| import torch.nn as nn |
| import torch.nn.functional as F |
| import torchvision.models as models |
| import torch |
| from einops import rearrange |
|
|
|
|
| class ConvBlock(nn.Module): |
| """ |
| Based on https://github.com/kevinlu1211/pytorch-unet-resnet-50-encoder/blob/master/u_net_resnet_50_encoder.py |
| |
| Helper module that consists of a Conv -> BN -> ReLU |
| """ |
|
|
| def __init__( |
| self, |
| in_channels, |
| out_channels, |
| padding=1, |
| kernel_size=3, |
| stride=1, |
| with_nonlinearity=True, |
| ): |
| super().__init__() |
| self.conv = nn.Conv2d( |
| in_channels, |
| out_channels, |
| padding=padding, |
| kernel_size=kernel_size, |
| stride=stride, |
| ) |
| self.bn = nn.BatchNorm2d(out_channels) |
| self.relu = nn.ReLU(inplace=True) if with_nonlinearity else None |
|
|
| def forward(self, x): |
| x = self.conv(x) |
| x = self.bn(x) |
| if self.relu is not None: |
| x = self.relu(x) |
| return x |
|
|
|
|
| class Bridge(nn.Module): |
| """ |
| Based on https://github.com/kevinlu1211/pytorch-unet-resnet-50-encoder/blob/master/u_net_resnet_50_encoder.py |
| """ |
|
|
| def __init__(self, in_channels, out_channels): |
| super().__init__() |
| self.bridge = nn.Sequential( |
| ConvBlock(in_channels, out_channels), ConvBlock(out_channels, out_channels) |
| ) |
|
|
| def forward(self, x): |
| return self.bridge(x) |
|
|
|
|
| class UpBlockForUNetWithResNet50(nn.Module): |
| """ |
| Based on https://github.com/kevinlu1211/pytorch-unet-resnet-50-encoder/blob/master/u_net_resnet_50_encoder.py |
| |
| Up block that encapsulates one up-sampling step which consists of Upsample -> ConvBlock -> ConvBlock |
| """ |
|
|
| def __init__( |
| self, |
| in_channels, |
| out_channels, |
| up_conv_in_channels=None, |
| up_conv_out_channels=None, |
| upsampling_method="conv_transpose", |
| ): |
| super().__init__() |
|
|
| if up_conv_in_channels == None: |
| up_conv_in_channels = in_channels |
| if up_conv_out_channels == None: |
| up_conv_out_channels = out_channels |
|
|
| if upsampling_method == "conv_transpose": |
| self.upsample = nn.ConvTranspose2d( |
| up_conv_in_channels, up_conv_out_channels, kernel_size=2, stride=2 |
| ) |
| elif upsampling_method == "bilinear": |
| self.upsample = nn.Sequential( |
| nn.Upsample(mode="bilinear", scale_factor=2), |
| nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1), |
| ) |
| self.conv_block_1 = ConvBlock(in_channels, out_channels) |
| self.conv_block_2 = ConvBlock(out_channels, out_channels) |
|
|
| def forward(self, up_x, down_x): |
| """ |
| :param up_x: this is the output from the previous up block |
| :param down_x: this is the output from the down block |
| :return: upsampled feature map |
| """ |
| x = self.upsample(up_x) |
| x = torch.cat([x, down_x], 1) |
| x = self.conv_block_1(x) |
| x = self.conv_block_2(x) |
| return x |
|
|
|
|
| class ModelResUNet_ft(nn.Module): |
| def __init__( |
| self, |
| res_base_model, |
| out_size, |
| imagenet_pretrain, |
| linear_probe=False, |
| use_base=True, |
| ): |
| super(ModelResUNet_ft, self).__init__() |
| self.resnet_dict = { |
| "resnet50": models.resnet50(weights=imagenet_pretrain), |
| } |
| resnet = self._get_res_basemodel(res_base_model) |
| self.use_base = use_base |
| if not self.use_base: |
| num_ftrs = int(resnet.fc.in_features / 2) |
| self.res_features = nn.Sequential(*list(resnet.children())[:-3]) |
| self.res_l1_anatomy = nn.Linear(num_ftrs, num_ftrs) |
| self.res_l2_anatomy = nn.Linear(num_ftrs, 256) |
| self.res_l1_pathology = nn.Linear(num_ftrs, num_ftrs) |
