone_features.size(1) == self.sam_prompt_embed_dim
        assert backbone_features.size(2) == self.sam_image_embedding_size
        assert backbone_features.size(3) == self.sam_image_embedding_size

        # a) Handle point prompts
        if point_inputs is not None:
            sam_point_coords = point_inputs["point_coords"]
            sam_point_labels = point_inputs["point_labels"]
            assert sam_point_coords.size(0) == B and sam_point_labels.size(0) == B
        else:
            # If no points are provide, pad with an empty point (with label -1)
            sam_point_coords = torch.zeros(B, 1, 2, device=device)
            sam_point_labels = -torch.ones(B, 1, dtype=torch.int32, device=device)

        # b) Handle mask prompts
        if mask_inputs is not None:
            # If mask_inputs is provided, downsize it into low-res mask input if needed
            # and feed it as a dense mask prompt into the SAM mask encoder
            assert len(mask_inputs.shape) == 4 and mask_inputs.shape[:2] == (B, 1)
            if mask_inputs.shape[-2:] != self.sam_prompt_encoder.mask_input_size:
                sam_mask_prompt = F.interpolate(
                    mask_inputs.float(),
                    size=self.sam_prompt_encoder.mask_input_size,
                    align_corners=False,
                    mode="bilinear",
                    antialias=True,  # use antialias for downsampling
                )
            else:
                sam_mask_prompt = mask_inputs
        else:
            # Otherwise, simply feed None (and SAM's prompt encoder will add
            # a learned `no_mask_embed` to indicate no mask input in this case).
            sam_mask_prompt = None

        sparse_embeddings, dense_embeddings = self.sam_prompt_encoder(
            points=(sam_point_coords, sam_point_labels),
            boxes=None,
            masks=sam_mask_prompt,
        )

        ## Extension: LLM prompt
        if language_embd is not None:
            # B N C
            assert sparse_embeddings.size(0) == language_embd.size(0)
            assert sparse_embeddings.size(2) == language_embd.size(2)
            sparse_embeddings = torch.cat([sparse_embeddings, language_embd], dim=1)

        (
            low_res_multimasks,
            ious,
            sam_output_tokens,
            object_score_logits,
        ) = self.sam_mask_decoder(
            image_embeddings=backbone_features,
            image_pe=self.sam_prompt_encoder.get_dense_pe(),
            sparse_prompt_embeddings=sparse_embeddings,
            dense_prompt_embeddings=dense_embeddings,
            multimask_output=multimask_output,
            repeat_image=False,  # the image is already batched
            high_res_features=high_res_features,
        )
        if self.pred_obj_scores:
            is_obj_appearing = object_score_logits > 0

            # Mask used for spatial memories is always a *hard* choice between obj and no obj,
            # consistent with the actual mask prediction
            # print('Do torch.where !!!')
            # low_res_multimasks = torch.where(
            #     is_obj_appearing[:, None, None],
            #     low_res_multimasks,
            #     NO_OBJ_SCORE,
            # )

        # convert masks from possibly bfloat16 (or float16) to float32
        # (older PyTorch versions before 2.1 don't support `interpolate` on bf16)
        low_res_multimasks = low_res_multimasks.float()
        high_res_multimasks = F.interpolate(
            low_res_multimasks,
            size=(self.image_size, self.image_size),
            mode="bilinear",
            align_corners=False,
        )

        sam_output_token = sam_output_tokens[:, 0]
        if multimask_output:
            # take the best mask prediction (with the highest IoU estimation)
            best_iou_inds = torch.argmax(ious, dim=-1)
            batch_inds = torch.arange(B, device=device)
            low_res_masks = low_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
            high_res_masks = high_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
            if sam_output_tokens.size(1) > 1:
                sam_output_token = sam_output_tokens[batch_inds, best_iou_inds]
        else:
            low_res_masks, high_res_masks = low_res_multimasks, high_res_multimasks

        # Extract object pointer from the SAM output token (with occlusion handling)
        obj_ptr = self.obj_ptr_proj(sam_output_token)
        if self.pred_obj_scores:
            # Allow *soft* no obj ptr, unlike for masks
            if self.soft_no_obj_ptr:
                # Only hard possible with gt
                assert not self.teacher_force_obj_scores_for_mem
                lambda_is_obj_appearing = object_score_logits.sigmoid()
            else:
                lambda_is_obj_appearing = is_obj_appearing.float()

            if self.fixed_no_obj_ptr:
                obj_ptr = lambda_is_obj_appearing * obj_ptr
            obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr

        return (
            low_res_multimasks,
            high_res_multimasks,
            ious,
            low_res_masks,
            high_res_masks,
            obj_ptr,
            object_score_logits,
        )


def _obj_id_to_idx(inference_state, obj_id):
    """Map client-side object id to model-side object index."""
    obj_idx = inference_state["obj_id_to_idx"].get(obj_id, None)
    if obj_idx is not None:
        return obj_idx

