环境

• 环境安装        

• pip install openmim
  mim install mmengine
  mim install mmcv
  mim install mmpretrain
  # 安装多模态模型
  pip install "mmpretrain[multimodal]"

• 验证环境

• In [1]: import mmengine
  
  In [2]: mmengine.__version__
  Out[2]: '0.7.3'
  
  In [3]: import mmcv
  
  In [4]: mmcv.__version__
  Out[4]: '2.0.0'
  
  In [5]: import mmpretrain
  
  In [6]: mmpretrain.__version__
  Out[6]: '1.0.0rc8'

• 初始应用
  • 导入MMPretrain相关包：模型获取，模型列举，模型推理

  • from mmpretrain import get_model, list_models, inference_model

  • 获取分类任务中resnet18相关模型

  • In [8]: list_models(task="Image Classification", pattern="resnet18")
    Out[8]: ['resnet18_8xb16_cifar10', 'resnet18_8xb32_in1k']

  • 获取模型冰查看模型类型及其对应的backbone

  • In [9]: model = get_model('resnet18_8xb32_in1k')
    
    In [10]: type(model)
    Out[10]: mmpretrain.models.classifiers.image.ImageClassifier
    
    In [11]: type(model.backbone)
    Out[11]: mmpretrain.models.backbones.resnet.ResNet

  • 模型推理

  • inference_model(mode, r"D:\workspace\.git\OpenMMLabCamp\bird.jpg", show=True)

    

     没有加载预训练权重导致上述问题

  • In [2]: list_models(task='Image Caption', pattern='blip')
    Out[2]: ['blip-base_3rdparty_caption', 'blip2-opt2.7b_3rdparty-zeroshot_caption']
    
    In [3]: inference_model('blip-base_3rdparty_caption', './mmpretrain/demo/cat-dog.png', show=True)

    推理结果如下：

  • 

• 训练结果如下：

• 2023/06/06 22:37:47 - mmengine - INFO - Exp name: dog_and_cat_20230606_223441
  2023/06/06 22:37:47 - mmengine - INFO - Saving checkpoint at 5 epochs
  2023/06/06 22:37:49 - mmengine - INFO - Epoch(val) [5][51/51]    accuracy/top1: 95.7527  data_time: 0.0022  time: 0.0252

• 在测试集上测试结果

• .\tools\test.py configs2/dog_and_cat.py --checkpoint work_dirs/dog_and_cat/epoch_5.pth
  

• 样本测试结果存入pkl

• .\tools\test.py mmpretrain configs2/dog_and_cat.py --checkpoint work_dirs/dog_and_cat/epoch_5.pth
  

• 分析测试结果

• .\tools\analysis_tools\analyze_results.py configs2/dog_and_cat.py  rersult.pkl --out-dir annalyze
  
  

• 生成混淆矩阵

• .\tools\analysis_tools\confusion_matrix.py configs2/dog_and_cat.py  rersult.pkl --show --include-values