| self.res_l2_pathology = nn.Linear(num_ftrs, 256) |
|
|
| self.mask_generator = nn.Linear(num_ftrs, num_ftrs) |
| self.back = nn.Linear(256, num_ftrs) |
| self.last_res = nn.Sequential(*list(resnet.children())[-3:-1]) |
| else: |
| self.res_features = nn.Sequential(*list(resnet.children())[:-3]) |
| self.d = { |
| "input": 3, |
| "conv1": 64, |
| "conv2": 256, |
| "conv3": 512, |
| "conv4": 1024, |
| "bridge": 1024, |
| "up1": 512, |
| "up2": 256, |
| "up3": 128, |
| "up4": 64, |
| } |
| self.downscale_factors = { |
| "input": 1, |
| "conv1": 2, |
| "conv2": 4, |
| "conv3": 8, |
| "conv4": 16, |
| "bridge": 16, |
| "up1": 8, |
| "up2": 4, |
| "up3": 2, |
| "up4": 1, |
| } |
| |
| self.bridge = Bridge(self.d["conv4"], self.d["bridge"]) |
| self.up_blocks = nn.ModuleList( |
| [ |
| UpBlockForUNetWithResNet50( |
| in_channels=self.d["up1"] + self.d["conv3"], |
| out_channels=self.d["up1"], |
| up_conv_in_channels=self.d["bridge"], |
| up_conv_out_channels=self.d["up1"], |
| ), |
| UpBlockForUNetWithResNet50( |
| in_channels=self.d["up2"] + self.d["conv2"], |
| out_channels=self.d["up2"], |
| up_conv_in_channels=self.d["up1"], |
| up_conv_out_channels=self.d["up2"], |
| ), |
| UpBlockForUNetWithResNet50( |
| in_channels=self.d["up3"] + self.d["conv1"], |
| out_channels=self.d["up3"], |
| up_conv_in_channels=self.d["up2"], |
| up_conv_out_channels=self.d["up3"], |
| ), |
| UpBlockForUNetWithResNet50( |
| in_channels=self.d["up4"] + self.d["input"], |
| out_channels=self.d["up4"], |
| up_conv_in_channels=self.d["up3"], |
| up_conv_out_channels=self.d["up4"], |
| ), |
| ] |
| ) |
| self.out_size = out_size |
| self.dropout = nn.Dropout(p=0.2) |
| self.seg_classifier = nn.Conv1d( |
| self.d["up4"], out_size, kernel_size=1, bias=True |
| ) |
|
|
| def _get_res_basemodel(self, res_model_name): |
| try: |
| res_model = self.resnet_dict[res_model_name] |
| print("Image feature extractor:", res_model_name) |
| return res_model |
| except: |
| raise ( |
| "Invalid model name. Check the config file and pass one of: resnet18 or resnet50" |
| ) |
|
|
| def image_encoder(self, xis): |
| |
| """ |
| 16 torch.Size([16, 1024, 14, 14]) |
| torch.Size([16, 196, 1024]) |
| torch.Size([3136, 1024]) |
| torch.Size([16, 196, 256]) |
| """ |
| batch_size = xis.shape[0] |
| res_fea = self.res_features(xis) |
| res_fea = rearrange(res_fea, "b d n1 n2 -> b (n1 n2) d") |
| x = rearrange(res_fea, "b n d -> (b n) d") |
| mask = self.mask_generator(x) |
| x_pathology = mask * x |
| x_pathology = self.res_l1_pathology(x_pathology) |
| x_pathology = F.relu(x_pathology) |
|
|
| x_pathology = self.res_l2_pathology(x_pathology) |
|
|
| out_emb_pathology = rearrange(x_pathology, "(b n) d -> b n d", b=batch_size) |
| out_emb_pathology = self.back(out_emb_pathology) |
| out_emb_pathology = rearrange(out_emb_pathology, "b (n1 n2) d -> b d n1 n2", n1=14, n2=14) |
|
|
| out_emb_pathology = out_emb_pathology.squeeze() |
|
|
| return out_emb_pathology |
|
|
| def forward(self, img): |
| x = img |
| down_embdding = [x] |
| for i in range(len(self.res_features)): |
| x = self.res_features[i](x) |
| if i == 2 or i == 4 or i == 5: |
| down_embdding.append(x) |
|
|
| o = self.bridge(x) |
|
|
| for i in range(len(self.up_blocks)): |
| o = self.up_blocks[i](o, down_embdding[len(down_embdding) - i - 1]) |
| o = self.dropout(o) |
| batch_size = o.shape[0] |
| h = o.shape[-2] |
| w = o.shape[-1] |
| class_number = o.shape[-3] |
| o = o.reshape(batch_size, class_number, h * w) |
| o = self.seg_classifier(o) |
| o = o.reshape(batch_size, self.out_size, h, w) |
| return o |
|
|