    # This is a new object id not sent to the server before. We only allow adding
    # new objects *before* the tracking starts.
    allow_new_object = not inference_state["tracking_has_started"]
    if allow_new_object:
        # get the next object slot
        obj_idx = len(inference_state["obj_id_to_idx"])
        inference_state["obj_id_to_idx"][obj_id] = obj_idx
        inference_state["obj_idx_to_id"][obj_idx] = obj_id
        inference_state["obj_ids"] = list(inference_state["obj_id_to_idx"])
        # set up input and output structures for this object
        inference_state["point_inputs_per_obj"][obj_idx] = {}
        inference_state["mask_inputs_per_obj"][obj_idx] = {}
        inference_state["output_dict_per_obj"][obj_idx] = {
            "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
            "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
        }
        inference_state["temp_output_dict_per_obj"][obj_idx] = {
            "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
            "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
        }
        return obj_idx
    else:
        raise RuntimeError(
            f"Cannot add new object id {obj_id} after tracking starts. "
            f"All existing object ids: {inference_state['obj_ids']}. "
            f"Please call 'reset_state' to restart from scratch."
        )


def _get_maskmem_pos_enc(inference_state, current_out):
    """
    `maskmem_pos_enc` is the same across frames and objects, so we cache it as
    a constant in the inference session to reduce session storage size.
    """
    model_constants = inference_state["constants"]
    # "out_maskmem_pos_enc" should be either a list of tensors or None
    out_maskmem_pos_enc = current_out["maskmem_pos_enc"]
    if out_maskmem_pos_enc is not None:
        if "maskmem_pos_enc" not in model_constants:
            assert isinstance(out_maskmem_pos_enc, list)
            # only take the slice for one object, since it's same across objects
            maskmem_pos_enc = [x[0:1].clone() for x in out_maskmem_pos_enc]
            model_constants["maskmem_pos_enc"] = maskmem_pos_enc
        else:
            maskmem_pos_enc = model_constants["maskmem_pos_enc"]
        # expand the cached maskmem_pos_enc to the actual batch size
        batch_size = out_maskmem_pos_enc[0].size(0)
        expanded_maskmem_pos_enc = [
            x.expand(batch_size, -1, -1, -1) for x in maskmem_pos_enc
        ]
    else:
        expanded_maskmem_pos_enc = None
    return expanded_maskmem_pos_enc


def _obj_idx_to_id(inference_state, obj_idx):
    """Map model-side object index to client-side object id."""
    return inference_state["obj_idx_to_id"][obj_idx]


def _get_obj_num(inference_state):
    """Get the total number of unique object ids received so far in this session."""
    return len(inference_state["obj_idx_to_id"])


class SAM2VideoPredictor(SAM2Base):
    """The predictor class to handle user interactions and manage inference states."""

    def __init__(
        self,
        fill_hole_area=0,
        # whether to apply non-overlapping constraints on the output object masks
        non_overlap_masks=False,
        # whether to clear non-conditioning memory of the surrounding frames (which may contain outdated information) after adding correction clicks;
        # note that this would only apply to *single-object tracking* unless `clear_non_cond_mem_for_multi_obj` is also set to True)
        clear_non_cond_mem_around_input=False,
        # whether to also clear non-conditioning memory of the surrounding frames (only effective when `clear_non_cond_mem_around_input` is True).
        clear_non_cond_mem_for_multi_obj=False,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.fill_hole_area = fill_hole_area
        self.non_overlap_masks = non_overlap_masks
        self.clear_non_cond_mem_around_input = clear_non_cond_mem_around_input
        self.clear_non_cond_mem_for_multi_obj = clear_non_cond_mem_for_multi_obj

    def _get_image_feature(self, inference_state, frame_idx, batch_size):
        """Compute the image features on a given frame."""
        # Look up in the cache first
        image, backbone_out = inference_state["cached_features"].get(
            frame_idx, (None, None)
        )
        if backbone_out is None:
            # Cache miss -- we will run inference on a single image
            # image = inference_state["images"][frame_idx].cuda().float().unsqueeze(0)
            image = inference_state["images"][frame_idx].cuda().unsqueeze(0)
            backbone_out = self.forward_image(image)
            # Cache the most recent frame's feature (for repeated interactions with
            # a frame; we can use an LRU cache for more frames in the future).
            inference_state["cached_features"] = {frame_idx: (image, backbone_out)}

