Thank you, can you please tell me which operation should I use for fcn architecture. I have used addition. I upscaled the vgg pool5 layers by 32 for fcn32 and gives good results. But when I add pool4 layers and (pool 5 up scale by 2) to get fcn16 output, it performs bad. According to the fcn architecture, fcn16 and fcn8 are supposed to work better.
this is my code,
import torch
import torch.nn as nn
import torchvision.models as models
from pytorch_model_summary import summary
vgg16 = models.vgg16(pretrained=True)
for param in vgg16.features.parameters():
param.requires_grad = False
#False Total params: 185,771,904 Trainable params: 171,057,216 Non-trainable params: 14,714,688
#true Total params: 185,771,904 Trainable params: 185,771,904 Non-trainable params: 0
class fcn(nn.Module):
def init(self):
super(fcn, self).init()
self.features = vgg16.features
self.classifier = nn.Sequential(
nn.Conv2d(512, 4096, 7),
nn.ReLU(inplace=True),
#nn.Dropout2d(),
nn.Conv2d(4096, 4096, 1),
nn.ReLU(inplace=True),
#nn.Dropout2d(),
nn.Conv2d(4096, 32, 1),
nn.ConvTranspose2d(32, 32, 224, stride=32)
)
def forward(self, x):
x = self.features(x)#/32
x = self.classifier(x)
#print(x.shape)
return x
class fcn16(nn.Module):
def init(self):
super(fcn16, self).init()
self.features = vgg16.features
self.classifier = nn.Sequential(
nn.Conv2d(512, 4096, 7),
nn.ReLU(inplace=True),
nn.Conv2d(4096, 4096, 1),
nn.ReLU(inplace=True),
nn.Conv2d(4096, 32, 1)
)
self.score_pool4 = nn.Conv2d(512, 32, 1)
self.upscore2 = nn.ConvTranspose2d(32, 32, 14, stride=2, bias=False)
self.upscore16 = nn.ConvTranspose2d(32, 32, 16, stride=16, bias=False)
def forward(self, x):
pool4 = self.features:-7#512 features /16
pool5 = self.features-7:#512 features /16/2=/32
pool5_upscored = self.upscore2(self.classifier(pool5))#32 class features stride2 /32*2=/16
pool4_scored = self.score_pool4(pool4)#32 features /16
combined = pool4_scored + pool5_upscored
#combined = torch.cat([pool4_scored, pool5_upscored])
res = self.upscore16(combined)# /1
return res
class fcn8(nn.Module):
def init(self):
super(fcn8, self).init()
self.features = vgg16.features
self.classifier = nn.Sequential(
nn.Conv2d(512, 4096, 7),
nn.ReLU(inplace=True),
nn.Conv2d(4096, 4096, 1),
nn.ReLU(inplace=True),
nn.Conv2d(4096, 32, 1)
)
self.score_pool4 = nn.Conv2d(512, 32, 1)
self.score_pool3 = nn.Conv2d(256, 32, 1)
self.upscore2 = nn.ConvTranspose2d(32, 32, 14, stride=2, bias=False)
self.upscore3 = nn.ConvTranspose2d(32, 32, 2, stride=2, bias=False)
#self.upscore16 = nn.ConvTranspose2d(32, 32, 16, stride=16, bias=False)
self.upscore8 = nn.ConvTranspose2d(32, 32, 8, stride=8, bias=False)
def forward(self, x):
pool3 = self.features:-14#256 features /8
pool4 = self.features-14:-7#512 features /8/2=16
pool5 = self.features-7:#512 features /16/2=/32
pool5_upscored = self.upscore2(self.classifier(pool5))#32 class features stride2 /322=/16
pool4_scored = self.score_pool4(pool4)#32 class features /16
pool3_scored = self.score_pool3(pool3)#32 class features /8
combined = pool4_scored + pool5_upscored #/16
#print(combined.shape)
combined_upscored = self.upscore3(combined)#32 class features stride2 /162=/8
#print(combined_upscored.shape)
combined2 = pool3_scored + combined_upscored
#print(combined2.shape)
#res = self.upscore16(combined)#/1
res = self.upscore8(combined2)#/1
#print(res.shape)
return res
Thank you