Import adapted Resnet model

vision
1

Hi everyone, I would like to use a pre-trained Resnet model, but with some modifications. I would like to change the input from 3 channels (RGB) to gray scale (1 input). I would also like to change the output from 1000 classes to 10. Furthermore, it would be interesting to implement a deconvnet to see the feature maps at the different filters.

I do all this changes in the original Resnet code, and save it as resnet_adapted in the same folder. How do I import this model pre-trained? Could I freeze certain layers, just like in normal transfer learning?

I have been advised to use:

import sys
sys.path.append(r'C:\Users\Jacobo\anaconda3\envs\deep learning 3\Lib\site-packages\torchvision\models')
import resnet_adapted

but I think I cannot obtain a pre-trained model from that, nor freeze specific layers.

Thanks in advance for the help!

2

Depending how much has changed in your custom Resnet implementation, you could either try to directly load the pretrained resnet state_dict with the strict=False option:

resnet_adapted = MyResnet()
pretrained = torchvision.models.resnet50(pretrained=True)
resnet_adapted.load_state_dict(pretrained.state_dict(), strict=False)

or load specific layers manually by iterating the keys in the pretrained state_dict.

Why do you think you cannot freeze specific layers in your model? Are you seeing any errors?