        # expand the features to have the same dimension as the number of objects
        expanded_image = image.expand(batch_size, -1, -1, -1)
        expanded_backbone_out = {
            "backbone_fpn": backbone_out["backbone_fpn"].copy(),
            "vision_pos_enc": backbone_out["vision_pos_enc"].copy(),
        }
        for i, feat in enumerate(expanded_backbone_out["backbone_fpn"]):
            expanded_backbone_out["backbone_fpn"][i] = feat.expand(
                batch_size, -1, -1, -1
            )
        for i, pos in enumerate(expanded_backbone_out["vision_pos_enc"]):
            pos = pos.expand(batch_size, -1, -1, -1)
            expanded_backbone_out["vision_pos_enc"][i] = pos

        features = self._prepare_backbone_features(expanded_backbone_out)
        features = (expanded_image,) + features
        return features


    def _run_single_frame_inference(
        self,
        inference_state,
        output_dict,
        frame_idx,
        batch_size,
        is_init_cond_frame,
        point_inputs,
        mask_inputs,
        reverse,
        run_mem_encoder,
        prev_sam_mask_logits=None,
        ## Extension: LLM prompt
        language_embd=None,
    ):
        """Run tracking on a single frame based on current inputs and previous memory."""
        # Retrieve correct image features
        (
            _,
            _,
            current_vision_feats,
            current_vision_pos_embeds,
            feat_sizes,
        ) = self._get_image_feature(inference_state, frame_idx, batch_size)

        # point and mask should not appear as input simultaneously on the same frame
        assert point_inputs is None or mask_inputs is None
        current_out = self.track_step(
            frame_idx=frame_idx,
            is_init_cond_frame=is_init_cond_frame,
            current_vision_feats=current_vision_feats,
            current_vision_pos_embeds=current_vision_pos_embeds,
            feat_sizes=feat_sizes,
            point_inputs=point_inputs,
            mask_inputs=mask_inputs,
            output_dict=output_dict,
            num_frames=inference_state["num_frames"],
            track_in_reverse=reverse,
            run_mem_encoder=run_mem_encoder,
            prev_sam_mask_logits=prev_sam_mask_logits,
            language_embd=language_embd,
        )

        # optionally offload the output to CPU memory to save GPU space
        storage_device = inference_state["storage_device"]
        maskmem_features = current_out["maskmem_features"]
        if maskmem_features is not None:
            maskmem_features = maskmem_features.to(torch.bfloat16)
            maskmem_features = maskmem_features.to(storage_device, non_blocking=True)
        pred_masks_gpu = current_out["pred_masks"]
        # potentially fill holes in the predicted masks
        if self.fill_hole_area > 0:
            pred_masks_gpu = fill_holes_in_mask_scores(
                pred_masks_gpu, self.fill_hole_area
            )
        pred_masks = pred_masks_gpu.to(storage_device, non_blocking=True)
        # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
        maskmem_pos_enc = _get_maskmem_pos_enc(inference_state, current_out)
        # object pointer is a small tensor, so we always keep it on GPU memory for fast access
        obj_ptr = current_out["obj_ptr"]
        # make a compact version of this frame's output to reduce the state size
        compact_current_out = {
            "maskmem_features": maskmem_features,
            "maskmem_pos_enc": maskmem_pos_enc,
            "pred_masks": pred_masks,
            "obj_ptr": obj_ptr,
        }
        return compact_current_out, pred_masks_gpu


    def _consolidate_temp_output_across_obj(
        self,
        inference_state,
        frame_idx,
        is_cond,
        run_mem_encoder,
        consolidate_at_video_res=False,
    ):
        """
        Consolidate the per-object temporary outputs in `temp_output_dict_per_obj` on
        a frame into a single output for all objects, including
        1) fill any missing objects either from `output_dict_per_obj` (if they exist in
           `output_dict_per_obj` for this frame) or leave them as placeholder values
           (if they don't exist in `output_dict_per_obj` for this frame);
        2) if specified, rerun memory encoder after apply non-overlapping constraints
           on the object scores.
        """
        batch_size = _get_obj_num(inference_state)
        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
        # Optionally, we allow consolidating the temporary outputs at the original
        # video resolution (to provide a better editing experience for mask prompts).
        if consolidate_at_video_res:
            assert not run_mem_encoder, "memory encoder cannot run at video resolution"
            consolidated_H = inference_state["video_height"]
            consolidated_W = inference_state["video_width"]
            consolidated_mask_key = "pred_masks_video_res"
        else:
            consolidated_H = consolidated_W = self.image_size // 4
            consolidated_mask_key = "pred_masks"

