David Nukrai

How many workers are you using? Try to add "--num_workers 4" to check if it solves both the problem of slow training and the gpu utility (usually is the same...

Thanks!! Works now. Much appreciate it.

Thanks. IMO we need two abstraction levels of metrics here; first, the one we have now, that enables numeric metrics per single run, and secondly, metrics of analyzing many metrics...

- [ ] Hi, thanks for the interest in our work. We describe in the paper both with and without modalities offset. When it is without the offset so yes,...

Hi, [here](https://github.com/DavidHuji/CapDec#datasets) are the instructions. Please let me know if you encounter any issue.

Hi, sorry for the confusion. The json (single_caption_per_sample_val) holds the captions data (per id) and it is generated in the script of parse_karpathy. So once you download the data from...

Thank you for your interest in our work. Generally, I recommend using the default hyperparameters as a starting point, which are the values defined in the arg-parser. The number of...

Hi, thanks for the comment. [Here](https://drive.google.com/drive/folders/13VHl-eD87hoD4R_EBFKOG5z6P7geULid?usp=sharing) are the weights for table 1 and fig.3 (also see [here](https://drive.google.com/drive/folders/13VHl-eD87hoD4R_EBFKOG5z6P7geULid?usp=sharing)). If you got significantly different results when you trained, I'll be happy to...

Hi, thanks for the comment. If you follow the instructions in the Readme, the default configuration should be sufficient. Regarding the number of epochs, we trained until the validation results...

It is an interesting question. We did not reproduce ZeroCap's results but rather took them from Magic's paper. I guess in Magic's paper they improved ZeroCap method somehow. Although the...