Fastai WandbCallback with multilabel datasets

• Weights and Biases version: 0.8.9
• Python version: 3.6.9
• Operating System: Linux-4.14.133-88.112.amzn1.x86_64-x86_64-with-Ubuntu-16.04-xenial

Description

I’m trying to see prediction samples for fastai image classification, where the images are multilabel. I’d like it to be ran on the validation data at the end of each epoch and to see captions of ground truth/predictions.

When I use a multilabel dataset, instead of ground truth/prediction captions for each image, I get a three copies of each image side by side, where each is labeled, “Input data”, “Prediction”, “Ground Truth” (but not the actual categories)

What I Did

callback_fns += [partial(WandbCallback, input_type='images', monitor='acc_thresholded')]

[austinmw]

@borisdayma Yes that’s the output shape. It’s difficult for me to create a minimum reproducible example, because my code is currently completely intertwined with aws sagemaker. However, a good example of an almost identical problem (fastai, cnn, multilabel) is fastai’s course v3 part 1 lesson 3 example using the planets dataset: https://nbviewer.jupyter.org/github/fastai/course-v3/blob/master/nbs/dl1/lesson3-planet.ipynb I think if it works with that it should work with mine as well.

[borisdayma]

Sorry I had a typo. What is the shape of output tensor? Is it (batch_size, 4) with each value representing probability of a class?

A simple example may make it easier to debug if possible. Maybe you are even adapting one of fastai notebooks that I can look at?