        # Initialize `consolidated_out`. Its "maskmem_features" and "maskmem_pos_enc"
        # will be added when rerunning the memory encoder after applying non-overlapping
        # constraints to object scores. Its "pred_masks" are prefilled with a large
        # negative value (NO_OBJ_SCORE) to represent missing objects.
        consolidated_out = {
            "maskmem_features": None,
            "maskmem_pos_enc": None,
            consolidated_mask_key: torch.full(
                size=(batch_size, 1, consolidated_H, consolidated_W),
                fill_value=NO_OBJ_SCORE,
                dtype=torch.float32,
                device=inference_state["storage_device"],
            ),
            "obj_ptr": torch.full(
                size=(batch_size, self.hidden_dim),
                fill_value=NO_OBJ_SCORE,
                dtype=torch.float32,
                device=inference_state["device"],
            ),
        }
        empty_mask_ptr = None
        for obj_idx in range(batch_size):
            obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
            obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
            out = obj_temp_output_dict[storage_key].get(frame_idx, None)
            # If the object doesn't appear in "temp_output_dict_per_obj" on this frame,
            # we fall back and look up its previous output in "output_dict_per_obj".
            # We look up both "cond_frame_outputs" and "non_cond_frame_outputs" in
            # "output_dict_per_obj" to find a previous output for this object.
            if out is None:
                out = obj_output_dict["cond_frame_outputs"].get(frame_idx, None)
            if out is None:
                out = obj_output_dict["non_cond_frame_outputs"].get(frame_idx, None)
            # If the object doesn't appear in "output_dict_per_obj" either, we skip it
            # and leave its mask scores to the default scores (i.e. the NO_OBJ_SCORE
            # placeholder above) and set its object pointer to be a dummy pointer.
            if out is None:
                # Fill in dummy object pointers for those objects without any inputs or
                # tracking outcomes on this frame (only do it under `run_mem_encoder=True`,
                # i.e. when we need to build the memory for tracking).
                if run_mem_encoder:
                    if empty_mask_ptr is None:
                        empty_mask_ptr = self._get_empty_mask_ptr(
                            inference_state, frame_idx
                        )
                    # fill object pointer with a dummy pointer (based on an empty mask)
                    consolidated_out["obj_ptr"][obj_idx : obj_idx + 1] = empty_mask_ptr
                continue
            # Add the temporary object output mask to consolidated output mask
            obj_mask = out["pred_masks"]
            consolidated_pred_masks = consolidated_out[consolidated_mask_key]
            if obj_mask.shape[-2:] == consolidated_pred_masks.shape[-2:]:
                consolidated_pred_masks[obj_idx : obj_idx + 1] = obj_mask
            else:
                # Resize first if temporary object mask has a different resolution
                resized_obj_mask = torch.nn.functional.interpolate(
                    obj_mask,
                    size=consolidated_pred_masks.shape[-2:],
                    mode="bilinear",
                    align_corners=False,
                )
                consolidated_pred_masks[obj_idx : obj_idx + 1] = resized_obj_mask
            consolidated_out["obj_ptr"][obj_idx : obj_idx + 1] = out["obj_ptr"]

        # Optionally, apply non-overlapping constraints on the consolidated scores
        # and rerun the memory encoder
        if run_mem_encoder:
            device = inference_state["device"]
            high_res_masks = torch.nn.functional.interpolate(
                consolidated_out["pred_masks"].to(device, non_blocking=True),
                size=(self.image_size, self.image_size),
                mode="bilinear",
                align_corners=False,
            )
            if self.non_overlap_masks_for_mem_enc:
                high_res_masks = self._apply_non_overlapping_constraints(high_res_masks)
            maskmem_features, maskmem_pos_enc = self._run_memory_encoder(
                inference_state=inference_state,
                frame_idx=frame_idx,
                batch_size=batch_size,
                high_res_masks=high_res_masks,
                is_mask_from_pts=True,  # these frames are what the user interacted with
            )
            consolidated_out["maskmem_features"] = maskmem_features
            consolidated_out["maskmem_pos_enc"] = maskmem_pos_enc

        return consolidated_out


    def _get_orig_video_res_output(self, inference_state, any_res_masks):
        """
        Resize the object scores to the original video resolution (video_res_masks)
        and apply non-overlapping constraints for final output.
        """
        device = inference_state["device"]
        video_H = inference_state["video_height"]
        video_W = inference_state["video_width"]
        any_res_masks = any_res_masks.to(device, non_blocking=True)
        if any_res_masks.shape[-2:] == (video_H, video_W):
            video_res_masks = any_res_masks
        else:
            video_res_masks = torch.nn.functional.interpolate(
                any_res_masks,
                size=(video_H, video_W),
                mode="bilinear",
                align_corners=False,
            )
        if self.non_overlap_masks:
            video_res_masks = self._apply_non_overlapping_constraints(video_res_masks)
        return any_res_masks, video_res_masks

