AttributeError: 'CrossEntropyLoss' object has no attribute 'label_smoothing'

[iimog]

When trying to use the provided fastai2 model for anything but making single predictions using predict, I get the above error message. This happens in fastai version 2.7.10 and pytorch version 1.13.1. I assume that the loss function in the saved model is too old, so it does not yet have the label_smoothing attribute.

[iimog]

I was able to fix this problem by replacing the loss_func with the current implementation like this:

from fastai.vision.all import *
trainedModel = load_learner("resnet34_5percent_size256_extremeTfms_ceLoss_fastai2.pkl")
trainedModel.loss_func = CrossEntropyLossFlat(axis=1)

As the previous loss function was also CrossEntropyLossFlat, this should not make any difference in training.

After that, it should be possible, to replace the dataloaders like this to continue training on own data:

def label_func(x):
    return str(x).replace("image","mask")

def get_parent(x):
    return Path(x).parent.name == 'val'

dbl = DataBlock(blocks=(ImageBlock, MaskBlock(codes = np.array(["background","left_ventricle","myocardium"]))),
        get_items = get_image_files,
        get_y = label_func,
        splitter = FuncSplitter(get_parent),
        item_tfms=Resize(512, method='crop'),
        batch_tfms=aug_transforms(do_flip=True,max_rotate=90,max_lighting=.4,max_zoom=1.2,size=256))

dls = dbl.dataloaders("data/images", bs=16)
trainedModel.dls = dls

[iimog]

I'll re-upload a version with fixed loss in a new release soon. Documentation for re-training should go to the repository and not into an (unrelated) issue.

[iimog]

New model is uploaded and documentation in the proper location.