""" Author: Mélanie Gaillochet Date: 2021-02-12 """ import torch import torch.nn as nn from Utils.unet_utils import double_conv_block_2d, max_pooling_2d, up_conv_2d, \ pad_to_shape class ModUNet2D(nn.Module): def __init__(self, config): """ This modified UNet differs from the original one with the use of leaky relu (instead of ReLU) and the addition of residual connections. The idea is to help deal with fine-grained details :param in_channels: # of input channels (ie: 3 if image in RGB) :param out_channels: # of output channels (# segmentation classes) """ super(ModUNet2D, self).__init__() self.in_channels = config["in_channels"] self.out_channels = config["out_channels"] self.num_init_filters = config["num_init_filters"] # Kernel size, stride, paddding, normalizatin and activation function will be passed as keyword arguments kwargs = config['structure'] # We set the dropout self.dropout = nn.Dropout(kwargs['dropout_rate']) # Encoder part self.enc_1 = double_conv_block_2d(self.in_channels, self.num_init_filters // 2, self.num_init_filters, **kwargs['conv_block']) self.pool_1 = max_pooling_2d(**kwargs['pooling']) self.enc_2 = double_conv_block_2d(self.num_init_filters, self.num_init_filters, self.num_init_filters * 2, **kwargs['conv_block']) self.pool_2 = max_pooling_2d(**kwargs['pooling']) self.enc_3 = double_conv_block_2d(self.num_init_filters * 2, self.num_init_filters * 2, self.num_init_filters * 4, **kwargs['conv_block']) self.pool_3 = max_pooling_2d(**kwargs['pooling']) # Center part self.center = double_conv_block_2d(self.num_init_filters * 4, self.num_init_filters * 4, self.num_init_filters * 8, **kwargs['conv_block']) # Decoder part self.up_1 = up_conv_2d(self.num_init_filters * 8, self.num_init_filters * 8, **kwargs['upconv']) self.dec_1 = double_conv_block_2d(self.num_init_filters * 12, self.num_init_filters * 4, self.num_init_filters * 4, **kwargs['conv_block']) self.up_2 = up_conv_2d(self.num_init_filters * 4, self.num_init_filters * 4, **kwargs['upconv']) self.dec_2 = double_conv_block_2d(self.num_init_filters * 6, self.num_init_filters * 2, self.num_init_filters * 2, **kwargs['conv_block']) self.up_3 = up_conv_2d(self.num_init_filters * 2, self.num_init_filters * 2, **kwargs['upconv']) self.dec_3 = double_conv_block_2d(self.num_init_filters * 3, self.num_init_filters, self.num_init_filters, **kwargs['conv_block']) # Output self.out = nn.Conv2d(self.num_init_filters, self.out_channels, kernel_size=1, padding=0) def forward(self, x): # Encoding # print('\nEncoder 1') enc_1 = self.enc_1(x) # -> [BS, 64, x, y, z], if num_init_filters=64 # print(enc_1.shape) out = self.pool_1(enc_1) # -> [BS, 64, x/2, y/2, z/2] # We put a dropout layer out = self.dropout(out) # print('\nEncoder 2') enc_2 = self.enc_2(out) # -> [BS, 128, x/2, y/2, z/2] # print(enc_2.shape) out = self.pool_2(enc_2) # -> [BS, 128, x/4, y/4, z/4] # We put a dropout layer out = self.dropout(out) # print('\nEncoder 3') enc_3 = self.enc_3(out) # -> [BS, 256, x/4, y/4, z/4] # print(enc_3.shape) out = self.pool_3(enc_3) # -> [BS, 256, x/8, y/8, z/8] # We put a dropout layer out = self.dropout(out) # Center # print('\nCenter') center = self.center(out) # -> [BS, 512, x/8, y/8, z/8] # print(center.shape) # Decoding # print('\nDecoder 1') out = self.up_1(center) # -> [BS, 512, x/4, y/4, z/4] # print(out.shape) out = pad_to_shape(out, enc_3.shape) # print(out.shape) out = torch.cat([out, enc_3], dim=1) # -> [BS, 768, x/4, y/4, z/4] # print(out.shape) out = self.dropout(out) dec_1 = self.dec_1(out) # -> [BS, 256, x/4, y/4, z/4] # print(dec_1.shape) # print('\nDecoder 2') out = self.up_2(dec_1) # -> [BS, 256, x/2, y/2, z/2] # print(out.shape) out = pad_to_shape(out, enc_2.shape) # print(out.shape) out = torch.cat([out, enc_2], dim=1) # -> [BS, 384, x/2, y/2, z/2] # print(out.shape) out = self.dropout(out) dec_2 = self.dec_2(out) # -> [BS, 128, x/2, y/2, z/2] # print(dec_2.shape) # print('\nDecoder 3') out = self.up_3(dec_2) # -> [BS, 128, x, y, z] # print(out.shape) out = pad_to_shape(out, enc_1.shape) # print(out.shape) out = torch.cat([out, enc_1], dim=1) # -> [BS, 192, x, y, z] # print(out.shape) out = self.dropout(out) dec_3 = self.dec_3(out) # -> [BS, 64, x, y, z] # print(dec_3.shape) # We put a dropout layer # out = self.dropout(dec_3) # Output # print('\nOutput') out = self.out(dec_3) # -> [BS, out_channels, x, y, z] # print(out.shape) return out, [enc_1, enc_2, enc_3, center, dec_1, dec_2, dec_3] if __name__ == '__main__': # from unet import UNet from torchinfo import summary config = {'in_channels': 1, 'out_channels': 1, 'num_init_filters': 64, "structure": { "dropout_rate": 0, "conv_block": { "normalization": "group_norm", "activation_fct": "leakyReLU", "kernal_size": 3, "stride": 1, "padding": 1 }, "pooling": { "kernal_size": 2, "stride": 2, "padding": 0 }, "upconv": { "kernal_size": 2, "stride": 2, "padding": 0 } } } model = ModUNet2D(config) # x = torch.randn(size=(20, 1, 512, 512), dtype=torch.float32) # with torch.no_grad(): # out, [enc_1, enc_2, enc_3, center] = model(x) # # print(f'Out: {out.shape}') # print(f'enc_1: {enc_1.shape}') # print(f'enc_2: {enc_2.shape}') # print(f'enc_3: {enc_3.shape}') # print(f'center: {center.shape}') batch_size = 2 summary = summary(model, (batch_size, 1, 256, 256))