mirror of
https://github.com/labmlai/annotated_deep_learning_paper_implementations.git
synced 2025-11-01 03:43:09 +08:00
183 lines
6.5 KiB
Python
183 lines
6.5 KiB
Python
#!/bin/python
|
|
import torch
|
|
import torch.nn as nn
|
|
import torchvision
|
|
import torchvision.transforms as transforms
|
|
import torch.optim as optim
|
|
from torchsummary import summary
|
|
|
|
#custom import
|
|
import numpy as np
|
|
import time
|
|
import os
|
|
|
|
|
|
# ResBlock
|
|
class ResBlock(nn.Module):
|
|
def __init__(self, num_features, use_batch_norm=False):
|
|
super(ResBlock, self).__init__()
|
|
self.num_features = num_features
|
|
self.conv_layer1 = nn.Conv2d(num_features, num_features, kernel_size=3, stride=1, padding=1)
|
|
self.relu_layer = nn.ReLU()
|
|
self.conv_layer2 = nn.Conv2d(num_features, num_features, kernel_size=3, stride=1, padding=1)
|
|
|
|
self.use_batch_norm = use_batch_norm
|
|
if self.use_batch_norm:
|
|
self.batch_norm_layer1 = nn.BatchNorm2d(self.num_features)
|
|
self.batch_norm_layer2 = nn.BatchNorm2d(self.num_features)
|
|
|
|
for m in self.modules():
|
|
if isinstance(m, nn.Conv2d):
|
|
nn.init.kaiming_normal_(m.weight)
|
|
# nn.init.xavier_uniform_(m.weight)
|
|
|
|
def forward(self, x):
|
|
residual = x
|
|
x = self.conv_layer1(x)
|
|
if self.use_batch_norm:
|
|
x = self.batch_norm_layer1(x)
|
|
|
|
x = self.relu_layer(x)
|
|
x = self.conv_layer2(x)
|
|
if self.use_batch_norm:
|
|
x = self.batch_norm_layer2(x)
|
|
|
|
x += residual
|
|
x = self.relu_layer(x)
|
|
return x
|
|
|
|
# ResNet
|
|
class ResNet(nn.Module):
|
|
def __init__(self, in_features, num_class, feature_channel_list, batch_norm= False, num_stacks=1, zero_init_residual=True):
|
|
super(ResNet, self).__init__()
|
|
self.in_features = in_features
|
|
self.num_in_channel = in_features[2]
|
|
self.num_class = num_class
|
|
self.feature_channel_list = feature_channel_list
|
|
self.num_residual_blocks = len(self.feature_channel_list)
|
|
self.num_stacks = num_stacks
|
|
self.batch_norm = batch_norm
|
|
self.shape_list = []
|
|
self.shape_list.append(in_features)
|
|
self.module_list = nn.ModuleList()
|
|
self.zero_init_residual= zero_init_residual
|
|
self.build_()
|
|
|
|
def build_(self):
|
|
#track filter shape
|
|
cur_shape = self.GetCurShape()
|
|
cur_shape = self.CalcConvOutShape(cur_shape, kernel_size=7, padding=1, stride=2, out_filters= self.feature_channel_list[0])
|
|
self.shape_list.append(cur_shape)
|
|
|
|
if len(self.in_features) == 2:
|
|
in_channels = 1
|
|
else:
|
|
in_channels = self.in_features[2]
|
|
|
|
# First Conv layer 7x7 stride=2, pad =1
|
|
self.module_list.append(nn.Conv2d(in_channels= in_channels,
|
|
out_channels= self.feature_channel_list[0],
|
|
kernel_size=7,
|
|
stride=2,
|
|
padding=3))
|
|
|
|
|
|
#batch norm
|
|
if self.batch_norm: #batch_norm
|
|
self.module_list.append(nn.BatchNorm2d(self.feature_channel_list[0]))
|
|
|
|
# ReLU()
|
|
self.module_list.append(nn.ReLU())
|
|
|
|
for i in range(self.num_residual_blocks-1):
|
|
in_size = self.feature_channel_list[i]
