From 46624a615429232cee01be670d925dd593ceb6a3 Mon Sep 17 00:00:00 2001
From: Jordan Gong <jordan.gong@protonmail.com>
Date: Wed, 23 Dec 2020 20:59:44 +0800
Subject: Refactor and refine auto-encoder

1. Wrap Conv2d 3x3-padding-1 to VGGConv2d
2. Wrap ConvTranspose2d 4x4-stride-4-padding-1 to DCGANConvTranspose2d
3. Turn off bias in conv since the employment of batch normalization
---
 models/auto_encoder.py | 22 ++++++++--------
 models/layers.py       | 68 ++++++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 78 insertions(+), 12 deletions(-)

(limited to 'models')

diff --git a/models/auto_encoder.py b/models/auto_encoder.py
index 1708bc9..bb4a377 100644
--- a/models/auto_encoder.py
+++ b/models/auto_encoder.py
@@ -2,6 +2,8 @@ import torch
 from torch import nn as nn
 from torch.nn import functional as F
 
+from models.layers import VGGConv2d, DCGANConvTranspose2d
+
 
 class Encoder(nn.Module):
     def __init__(self, in_channels: int, opt):
@@ -12,20 +14,20 @@ class Encoder(nn.Module):
 
         # Cx[HxW]
         # Conv1 3x64x32 -> 64x64x32
-        self.conv1 = nn.Conv2d(in_channels, nf, kernel_size=3, padding=1)
+        self.conv1 = VGGConv2d(in_channels, nf)
         self.batch_norm1 = nn.BatchNorm2d(nf)
         # MaxPool1 64x64x32 -> 64x32x16
         self.max_pool1 = nn.AdaptiveMaxPool2d((32, 16))
         # Conv2 64x32x16 -> 256x32x16
-        self.conv2 = nn.Conv2d(nf, nf * 4, kernel_size=3, padding=1)
+        self.conv2 = VGGConv2d(nf, nf * 4)
         self.batch_norm2 = nn.BatchNorm2d(nf * 4)
         # MaxPool2 256x32x16 -> 256x16x8
         self.max_pool2 = nn.AdaptiveMaxPool2d((16, 8))
         # Conv3 256x16x8 -> 512x16x8
-        self.conv3 = nn.Conv2d(nf * 4, nf * 8, kernel_size=3, padding=1)
+        self.conv3 = VGGConv2d(nf * 4, nf * 8)
         self.batch_norm3 = nn.BatchNorm2d(nf * 8)
         # Conv4 512x16x8 -> 512x16x8 (for large dataset)
-        self.conv4 = nn.Conv2d(nf * 8, nf * 8, kernel_size=3, padding=1)
+        self.conv4 = VGGConv2d(nf * 8, nf * 8)
         self.batch_norm4 = nn.BatchNorm2d(nf * 8)
         # MaxPool3 512x16x8 -> 512x4x2
         self.max_pool3 = nn.AdaptiveMaxPool2d((4, 2))
@@ -61,20 +63,16 @@ class Decoder(nn.Module):
         self.fc = nn.Linear(self.em_dim, nf * 8 * 2 * 4)
         self.batch_norm_fc = nn.BatchNorm1d(nf * 8 * 2 * 4)
         # TransConv1 512x4x2 -> 256x8x4
-        self.trans_conv1 = nn.ConvTranspose2d(nf * 8, nf * 4, kernel_size=4,
-                                              stride=2, padding=1)
+        self.trans_conv1 = DCGANConvTranspose2d(nf * 8, nf * 4)
         self.batch_norm1 = nn.BatchNorm2d(nf * 4)
         # TransConv2 256x8x4 -> 128x16x8
-        self.trans_conv2 = nn.ConvTranspose2d(nf * 4, nf * 2, kernel_size=4,
-                                              stride=2, padding=1)
+        self.trans_conv2 = DCGANConvTranspose2d(nf * 4, nf * 2)
         self.batch_norm2 = nn.BatchNorm2d(nf * 2)
         # TransConv3 128x16x8 -> 64x32x16
-        self.trans_conv3 = nn.ConvTranspose2d(nf * 2, nf, kernel_size=4,
-                                              stride=2, padding=1)
+        self.trans_conv3 = DCGANConvTranspose2d(nf * 2, nf)
         self.batch_norm3 = nn.BatchNorm2d(nf)
         # TransConv4 3x32x16
-        self.trans_conv4 = nn.ConvTranspose2d(nf, out_channels, kernel_size=4,
-                                              stride=2, padding=1)
+        self.trans_conv4 = DCGANConvTranspose2d(nf, out_channels)
 
     def forward(self, fa, fgs, fgd):
         x = torch.cat((fa, fgs, fgd), dim=1).view(-1, self.em_dim)
diff --git a/models/layers.py b/models/layers.py
index 2be93ad..f3ccbeb 100644
--- a/models/layers.py
+++ b/models/layers.py
@@ -4,6 +4,74 @@ import torch
 import torch.nn as nn
 
 
+class BasicConv2d(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            kernel_size: Union[int, Tuple[int, int]],
+            **kwargs
+    ):
+        super(BasicConv2d, self).__init__()
+        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size,
+                              bias=False, **kwargs)
+
+    def forward(self, x):
+        return self.conv(x)
+
+
+class VGGConv2d(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            kernel_size: Union[int, Tuple[int, int]] = 3,
+            padding: int = 1,
+            **kwargs
+    ):
+        super(VGGConv2d, self).__init__()
+        self.conv = BasicConv2d(in_channels, out_channels, kernel_size,
+                                padding=padding, **kwargs)
+
+    def forward(self, x):
+        return self.conv(x)
+
+
+class BasicConvTranspose2d(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            kernel_size: Union[int, Tuple[int, int]],
+            **kwargs
+    ):
+        super(BasicConvTranspose2d, self).__init__()
+        self.trans_conv = nn.ConvTranspose2d(in_channels, out_channels,
+                                             kernel_size, bias=False, **kwargs)
+
+    def forward(self, x):
+        return self.trans_conv(x)
+
+
+class DCGANConvTranspose2d(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            kernel_size: Union[int, Tuple[int, int]] = 4,
+            stride: int = 2,
+            padding: int = 1,
+            **kwargs
+    ):
+        super(DCGANConvTranspose2d).__init__()
+        self.trans_conv = BasicConvTranspose2d(in_channels, out_channels,
+                                               kernel_size, stride=stride,
+                                               padding=padding, **kwargs)
+
+    def forward(self, x):
+        return self.trans_conv(x)
+
+
 class FocalConv2d(nn.Module):
     def __init__(
             self,
-- 
cgit v1.2.3