add swin_transformer

trojanvision/models/torchvision/__init__.py

@@ -9,10 +9,12 @@
 from .mobilenet import MobileNet
 from .resnet import ResNet
 from .shufflenetv2 import ShuffleNetV2
+from .swin_transformer import SwinTransformer
 from .vgg import VGG
 
 __all__ = ['AlexNet', 'DenseNet', 'EfficientNet', 'MNASNet',
-           'MobileNet', 'ResNet', 'ShuffleNetV2', 'VGG']
+           'MobileNet', 'ResNet', 'ShuffleNetV2', 'SwinTransformer',
+           'VGG']
 
 class_dict: dict[str, type[ImageModel]] = {
     'alexnet': AlexNet,
@@ -23,5 +25,6 @@
     'resnet': ResNet,
     'resnext': ResNet,
     'shufflenetv2': ShuffleNetV2,
+    'swin': SwinTransformer,
     'vgg': VGG,
 }

trojanvision/models/torchvision/swin_transformer.py

@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+
+from trojanvision.models.imagemodel import _ImageModel, ImageModel
+
+import torch
+import torch.nn as nn
+import torchvision.models
+from torchvision.models.swin_transformer import (Swin_T_Weights, Swin_S_Weights, Swin_B_Weights,
+                                                 Swin_V2_T_Weights, Swin_V2_S_Weights, Swin_V2_B_Weights)
+from collections import OrderedDict
+
+from torchvision.models.swin_transformer import _swin_transformer, SwinTransformerBlockV2, PatchMergingV2
+from typing import Optional, Any
+
+
+class _SwinTransformer(_ImageModel):
+
+    def __init__(self, name: str = 'swin_v2_t', **kwargs):
+        super().__init__(**kwargs)
+        if 'comp' in name:
+            ModelClass = eval(name)
+        else:
+            ModelClass = getattr(torchvision.models, name)
+        _model: torchvision.models.SwinTransformer = ModelClass(num_classes=self.num_classes)
+
+        self.features = _model.features
+        self.features.add_module('norm', _model.norm)
+        self.features.add_module('permute', _model.permute)
+        self.classifier = nn.Sequential(OrderedDict([
+            ('fc', _model.head),
+        ]))
+
+
+class SwinTransformer(ImageModel):
+    available_models = {'swin_t', 'swin_s', 'swin_b',
+                        'swin_v2_t', 'swin_v2_s', 'swin_v2_b',
+                        'swin_t_comp', 'swin_v2_t_comp',
+                        }
+    weights = {
+        'swin_t': Swin_T_Weights,
+        'swin_s': Swin_S_Weights,
+        'swin_b': Swin_B_Weights,
+        'swin_v2_t': Swin_V2_T_Weights,
+        'swin_v2_s': Swin_V2_S_Weights,
+        'swin_v2_b': Swin_V2_B_Weights,
+    }
+
+    def __init__(self, name: str = 'swin_v2_t',
+                 model: type[_SwinTransformer] = _SwinTransformer, **kwargs):
+        super().__init__(name=name, model=model, **kwargs)
+
+    def get_official_weights(self, **kwargs) -> OrderedDict[str, torch.Tensor]:
+        _dict = super().get_official_weights(**kwargs)
+        _dict['features.norm.weight'] = _dict['norm.weight']
+        _dict['features.norm.bias'] = _dict['norm.bias']
+        del _dict['norm.weight']
+        del _dict['norm.bias']
+        _dict['classifier.fc.weight'] = _dict['head.weight']
+        _dict['classifier.fc.bias'] = _dict['head.bias']
+        del _dict['head.weight']
+        del _dict['head.bias']
+        return _dict
+
+
+def swin_t_comp(*, weights: Optional[Swin_T_Weights] = None, progress: bool = True, **kwargs: Any) -> SwinTransformer:
+    weights = Swin_T_Weights.verify(weights)
+
+    return _swin_transformer(
+        patch_size=[2, 2],
+        embed_dim=96,
+        depths=[2, 2, 2, 2],
+        num_heads=[3, 6, 12, 24],
+        window_size=[4, 4],
+        stochastic_depth_prob=0.2,
+        weights=weights,
+        progress=progress,
+        **kwargs,
+    )
+
+
+def swin_v2_t_comp(*, weights: Optional[Swin_V2_T_Weights] = None, progress: bool = True, **kwargs: Any) -> SwinTransformer:
+    weights = Swin_V2_T_Weights.verify(weights)
+
+    return _swin_transformer(
+        patch_size=[2, 2],
+        embed_dim=96,
+        depths=[2, 2],
+        num_heads=[3, 6, 12, 24],
+        window_size=[4, 4],
+        stochastic_depth_prob=0.2,
+        weights=weights,
+        progress=progress,
+        block=SwinTransformerBlockV2,
+        downsample_layer=PatchMergingV2,
+        **kwargs,
+    )