Clarification for the Mask RCNN tutorial

vision
1

Hi,
In this tutorial: https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html

There is a code line that has an error:

        if self.transforms is not None:
            img, target = self.transforms(img, target)

According to the definition in the previous code section, self.transforms is a Compose object, so it accepts only one argument:

def get_transform(train):
    transforms = []
    transforms.append(T.ToTensor())
    if train:
        transforms.append(T.RandomHorizontalFlip(0.5))
    return T.Compose(transforms)

Can you please clarify what’s the correct syntax that one should use?

2

Also, I think the ToTensor should come after RandomHorizontalFlip, because RandomHorizontalFlip accepts PIL images.

3

You had me looking there, too! :slight_smile:

What happens is that the tutorial ships with its own little transforms module to bridge that gap.

Best regards

Thomas

4

Yep,thanks.I already requested to delete this thread. Will read more carefully next time

5

Oh, the next person with the question might find it useful. :slight_smile:

1 Like
6

Thanks @tom
Do you know how to save the model after training? It’s a JIT ScriptModule, and it doesn’t support pickle.dump or torch.save . There are open GitHub issues on it. I wonder if you know of a workaround.

7

It is considered best practice to save the state_dict, i.e. the parameters, rather than the full model. (You need the source anyway to re-instantiate the model.) This is what all pre-trained torchvision models do, too.
The obvious alternative is to trace the model to get a JIT model and save that.

Best regards

Thomas