    def init_state(
        self,
        images
    ):
        """Initialize a inference state."""
        inference_state = {}
        inference_state["images"] = images
        inference_state["num_frames"] = len(images)
        # whether to offload the video frames to CPU memory
        # turning on this option saves the GPU memory with only a very small overhead
        inference_state["offload_video_to_cpu"] = False
        # whether to offload the inference state to CPU memory
        # turning on this option saves the GPU memory at the cost of a lower tracking fps
        # (e.g. in a test case of 768x768 model, fps dropped from 27 to 24 when tracking one object
        # and from 24 to 21 when tracking two objects)
        inference_state["offload_state_to_cpu"] = False
        # the original video height and width, used for resizing final output scores
        inference_state["video_height"] = self.image_size
        inference_state["video_width"] = self.image_size
        inference_state["device"] = torch.device("cuda")
        inference_state["storage_device"] = torch.device("cuda")
        # inputs on each frame
        inference_state["point_inputs_per_obj"] = {}
        inference_state["mask_inputs_per_obj"] = {}
        # visual features on a small number of recently visited frames for quick interactions
        inference_state["cached_features"] = {}
        # values that don't change across frames (so we only need to hold one copy of them)
        inference_state["constants"] = {}
        # mapping between client-side object id and model-side object index
        inference_state["obj_id_to_idx"] = OrderedDict()
        inference_state["obj_idx_to_id"] = OrderedDict()
        inference_state["obj_ids"] = []
        # A storage to hold the model's tracking results and states on each frame
        inference_state["output_dict"] = {
            "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
            "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
        }
        # Slice (view) of each object tracking results, sharing the same memory with "output_dict"
        inference_state["output_dict_per_obj"] = {}
        # A temporary storage to hold new outputs when user interact with a frame
        # to add clicks or mask (it's merged into "output_dict" before propagation starts)
        inference_state["temp_output_dict_per_obj"] = {}
        # Frames that already holds consolidated outputs from click or mask inputs
        # (we directly use their consolidated outputs during tracking)
        inference_state["consolidated_frame_inds"] = {
            "cond_frame_outputs": set(),  # set containing frame indices
            "non_cond_frame_outputs": set(),  # set containing frame indices
        }
        # metadata for each tracking frame (e.g. which direction it's tracked)
        inference_state["tracking_has_started"] = False
        inference_state["frames_already_tracked"] = {}
        return inference_state

    def add_language_embd(
            self,
            inference_state,
            frame_idx,
            obj_id,
            language_embd,
            inference=False,
    ):
        obj_idx = _obj_id_to_idx(inference_state, obj_id)

        is_init_cond_frame = frame_idx not in inference_state["frames_already_tracked"]
        # whether to track in reverse time order
        if is_init_cond_frame:
            reverse = False
        else:
            reverse = inference_state["frames_already_tracked"][frame_idx]["reverse"]

        obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
        obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
        # Add a frame to conditioning output if it's an initial conditioning frame or
        # if the model sees all frames receiving clicks/mask as conditioning frames.
        is_cond = is_init_cond_frame or self.add_all_frames_to_correct_as_cond
        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"

        # Get any previously predicted mask logits on this object and feed it along with
        # the new clicks into the SAM mask decoder.
        prev_sam_mask_logits = None
        # lookup temporary output dict first, which contains the most recent output
        # (if not found, then lookup conditioning and non-conditioning frame output)
        prev_out = obj_temp_output_dict[storage_key].get(frame_idx)
        if prev_out is None:
            prev_out = obj_output_dict["cond_frame_outputs"].get(frame_idx)
            if prev_out is None:
                prev_out = obj_output_dict["non_cond_frame_outputs"].get(frame_idx)

        if prev_out is not None and prev_out["pred_masks"] is not None:
            prev_sam_mask_logits = prev_out["pred_masks"].cuda(non_blocking=True)
            # Clamp the scale of prev_sam_mask_logits to avoid rare numerical issues.
            prev_sam_mask_logits = torch.clamp(prev_sam_mask_logits, -32.0, 32.0)

        current_out, pred_mask_gpu = self._run_single_frame_inference(
            inference_state=inference_state,
            output_dict=obj_output_dict,  # run on the slice of a single object
            frame_idx=frame_idx,
            batch_size=1,  # run on the slice of a single object
            is_init_cond_frame=is_init_cond_frame,
            point_inputs=None,
            mask_inputs=None,
            reverse=reverse,
            # Skip the memory encoder when adding clicks or mask. We execute the memory encoder
            # at the beginning of `propagate_in_video` (after user finalize their clicks). This
            # allows us to enforce non-overlapping constraints on all objects before encoding
            # them into memory.
            run_mem_encoder=False,
            prev_sam_mask_logits=prev_sam_mask_logits,
            ## Extension: LLM prompt
            language_embd=language_embd,
        )
        # Add the output to the output dict (to be used as future memory)
        obj_temp_output_dict[storage_key][frame_idx] = current_out

