Merge pull request #20 from daniel-code/feat/pytorch_models

Feat/pytorch models

Author: daniel-code

README.md

@@ -87,17 +87,19 @@ python train.py -r "datasets/final/train"
 python train.py -r "datasets/final/train" --user-pretrained-weight --finetune-last-layer --use-lr-scheduler --use-auto-augment
 ```
 
-- Training with different model types. See more details in `scripts/different_models.sh`
+- Training with different model types. See more details in `scripts/different_models.sh`.
+  Support [pytorch built-in model types](https://pytorch.org/vision/main/models.html#classification).
 
 ```commandline
 python train.py -r "datasets/final/train" --model-type resnext50_32x4d
 ```
 
-Support model types:
+- Training with different image size. Some model has image resolution constraint, e.g. vit, only accept image size by (
+  244, 244).
 
-- ResNet: resnet18, resnet34, resnet_50, resnet_101
-- ResNext: resnext50_32x4d, resnext101_32x8d
-- Swin: swin_t, swin_s, swin_b
+```commandline
+python train.py -r "datasets/final/train" --model-type vit_b_16  --image-size 224 224
+```
 
 After training, the model weight will export to `model_weights/<model-type>_<exp_time>`.
 Use `tensorboard --logdir model_weights` to browse training log.
@@ -159,7 +161,8 @@ Options:
 python analysis.py -r "datasets/final/train" --model-path "model_weights/<model-type>_<exp_time>/model.pt"
 ```
 
-By default, the `reports/test.png` is AUC of ROC curve and confusion matrix, and the `reports/test_images.jpg` shows the fail cases.
+By default, the `reports/test.png` is AUC of ROC curve and confusion matrix, and the `reports/test_images.jpg` shows the
+fail cases.
 
 ## Inference
 

dogs_cats_classifier/models/__init__.py

@@ -1,5 +1,4 @@
-from .resnet import ResNet
-from .resnext import ResNext
-from .swin import Swin
+from .base import ModelBase
+from .torch_model_wrapper import is_torch_builtin_models, TorchModelWrapper
 
-__all__ = ['ResNet', 'Swin', 'ResNext']
+__all__ = ['ModelBase', 'TorchModelWrapper', 'is_torch_builtin_models']

dogs_cats_classifier/models/base.py

@@ -26,8 +26,6 @@ def __init__(self,
         self.user_pretrained_weight = user_pretrained_weight
         self.finetune_last_layer = finetune_last_layer
 
-        self.models_mapping = self._setup_models_mapping()
-        assert model_type in self.models_mapping, f'{model_type} is not available. There is available model types: {list(self.models_mapping.keys())}'
         self.model = self._setup_model(model_type=model_type)
 
         if finetune_last_layer:
@@ -39,9 +37,6 @@ def __init__(self,
 
         self.example_input_array = torch.zeros((1, 3, input_shape[0], input_shape[1]), dtype=torch.float32)
 
-    def _setup_models_mapping(self) -> dict:
-        raise NotImplementedError
-
     def _setup_model(self, model_type) -> torch.nn.Module:
         raise NotImplementedError
 

dogs_cats_classifier/models/resnet.py

@@ -0,0 +1,40 @@
+import types
+
+from .base import ModelBase
+from torch.nn import Module, Sequential, Linear
+
+
+def is_torch_builtin_models(model_type: str):
+    model_type = model_type.lower()
+    model_module = __import__('torchvision.models', fromlist=(model_type, ), level=0)
+    if model_type in dir(model_module):
+        model_func = getattr(model_module, model_type)
+        return True if isinstance(model_func, types.FunctionType) else False
+
+    return False
+
+
+class TorchModelWrapper(ModelBase):
+    def _setup_model(self, model_type) -> Module:
+        # dynamic import module
+        model_module = __import__('torchvision.models', fromlist=(model_type, ), level=0)
+        model_func = getattr(model_module, model_type)
+        model = model_func(weights='DEFAULT' if self.user_pretrained_weight else None)
+
+        # get last layer's name
+        layer_name = list(model.named_children())[-1][0]
+
+        # check last layer
+        last_layer = model.get_submodule(layer_name)
+        if isinstance(last_layer, Sequential):
+            last_layer = last_layer[-1]
+        in_features = last_layer.in_features
+
+        # replace the last layer
+        last_layer = getattr(model, layer_name)
+        if isinstance(last_layer, Sequential):
+            last_layer[-1] = Linear(in_features=in_features, out_features=self.num_classes)
+        else:
+            setattr(model, layer_name, Linear(in_features=in_features, out_features=self.num_classes))
+
+        return model

dogs_cats_classifier/models/resnext.py

@@ -3,7 +3,7 @@
 setup(
     name='dogs_cats_classifier',
     packages=find_packages(),
-    version='0.1.0',
+    version='0.1.1',
     description='Create an algorithm to distinguish dogs from cats',
     author='YanRu',
     license="MIT",

dogs_cats_classifier/models/swin.py

@@ -7,7 +7,7 @@
 from pytorch_lightning.callbacks import LearningRateMonitor
 
 from dogs_cats_classifier.data import DogsCatsImagesDataModule
-from dogs_cats_classifier.models import ResNet, Swin, ResNext
+from dogs_cats_classifier.models import TorchModelWrapper, is_torch_builtin_models
 from dogs_cats_classifier.utils import Evaluator
 from datetime import datetime
 
@@ -78,16 +78,12 @@ def main(batch_size, max_epochs, num_workers, image_size, dataset_root, fast_dev
     print(dogs_cats_datamodule)
 
     # prepare model
-    if 'swin' in model_type:
-        model = Swin
-    elif 'resnext' in model_type:
-        model = ResNext
-    elif 'resnet' in model_type:
-        model = ResNet
+    if is_torch_builtin_models(model_type):
+        model_class = TorchModelWrapper
     else:
         raise ValueError(f'{model_type} is not available.')
 
-    model = model(
+    model = model_class(
         num_classes=1,
         model_type=model_type,
         input_shape=image_size,
@@ -117,9 +113,10 @@ def main(batch_size, max_epochs, num_workers, image_size, dataset_root, fast_dev
     torch.jit.save(script_model, os.path.join(output_path, 'model.pt'))
 
     # evaluation
-    dogs_cats_datamodule.setup()
-    evaluator = Evaluator(model=model, output_path=output_path)
-    evaluator.evaluate(dataloader=dogs_cats_datamodule.test_dataloader(), title=f'{model_type}_test', verbose=False)
+    if not fast_dev_run:
+        dogs_cats_datamodule.setup()
+        evaluator = Evaluator(model=model, output_path=output_path)
+        evaluator.evaluate(dataloader=dogs_cats_datamodule.test_dataloader(), title=f'{model_type}_test', verbose=False)
 
 
 if __name__ == '__main__':