|
|
out_size = self.feature_channel_list[i+1]
|
|
|
|
res_block = ResBlock(in_size, use_batch_norm=True)
|
|
|
|
# #Stacking Residual blocks
|
|
for num in range(self.num_stacks):
|
|
self.module_list.append(res_block)
|
|
|
|
# # Intermediate Conv and ReLU()
|
|
self.module_list.append(nn.Conv2d(in_channels=in_size,
|
|
out_channels= out_size,
|
|
kernel_size=3,
|
|
padding=1,
|
|
stride=2))
|
|
|
|
# track filter shape
|
|
cur_shape = self.CalcConvOutShape(cur_shape, kernel_size=3, padding=1,
|
|
stride=2, out_filters=out_size)
|
|
|
|
self.shape_list.append(cur_shape)
|
|
|
|
# # batch norm
|
|
if self.batch_norm: # batch_norm
|
|
self.module_list.append(nn.BatchNorm2d(out_size))
|
|
|
|
self.module_list.append(nn.ReLU())
|
|
|
|
# print("shape list", self.shape_list)
|
|
|
|
#TODO include in the main loop
|
|
#Last Residual block
|
|
res_block = ResBlock(out_size, use_batch_norm=True)
|
|
for num in range(self.num_stacks):
|
|
self.module_list.append(res_block)
|
|
|
|
#Last AvgPool layer
|
|
# self.module_list.append(nn.AvgPool2d(kernel_size=2, stride=2, padding=0))
|
|
self.module_list.append(nn.MaxPool2d(kernel_size=2, stride=2, padding=0))
|
|
|
|
# track filter shape
|
|
cur_shape = self.CalcConvOutShape(cur_shape, kernel_size=2, padding=0, stride=2, out_filters=out_size)
|
|
self.shape_list.append(cur_shape)
|
|
|
|
s = self.GetCurShape()
|
|
in_features = s[0] * s[1] * s[2]
|
|
|
|
# Initialization
|
|
for m in self.modules():
|
|
if isinstance(m, nn.Conv2d):
|
|
nn.init.kaiming_normal_(m.weight)
|
|
# nn.init.xavier_uniform_(m.weight)
|
|
|
|
# if self.zero_init_residual:
|
|
# for m in self.modules():
|
|
# if isinstance(m, ResBlock):
|
|
# nn.init.constant_(m.batch_norm_layer1.weight, 0)
|
|
# nn.init.constant_(m.batch_norm_layer2.weight, 0)
|
|
|
|
def GetCurShape(self):
|
|
return self.shape_list[-1]
|
|
|
|
def CalcConvFormula(self, W, K, P, S):
|
|
return int(np.floor(((W - K + 2 * P) / S) + 1))
|
|
|
|
# https://stackoverflow.com/questions/53580088/calculate-the-output-size-in-convolution-layer
|
|
# Calculate the output shape after applying a convolution
|
|
def CalcConvOutShape(self, in_shape, kernel_size, padding, stride, out_filters):
|
|
# Multiple options for different kernel shapes
|
|
if type(kernel_size) == int:
|
|
out_shape = [self.CalcConvFormula(in_shape[i], kernel_size, padding, stride) for i in range(2)]
|
|
else:
|
|
out_shape = [self.CalcConvFormula(in_shape[i], kernel_size[i], padding, stride) for i in range(2)]
|
|
|
|
return (out_shape[0], out_shape[1], out_filters) # , batch_size... but not necessary.
|
|
|
|
def AddMLP(self, MLP):
|
|
if MLP:
|
|
self.module_list.append(MLP)
|
|
|
|
# def MLP(self, in_features, num_classes, use_batch_norm=False, use_dropout=False, use_softmax=False):
|
|
# return nn.ReLU(nn.Linear(in_features, num_classes))
|
|
|
|
def forward(self, x):
|
|
for mod_name in self.module_list:
|
|
x = mod_name(x)
|
|
x = x.view(x.size(0), -1) # flat #TODO check if it works
|
|
return x
|
|
|
|
|