        # Resize the output mask to the original video resolution
        obj_ids = inference_state["obj_ids"]
        if inference:
            _consolidated_out = self._consolidate_temp_output_across_obj(
                inference_state,
                frame_idx,
                is_cond=is_cond,
                run_mem_encoder=False,
                consolidate_at_video_res=False,
            )
            # _, video_res_masks = self._get_orig_video_res_output(
            #     inference_state, consolidated_out["pred_masks_video_res"]
            # )
        return frame_idx, obj_ids, pred_mask_gpu


    def _clear_non_cond_mem_around_input(self, inference_state, frame_idx):
        """
        Remove the non-conditioning memory around the input frame. When users provide
        correction clicks, the surrounding frames' non-conditioning memories can still
        contain outdated object appearance information and could confuse the model.

        This method clears those non-conditioning memories surrounding the interacted
        frame to avoid giving the model both old and new information about the object.
        """
        r = self.memory_temporal_stride_for_eval
        frame_idx_begin = frame_idx - r * self.num_maskmem
        frame_idx_end = frame_idx + r * self.num_maskmem
        output_dict = inference_state["output_dict"]
        non_cond_frame_outputs = output_dict["non_cond_frame_outputs"]
        for t in range(frame_idx_begin, frame_idx_end + 1):
            non_cond_frame_outputs.pop(t, None)
            for obj_output_dict in inference_state["output_dict_per_obj"].values():
                obj_output_dict["non_cond_frame_outputs"].pop(t, None)

    def _run_memory_encoder(
        self, inference_state, frame_idx, batch_size, high_res_masks, is_mask_from_pts
    ):
        """
        Run the memory encoder on `high_res_masks`. This is usually after applying
        non-overlapping constraints to object scores. Since their scores changed, their
        memory also need to be computed again with the memory encoder.
        """
        # Retrieve correct image features
        _, _, current_vision_feats, _, feat_sizes = self._get_image_feature(
            inference_state, frame_idx, batch_size
        )
        maskmem_features, maskmem_pos_enc = self._encode_new_memory(
            current_vision_feats=current_vision_feats,
            feat_sizes=feat_sizes,
            pred_masks_high_res=high_res_masks,
            is_mask_from_pts=is_mask_from_pts,
        )

        # optionally offload the output to CPU memory to save GPU space
        storage_device = inference_state["storage_device"]
        maskmem_features = maskmem_features.to(torch.bfloat16)
        maskmem_features = maskmem_features.to(storage_device, non_blocking=True)
        # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
        maskmem_pos_enc = _get_maskmem_pos_enc(
            inference_state, {"maskmem_pos_enc": maskmem_pos_enc}
        )
        return maskmem_features, maskmem_pos_enc

    def _add_output_per_object(
        self, inference_state, frame_idx, current_out, storage_key
    ):
        """
        Split a multi-object output into per-object output slices and add them into
        `output_dict_per_obj`. The resulting slices share the same tensor storage.
        """
        maskmem_features = current_out["maskmem_features"]
        assert maskmem_features is None or isinstance(maskmem_features, torch.Tensor)

        maskmem_pos_enc = current_out["maskmem_pos_enc"]
        assert maskmem_pos_enc is None or isinstance(maskmem_pos_enc, list)

        output_dict_per_obj = inference_state["output_dict_per_obj"]
        for obj_idx, obj_output_dict in output_dict_per_obj.items():
            obj_slice = slice(obj_idx, obj_idx + 1)
            obj_out = {
                "maskmem_features": None,
                "maskmem_pos_enc": None,
                "pred_masks": current_out["pred_masks"][obj_slice],
                "obj_ptr": current_out["obj_ptr"][obj_slice],
            }
            if maskmem_features is not None:
                obj_out["maskmem_features"] = maskmem_features[obj_slice]
            if maskmem_pos_enc is not None:
                obj_out["maskmem_pos_enc"] = [x[obj_slice] for x in maskmem_pos_enc]
            obj_output_dict[storage_key][frame_idx] = obj_out

