Changing the first Conv layer of a resnet34

vision
1

Hi experts and community folks, here I am trying to do some hack using list.children()

import fastai
from torchsummary import summary
from fastai.vision.models.unet import *
from fastai.callbacks import *
from fastai.utils.mem import *

arch = torchvision.models.resnet34(pretrained=False)

li = list(model.children()) #list of the children

print( li[0][0][0] ) -> gives Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)

So I do is li[0][0][0] = nn.Conv2d(1, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)

So here my input will be one channel which I want as the input to the resnet.

summary(model.to(‘cuda’),input_size=(1, 320, 320)) -> Here I have this error

NotImplementedError Traceback (most recent call last)
in
----> 1 summary(model.to(‘cuda’),input_size=(1, 320, 320))

~/anaconda3/envs/harsimar/lib/python3.6/site-packages/torchsummary/torchsummary.py in summary(model, input_size, batch_size, device)
70 # make a forward pass
71 # print(x.shape)
—> 72 model(*x)
73
74 # remove these hooks

~/anaconda3/envs/harsimar/lib/python3.6/site-packages/torch/nn/modules/module.py in call(self, *input, **kwargs)
530 result = self._slow_forward(*input, **kwargs)
531 else:
–> 532 result = self.forward(*input, **kwargs)
533 for hook in self._forward_hooks.values():
534 hook_result = hook(self, input, result)

~/anaconda3/envs/harsimar/lib/python3.6/site-packages/torch/nn/modules/container.py in forward(self, input)
98 def forward(self, input):
99 for module in self:
–> 100 input = module(input)
101 return input
102

~/anaconda3/envs/harsimar/lib/python3.6/site-packages/torch/nn/modules/module.py in call(self, *input, **kwargs)
530 result = self._slow_forward(*input, **kwargs)
531 else:
–> 532 result = self.forward(*input, **kwargs)
533 for hook in self._forward_hooks.values():
534 hook_result = hook(self, input, result)

~/anaconda3/envs/harsimar/lib/python3.6/site-packages/torch/nn/modules/module.py in forward(self, *input)
94 registered hooks while the latter silently ignores them.
95 “”"
—> 96 raise NotImplementedError
97
98 def register_buffer(self, name, tensor):

NotImplementedError:

Any views on this?

2

If you want to change the first layer, you can directly replace it via:

model = torchvision.models.resnet34(pretrained=False)
model.conv1 = nn.Conv2d(1, 64, (7, 7), (2, 2), (3, 3), bias=False)

x = torch.randn(1, 1, 224, 224)
out = model(x)

I’m not sure what might go wrong in your code, as it’s a bit hard to read.
You can post code snippets by wrapping them into three backticks ```, which makes debugging a bit easier. :wink:

1 Like
3

[quote=“Harsimar_Singh, post:1, topic:84215”]
So I did this first which resulted in error -

import fastai
from torchsummary import summary
from fastai.vision.models.unet import *
from fastai.callbacks import *
from fastai.utils.mem import *

arch = torchvision.models.resnet34(pretrained=False)

li = list(model.children()) #list of the children

print( li[0][0][0] ) -> gives Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)

So I do is li[0][0][0] = nn.Conv2d(1, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)

So here my input will be one channel which I want as the input to the resnet.

summary(model.to(‘cuda’),input_size=(1, 320, 320)) -> Here I have this error which is NotImplemented Error

After your suggestions I did this

def build_model(args):
    arch = torchvision.models.resnet34(pretrained=False)
    arch.conv1 = nn.Conv2d(1, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
    li = list(arch.children())
    arch = nn.Sequential(*li[:8])
    model = DynamicUnet(encoder=arch,n_classes = 1,img_size = (320,320)) #fastai unet converter class using a pertained network as encoder 
    model.to(args.device)
    return model

And this worked beautifully !! Thank you so much :smiley: :smiley:

1 Like