While implementing some of the paper from segmentation, I found it weird when changing conv2d parameters.
First I tried using .parameter to change parameters of conv2d, but this didn’t work and made strange results.
Code
import torch
import torch.nn as nn
import torchvision.models as models
vgg16 = models.vgg16(pretrained=True)
features = list(vgg16.features.children())
conv5 = nn.Sequential(*features[24:30])
for i in range(len(conv5)):
if isinstance(conv5[i], nn.Conv2d):
conv5[i].dilation = (2, 2)
conv5[i].padding = (2, 2)
conv5[i].padding_mode = 'reflect'
conv5_input = torch.randn(8, 512, 64, 64)
conv5_output = conv5(conv5_input)
print(conv5_input.shape, conv5_output.shape)
Result
torch.Size([8, 512, 64, 64]) torch.Size([8, 512, 58, 58])
I googled a lot and I finally solved by loading state_dict and this worked
Code
import torch
import torch.nn as nn
import torchvision.models as models
vgg16 = models.vgg16(pretrained=True)
features = list(vgg16.features.children())
fixed_1 = nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(2, 2), dilation=(2, 2), padding_mode='reflect')
fixed_2 = nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(2, 2), dilation=(2, 2), padding_mode='reflect')
fixed_3 = nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(2, 2), dilation=(2, 2), padding_mode='reflect')
w_conv51 = features[24].state_dict()
w_conv52 = features[26].state_dict()
w_conv53 = features[28].state_dict()
fixed_1.load_state_dict({'weight':w_conv51['weight'], 'bias':w_conv51['bias']})
fixed_2.load_state_dict({'weight':w_conv52['weight'], 'bias':w_conv52['bias']})
fixed_3.load_state_dict({'weight':w_conv53['weight'], 'bias':w_conv53['bias']})
fixed = nn.Sequential(
fixed_1, nn.ReLU(),
fixed_2, nn.ReLU(),
fixed_3, nn.ReLU())
fixed_input = torch.randn(8, 512, 64, 64)
fixed_output = fixed(fixed_input)
print(fixed_input.shape, fixed_output.shape)
Result
torch.Size([8, 512, 64, 64]) torch.Size([8, 512, 64, 64])
I have 2 questions for this example. (I googled, but I cannot find what I really want…)
-
Why first trial (.parameter to change conv2d layers with weight preserved) didn’t work and made strange result? (the network details are same when I print each network)
-
Is there more cool way to change conv2d parameters with preserving the weight? (more better way than what I did on final)
