How can I change pretrained resnet model to accept single channel image?

[vivekstorm (Vivek)]

I’m trying to use per-trained ResNet-18 model for binary classification with modification in input channel and kernel size of 1st Conv layer. What is the best way to do this? Also does this approach uses pretrained weight for conv1 layer?Currently I’m trying like this

model=models.resnet18(pretrained=True)
model.conv1=nn.Conv2d(1,64,kernel_size=(3,3),stride=(2,2),padding=(3,3),bias=False)
num_ftrs=model.fc.in_features
model.fc=nn.Linear(num_ftrs,2)

[Nishanth_Sasankan]

someone told me this, but I have not tried it yet

Grayscale images for resenet and deeplabv3

You can always define a custom resnet and change the first layer to adapt for your input shape. For example, with resnet18: import torch import torchvision resnet = torchvision.models.resnet18() input = torch.randn((16,3,244,244)) output = resnet(input) print(output.shape) # this fails becasue resnet expects 3 channels #input = torch.randn((16,1,244,244)) #output = resnet(input) #print(output.shape) import torch.nn as nn class MyResNet(nn.Module): def __init__(self, in_channels=1): …

[LeviViana]

import torch
from torchvision.models import resnet18
net = resnet18(pretrained=False) # You should put 'True' here 
net.conv1 = torch.nn.Conv1d(1, 64, (7, 7), (2, 2), (3, 3), bias=False)
batch = torch.rand(4, 1, 224, 224) # in the NCHW format
net(batch).size()