    @torch.inference_mode()
    def propagate_in_video_preflight(self, inference_state):
        """Prepare inference_state and consolidate temporary outputs before tracking."""
        # Tracking has started and we don't allow adding new objects until session is reset.
        inference_state["tracking_has_started"] = True
        batch_size = _get_obj_num(inference_state)

        # Consolidate per-object temporary outputs in "temp_output_dict_per_obj" and
        # add them into "output_dict".
        temp_output_dict_per_obj = inference_state["temp_output_dict_per_obj"]
        output_dict = inference_state["output_dict"]
        # "consolidated_frame_inds" contains indices of those frames where consolidated
        # temporary outputs have been added (either in this call or any previous calls
        # to `propagate_in_video_preflight`).
        consolidated_frame_inds = inference_state["consolidated_frame_inds"]
        for is_cond in [False, True]:
            # Separately consolidate conditioning and non-conditioning temp outptus
            storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
            # Find all the frames that contain temporary outputs for any objects
            # (these should be the frames that have just received clicks for mask inputs
            # via `add_new_points` or `add_new_mask`)
            temp_frame_inds = set()
            for obj_temp_output_dict in temp_output_dict_per_obj.values():
                temp_frame_inds.update(obj_temp_output_dict[storage_key].keys())
            consolidated_frame_inds[storage_key].update(temp_frame_inds)
            # consolidate the temprary output across all objects on this frame
            for frame_idx in temp_frame_inds:
                consolidated_out = self._consolidate_temp_output_across_obj(
                    inference_state, frame_idx, is_cond=is_cond, run_mem_encoder=True
                )
                # merge them into "output_dict" and also create per-object slices
                output_dict[storage_key][frame_idx] = consolidated_out
                self._add_output_per_object(
                    inference_state, frame_idx, consolidated_out, storage_key
                )
                clear_non_cond_mem = self.clear_non_cond_mem_around_input and (
                    self.clear_non_cond_mem_for_multi_obj or batch_size <= 1
                )
                if clear_non_cond_mem:
                    # clear non-conditioning memory of the surrounding frames
                    self._clear_non_cond_mem_around_input(inference_state, frame_idx)

            # clear temporary outputs in `temp_output_dict_per_obj`
            for obj_temp_output_dict in temp_output_dict_per_obj.values():
                obj_temp_output_dict[storage_key].clear()

        # edge case: if an output is added to "cond_frame_outputs", we remove any prior
        # output on the same frame in "non_cond_frame_outputs"
        for frame_idx in output_dict["cond_frame_outputs"]:
            output_dict["non_cond_frame_outputs"].pop(frame_idx, None)
        for obj_output_dict in inference_state["output_dict_per_obj"].values():
            for frame_idx in obj_output_dict["cond_frame_outputs"]:
                obj_output_dict["non_cond_frame_outputs"].pop(frame_idx, None)
        for frame_idx in consolidated_frame_inds["cond_frame_outputs"]:
            assert frame_idx in output_dict["cond_frame_outputs"]
            consolidated_frame_inds["non_cond_frame_outputs"].discard(frame_idx)

        # Make sure that the frame indices in "consolidated_frame_inds" are exactly those frames
        # with either points or mask inputs (which should be true under a correct workflow).
        all_consolidated_frame_inds = (
            consolidated_frame_inds["cond_frame_outputs"]
            | consolidated_frame_inds["non_cond_frame_outputs"]
        )
        input_frames_inds = set()
        for point_inputs_per_frame in inference_state["point_inputs_per_obj"].values():
            input_frames_inds.update(point_inputs_per_frame.keys())
        for mask_inputs_per_frame in inference_state["mask_inputs_per_obj"].values():
            input_frames_inds.update(mask_inputs_per_frame.keys())

        # with language embd as input, there may not be point or box
        # assert all_consolidated_frame_inds == input_frames_inds

    @torch.inference_mode()
    def propagate_in_video(
        self,
        inference_state,
        start_frame_idx=None,
        max_frame_num_to_track=None,
        reverse=False,
    ):
        """Propagate the input points across frames to track in the entire video."""
        self.propagate_in_video_preflight(inference_state)

        output_dict = inference_state["output_dict"]
        consolidated_frame_inds = inference_state["consolidated_frame_inds"]
        obj_ids = inference_state["obj_ids"]
        num_frames = inference_state["num_frames"]
        batch_size = _get_obj_num(inference_state)
        if len(output_dict["cond_frame_outputs"]) == 0:
            raise RuntimeError("No points are provided; please add points first")
        clear_non_cond_mem = self.clear_non_cond_mem_around_input and (
            self.clear_non_cond_mem_for_multi_obj or batch_size <= 1
        )

