202 lines
7.2 KiB
Markdown
202 lines
7.2 KiB
Markdown
---
|
||
tasks:
|
||
- image-classification
|
||
model-type:
|
||
- Transformer
|
||
domain:
|
||
- cv
|
||
frameworks:
|
||
- pytorch
|
||
backbone:
|
||
- ViT-Base
|
||
metrics:
|
||
- accuracy
|
||
finetune-support: True
|
||
customized-quickstart: True
|
||
integrating: False
|
||
license: Apache License 2.0
|
||
tags:
|
||
- Alibaba
|
||
- Image classification
|
||
- Transformer
|
||
- Dailylife tags
|
||
datasets:
|
||
evaluation:
|
||
- online dataset
|
||
train:
|
||
- tany0699/dailytags
|
||
test:
|
||
- tany0699/dailytags
|
||
|
||
widgets:
|
||
- task: image-classification
|
||
inputs:
|
||
- type: image
|
||
examples:
|
||
- name: 1
|
||
inputs:
|
||
- name: image
|
||
data: git://resources/test.jpg
|
||
---
|
||
|
||
# 日常物体识别模型介绍
|
||
自建1300类常见物体标签体系，覆盖常见的日用品，动物，植物，家具，设备，食物等物体，标签从海量中文互联网社区语料进行提取，保留了出现频率较高的常见物体名称。模型结构采用最新的ViT-Base结构。
|
||
创空间快速可视化展示: [ViT图像分类-中文-日常物品](https://modelscope.cn/studios/tany0699/cv_vit-base_image-classification_Dailylife-labels/summary)
|
||
|
||
本系列还有如下模型，欢迎试用:
|
||
- [ViT图像分类-通用](https://modelscope.cn/models/damo/cv_vit-base_image-classification_ImageNet-labels/summary)
|
||
- [NextViT实时图像分类-中文-日常物品](https://modelscope.cn/models/damo/cv_nextvit-small_image-classification_Dailylife-labels/summary)
|
||
- [ConvNeXt图像分类-中文-垃圾分类](https://modelscope.cn/models/damo/cv_convnext-base_image-classification_garbage/summary)
|
||
- [BEiTv2图像分类-通用-base](https://modelscope.cn/models/damo/cv_beitv2-base_image-classification_patch16_224_pt1k_ft22k_in1k/summary)
|
||
- [BEiTv2图像分类-通用-large](https://modelscope.cn/models/damo/cv_beitv2-large_image-classification_patch16_224_pt1k_ft22k_in1k/summary)
|
||
|
||
## 模型描述
|
||
|
||
采用Transformer经典的[ViT-Base](https://github.com/google-research/vision_transformer)结构, 并采用了DeiT的知识蒸馏方式进行训练。
|
||
<img src="./resources/overview.jpg" alt="overview"/>
|
||
|
||
## 期望模型使用方式以及适用范围
|
||
|
||
本模型适用范围较广，覆盖大部分日常生活常见的物品类目，包括日用品，动物，植物，家具，设备，食物等。也可作为下游任务的预训练backbone。
|
||
|
||
### 如何使用
|
||
|
||
在ModelScope框架上，提供输入图片，即可通过简单的Pipeline调用来使用。
|
||
|
||
#### 代码范例
|
||
```python
|
||
from modelscope.pipelines import pipeline
|
||
from modelscope.utils.constant import Tasks
|
||
|
||
img_path = 'https://modelscope.oss-cn-beijing.aliyuncs.com/test/images/bird.JPEG'
|
||
image_classification = pipeline(Tasks.image_classification,
|
||
model='damo/cv_vit-base_image-classification_Dailylife-labels')
|
||
result = image_classification(img_path)
|
||
print(result)
|
||
```
|
||
|
||
### 模型局限性以及可能的偏差
|
||
|
||
- 支持1300类常见物体识别
|
||
|
||
|
||
## 训练数据介绍
|
||
|
||
- 140万包含常见物体的图像集
|
||
|
||
|
||
## 模型训练流程
|
||
|
||
- 主要训练参数参考[DeiT论文](https://arxiv.org/abs/2012.12877)的设置，除了weight decay在复现时设置为0.1，模型训练未使用pretrained参数进行初始化。
|
||
|
||
### 预处理
|
||
|
||
测试时主要的预处理如下：
|
||
- Resize：先将原始图片的短边缩放至256
|
||
- Normalize：图像归一化，减均值除以标准差
|
||
- CenterCrop：裁切为224x224
|
||
|
||
## 数据评估及结果
|
||
|
||
模型在自建测试集进行测试，结果如下:
|
||
|
||
| Model | top-1 acc | top-5 acc | #params | Remark |
|
||
|:--------:|:-------:|:--------:|:-------:|--------------|
|
||
| ViT-base | 74.3 | 95.3 | 86M | modelscope |
|
||
|
||
|
||
## 模型训练
|
||
