Expand 3d_classification example to multi-class use case

[baderstine]

The current 3d_classification example only allows a binary label (outcome).

Could the example be easily expanded or updated to demonstrate a multi-class label (outcome) use case?

[baderstine]

In my particular use case, I have 14 classes coded as integer values between 0-13.
When I attempt to use a multi-class labels object having these values, I got an error due to the loss = loss_function(outputs, labels) call:

RuntimeError: 0D or 1D target tensor expected, multi-target not supported

A search suggested that "When using PyTorch's nn.CrossEntropyLoss(), the target labels should not be one-hot encoded; instead, they should be integers representing the class indices." So my first attempted fix was to comment out the line that one-hot encodes the labels and leave them as a list of integers:
# labels = torch.nn.functional.one_hot(torch.as_tensor(labels)).float()

However, this change results in Target X is out of bounds errors, where X is one of the class integers. The specific value of X seems to change each time I run the code... I assume this is due to batch randomization.

[baderstine]

Ah... possibly this is as simple as keeping all the code the same but changing one line:
model = monai.networks.nets.DenseNet121(spatial_dims=3, in_channels=1, out_channels=X).to(device)
where X is the number of classes you are trying to predict (14 in my case).

[baderstine]

The tutorial doesn't explain where the final model is stored or how it could be extracted and deployed. Is there a separate tutorial for this? I would also like to run the model on the full training/test set to get predicted versus actual class for each input image and enable accuracy calculations by class. Can you include code in the tutorial for this kind of post-training model evaluation?

[baderstine]

This bit of code might be used to load the existing best model and run in a evaluation loop on a different evaluation ImageDataset and DataLoader :

    checkpoint_dir = os.path.join(path_checkpoint, 'best_metric_model_classification3d_array.pth')
    state_dict = torch.load(checkpoint_dir, map_location=device) 
    model.load_state_dict(state_dict)
    model.eval()