JFan

Looks like tl.CrossEntropyLoss() is depreciated: https://trax-ml.readthedocs.io/en/latest/trax.layers.html#trax.layers.metrics.CrossEntropyLoss However, it looks like tl.WeightedCategoryCrossEntropy() might also have a memory leak (at least on TPU) so hold off on switching to trax 1.3.7 for...

According to the output, it appears that your GPU has ``2.00GiB`` of vram, and unfortunately, for large images being converted to numpy arrays, you would need significantly more memory. I'm...

In most cases, a 2-4gb card is enough (like running a 1920x1080 image), but an image with the longer side being larger than 5000 might need 5-6gb to even be...

The first print messages don't occur until the first checkpoint is written. Be patient, and be happy that it didn't just print the UID and then say it's done when...

Yep, just wait, it'll print the iterations when it has finished one

@spot92 Try changing `--checkpoint-iterations` to a lower value. It prints whenever it finishes that number of iterations

I'm seeing that there are a high number of epochs, while having a low number of iterations. Make sure you have the whole train2014 folder, and try lowering your epoch...

The models for each repo are inherently different. It would take insane amounts of time to manually "fix" the files to be interchangeable. If you're looking for a neural style...

@spot92 https://colab.research.google.com/ will help if you don't have linux And yes, that repo is significantly faster, with training times that can take at most 3 minutes on GPU

What are your computer's specs?