        # set start index, end index, and processing order
        if start_frame_idx is None:
            # default: start from the earliest frame with input points
            start_frame_idx = min(output_dict["cond_frame_outputs"])
        if max_frame_num_to_track is None:
            # default: track all the frames in the video
            max_frame_num_to_track = num_frames
        if reverse:
            end_frame_idx = max(start_frame_idx - max_frame_num_to_track, 0)
            if start_frame_idx > 0:
                processing_order = range(start_frame_idx, end_frame_idx - 1, -1)
            else:
                processing_order = []  # skip reverse tracking if starting from frame 0
        else:
            end_frame_idx = min(
                start_frame_idx + max_frame_num_to_track, num_frames - 1
            )
            processing_order = range(start_frame_idx, end_frame_idx + 1)

        for frame_idx in tqdm(processing_order, desc="propagate in video"):
            # We skip those frames already in consolidated outputs (these are frames
            # that received input clicks or mask). Note that we cannot directly run
            # batched forward on them via `_run_single_frame_inference` because the
            # number of clicks on each object might be different.
            if frame_idx in consolidated_frame_inds["cond_frame_outputs"]:
                storage_key = "cond_frame_outputs"
                current_out = output_dict[storage_key][frame_idx]
                pred_masks = current_out["pred_masks"]
                if clear_non_cond_mem:
                    # clear non-conditioning memory of the surrounding frames
                    self._clear_non_cond_mem_around_input(inference_state, frame_idx)
            elif frame_idx in consolidated_frame_inds["non_cond_frame_outputs"]:
                storage_key = "non_cond_frame_outputs"
                current_out = output_dict[storage_key][frame_idx]
                pred_masks = current_out["pred_masks"]
            else:
                storage_key = "non_cond_frame_outputs"
                current_out, pred_masks = self._run_single_frame_inference(
                    inference_state=inference_state,
                    output_dict=output_dict,
                    frame_idx=frame_idx,
                    batch_size=batch_size,
                    is_init_cond_frame=False,
                    point_inputs=None,
                    mask_inputs=None,
                    reverse=reverse,
                    run_mem_encoder=True,
                )
                output_dict[storage_key][frame_idx] = current_out
            # Create slices of per-object outputs for subsequent interaction with each
            # individual object after tracking.
            self._add_output_per_object(
                inference_state, frame_idx, current_out, storage_key
            )
            inference_state["frames_already_tracked"][frame_idx] = {"reverse": reverse}

            # Resize the output mask to the original video resolution (we directly use
            # the mask scores on GPU for output to avoid any CPU conversion in between)
            _, video_res_masks = self._get_orig_video_res_output(
                inference_state, pred_masks
            )
            yield frame_idx, obj_ids, video_res_masks

def fill_holes_in_mask_scores(mask, max_area):
    """
    A post processor to fill small holes in mask scores with area under `max_area`.
    """
    # Holes are those connected components in background with area <= self.max_area
    # (background regions are those with mask scores <= 0)
    assert max_area > 0, "max_area must be positive"
    labels, areas = get_connected_components(mask <= 0)
    is_hole = (labels > 0) & (areas <= max_area)
    # We fill holes with a small positive mask score (0.1) to change them to foreground.
    mask = torch.where(is_hole, 0.1, mask)
    return mask

def get_connected_components(mask):
    """
    Get the connected components (8-connectivity) of binary masks of shape (N, 1, H, W).

    Inputs:
    - mask: A binary mask tensor of shape (N, 1, H, W), where 1 is foreground and 0 is
            background.

    Outputs:
    - labels: A tensor of shape (N, 1, H, W) containing the connected component labels
              for foreground pixels and 0 for background pixels.
    - counts: A tensor of shape (N, 1, H, W) containing the area of the connected
              components for foreground pixels and 0 for background pixels.
    """
    from torch.utils.cpp_extension import load
    os.system("wget https://github.com/facebookresearch/sam2/blob/main/sam2/csrc/connected_components.cu")
    get_connected_componnets = load(
        name="get_connected_componnets",
        sources=["./connected_components.cu"],
        verbose=True,
        extra_cuda_cflags=[
            "-DCUDA_HAS_FP16=1",
            "-D__CUDA_NO_HALF_OPERATORS__",
            "-D__CUDA_NO_HALF_CONVERSIONS__",
            "-D__CUDA_NO_HALF2_OPERATORS__",
        ]
    )

    return get_connected_componnets.get_connected_componnets(mask.to(torch.uint8).contiguous())