generator.py

import torch
from torch import nn
from .modules import Conv2dBlock, Concat

class SkipEncoderDecoder(nn.Module):
    def __init__(self, input_depth, num_channels_down = [128] * 5, num_channels_up = [128] * 5, num_channels_skip = [128] * 5):
        super(SkipEncoderDecoder, self).__init__()

        self.model = nn.Sequential()
        model_tmp = self.model

        for i in range(len(num_channels_down)):

            deeper = nn.Sequential()
            skip = nn.Sequential()

            if num_channels_skip[i] != 0:
                model_tmp.add_module(str(len(model_tmp) + 1), Concat(1, skip, deeper))
            else:
                model_tmp.add_module(str(len(model_tmp) + 1), deeper)
            
            model_tmp.add_module(str(len(model_tmp) + 1), nn.BatchNorm2d(num_channels_skip[i] + (num_channels_up[i + 1] if i < (len(num_channels_down) - 1) else num_channels_down[i])))

            if num_channels_skip[i] != 0:
                skip.add_module(str(len(skip) + 1), Conv2dBlock(input_depth, num_channels_skip[i], 1, bias = False))
                
            deeper.add_module(str(len(deeper) + 1), Conv2dBlock(input_depth, num_channels_down[i], 3, 2, bias = False))
            deeper.add_module(str(len(deeper) + 1), Conv2dBlock(num_channels_down[i], num_channels_down[i], 3, bias = False))

            deeper_main = nn.Sequential()

            if i == len(num_channels_down) - 1:
                k = num_channels_down[i]
            else:
                deeper.add_module(str(len(deeper) + 1), deeper_main)
                k = num_channels_up[i + 1]

            deeper.add_module(str(len(deeper) + 1), nn.Upsample(scale_factor = 2, mode = 'nearest'))

            model_tmp.add_module(str(len(model_tmp) + 1), Conv2dBlock(num_channels_skip[i] + k, num_channels_up[i], 3, 1, bias = False))
            model_tmp.add_module(str(len(model_tmp) + 1), Conv2dBlock(num_channels_up[i], num_channels_up[i], 1, bias = False))

            input_depth = num_channels_down[i]
            model_tmp = deeper_main

        self.model.add_module(str(len(self.model) + 1), nn.Conv2d(num_channels_up[0], 3, 1, bias = True))
        self.model.add_module(str(len(self.model) + 1), nn.Sigmoid())
    
    def forward(self, x):
        return self.model(x)


def input_noise(INPUT_DEPTH, spatial_size, scale = 1./10):
    shape = [1, INPUT_DEPTH, spatial_size[0], spatial_size[1]]
    return torch.rand(*shape) * scale