使用托管在modelscope DatasetHub上的小型数据集[mini_imagenet100](https://modelscope.cn/datasets/tany0699/mini_imagenet100/summary)进行finetune训练的示例代码:
|
||
|
||
```python
|
||
from modelscope.msdatasets import MsDataset
|
||
from modelscope.metainfo import Trainers
|
||
from modelscope.trainers import build_trainer
|
||
import tempfile
|
||
|
||
model_id = 'damo/cv_vit-base_image-classification_Dailylife-labels'
|
||
|
||
# 加载数据
|
||
ms_train_dataset = MsDataset.load(
|
||
'mini_imagenet100', namespace='tany0699',
|
||
subset_name='default', split='train') # 加载训练集
|
||
|
||
ms_val_dataset = MsDataset.load(
|
||
'mini_imagenet100', namespace='tany0699',
|
||
subset_name='default', split='validation') # 加载验证集
|
||
|
||
tmp_dir = tempfile.TemporaryDirectory().name # 使用临时目录作为工作目录
|
||
|
||
# 修改配置文件
|
||
def cfg_modify_fn(cfg):
|
||
cfg.train.dataloader.batch_size_per_gpu = 32 # batch大小
|
||
cfg.train.dataloader.workers_per_gpu = 2 # 每个gpu的worker数目
|
||
cfg.train.max_epochs = 1 # 最大训练epoch数
|
||
cfg.model.mm_model.head.num_classes = 100 # 分类数
|
||
cfg.model.mm_model.train_cfg.augments[0].num_classes = 100 # 分类数
|
||
cfg.model.mm_model.train_cfg.augments[1].num_classes = 100 # 分类数
|
||
cfg.train.optimizer.lr = 1e-4 # 学习率
|
||
cfg.train.lr_config.warmup_iters = 1 # 预热次数
|
||
cfg.train.evaluation.metric_options = {'topk': (1, 5)} # 训练时的评估指标
|
||
cfg.evaluation.metric_options = {'topk': (1, 5)} # 评估时的评估指标
|
||
return cfg
|
||
|
||
# 构建训练器
|
||
kwargs = dict(
|
||
model=model_id, # 模型id
|
||
work_dir=tmp_dir, # 工作目录
|
||
train_dataset=ms_train_dataset, # 训练集
|
||
eval_dataset=ms_val_dataset, # 验证集
|
||
cfg_modify_fn=cfg_modify_fn, # 用于修改训练配置文件的回调函数
|
||
model_revision='v1.0.2'
|
||
)
|
||
trainer = build_trainer(name=Trainers.image_classification, default_args=kwargs)
|
||
|
||
# 进行训练
|
||
trainer.train()
|
||
|
||
# 进行评估
|
||
result = trainer.evaluate()
|
||
print('result:', result)
|
||
```
|
||
训练说明见示例代码中的注释，更详细的训练说明和用法见官方的[训练文档](https://modelscope.cn/docs/%E6%A8%A1%E5%9E%8B%E7%9A%84%E8%AE%AD%E7%BB%83Train)。训练过程产生的log和模型权重文件保存在work_dir工作目录中，并以前缀为'best_'的文件保存了在验证集上最优精度的权重。evaluate()默认使用精度最高的模型权重进行评估。
|
||
|
||
|
||
## 模型评估
|
||
使用训练好的模型对需要评估的数据集进行精度评估示例代码如下:
|
||
|
||
```python
|
||
from modelscope.msdatasets import MsDataset
|
||
from modelscope.metainfo import Trainers
|
||
from modelscope.trainers import build_trainer
|
||
import tempfile
|
||
|
||
model_id = 'damo/cv_vit-base_image-classification_Dailylife-labels'
|
||
|
||
# 加载用于评估的数据集
|
||
ms_val_dataset = MsDataset.load(
|
||
'dailytags', namespace='tany0699',
|
||
subset_name='default', split='validation')
|
||
|
||
tmp_dir = tempfile.TemporaryDirectory().name # 使用临时目录作为工作目录
|
||
|
||
# 构建训练器
|
||
kwargs = dict(
|
||
model=model_id, # 模型id
|
||
work_dir=tmp_dir, # 工作目录
|
||
train_dataset=None,
|
||
eval_dataset=ms_val_dataset, # 评估的数据集
|
||
model_revision='v1.0.2'
|
||
)
|
||
trainer = build_trainer(name=Trainers.image_classification, default_args=kwargs)
|
||
|
||
# 开始评估
|
||
result = trainer.evaluate()
|
||
print('result:', result)
|
||
```
|
||
评估过程默认使用模型中自带的预训练权重进行评估。
|
||
|
||
#### Clone with HTTP
|
||
```bash
|
||
git clone https://www.modelscope.cn/damo/cv_vit-base_image-classification_Dailylife-labels.git
|
||
``` |