[FIXED] How to use the Inception model for transfer learning in PyTorch?

Issue

I have created a PyTorch torchvision model for transfer learning, using the pre-built ResNet50 base model, like this:

        # Create base model from torchvision.models
        model = resnet50(pretrained=True)
        num_features = model.fc.in_features

        # Define the network head and attach it to the model
        model_head = nn.Sequential(
            nn.Linear(num_features, 512),
            nn.ReLU(),
            nn.Dropout(0.25),
            nn.Linear(512, 256),
            nn.ReLU(),
            nn.Dropout(0.5),
            nn.Linear(256, num_classes),
        )
        model.fc = model_head

Now I wanted to use the Ineception v3 model instead as base, so I switched from resnet50() above to inception_v3(), the rest stayed as is. However, during training I get the following error:

TypeError: cross_entropy_loss(): argument 'input' (position 1) must be Tensor, not InceptionOutputs

So how can one use the Inception v3 model from torchvision.models as base model for transfer learning?

Solution

From PyTorch documentation about Inceptionv3 architecture:

This network is unique because it has two output layers when training. The primary output is a linear layer at the end of the network. The second output is known as an auxiliary output and is contained in the AuxLogits part of the network.

Have a look at this tutorial: https://pytorch.org/tutorials/beginner/finetuning_torchvision_models_tutorial.html#inception-v3 There you can find how to use transfer learning for several models, included ResNet and Inception.

Answered By - caumente