Martin Krasser

Thanks @ruizengalways, glad you like it!

Try `import tensorflow as tf`.

Try `lm("new_model.h5", custom_objects = { "tf": tf })`

What error do you get?

This error occurs when `align_image(img)` doesn't detect a face and returns `None`. The notebook doesn't handle that situation at the moment. I'll leave this ticket open so that I don't...

This project uses [Karpathy splits](https://github.com/krasserm/fairseq-image-captioning/tree/master/splits) for train, validation and test splitting. Karpathy splits are defined over MS-COCO train and validation sets only, hence test2014 is not loaded. Why do you...

> Is this because the GPU memory is not enough? Yes, this is the reason. The settings documented in the README are appropriate for 2 GTX 1080 cards (8 GB...

Did you pre-train the model with [CE loss](https://github.com/krasserm/fairseq-image-captioning/blob/master/README.md#cross-entropy-loss) before running SCST?

Convergence improves with higher --max-sentences values (but also requires more memory). A value of 5 should work fine on 11 GB cards. Regarding `checkpoint_best.pt`, this is the checkpoint with the...

What is the frequency of OOMs when you run with --max-sentences 5 or 8?