Why the output of children part of a network has low resolution?

Hello. I train one network and save it, then load this network and I partition network to three module and give one images to input of first module and take output from end module but the output of this three module has low resolution… I dont understand why this happened?!!

model =  ConvAutoencoder()
max_epochs =10
outputs = train(model, num_epochs=max_epochs)

torch.save(model, 'model1.pth')
model1 = torch.load('model1.pth')

imgs3, labels = next(iter(test_loader))

f= model1(imgs3)

plt.imshow(f[13][0].detach().numpy(), cmap="gray") 

l1 = nn.Sequential(*list(model1.children())[:3])

l2 = nn.Sequential(*list(model1.children())[3:4])

l3 = nn.Sequential(*list(model1.children())[4:])

imgs2, labels = next(iter(test_loader))

y1= l1(imgs2)



plt.subplot(1, 2, 1)
plt.imshow(imgs2[13][0].detach().numpy(), cmap="gray") 
plt.subplot(1, 2, 2)
plt.imshow(y3[13][0].detach().numpy(), cmap="gray") 

without shift2

pls reply this question

please reply to this question.

Based on the posted code I assume the left image represents the input while the right one the model output?
If that’s the case, I guess your model isn’t able to create sharp images and you could check the literature for new architectures, which could avoid the blurry output.

1 Like