ZeroDivisionError
wavymulder opened this issue · comments
I'm attempting to get DLAS training ESRGAN. I'm on Windows 10 and using Anaconda.
Here's what I'm getting:
Traceback (most recent call last):
File "train.py", line 361, in <module>
trainer.init(args.opt, opt, args.launcher)
File "train.py", line 114, in init
self.total_epochs = int(math.ceil(total_iters / train_size))
ZeroDivisionError: division by zero
Not sure how to fix my train_size being zero. Is my path to my images wrong? Thanks.
Yup my path was wrong, my mistake sorry!
Hey there,
I'm really sorry, but I cannot spare the effort to actually support DLAS for wider use. Despite the fact that I started down the road of writing documentation and providing "recipes" like exists for ESRGAN, I don't maintain a test suite that ensures that these things don't get broken and it would be impossible for me to ensure backwards compatibility.
With that being said, I believe what is happening in your case is you've misconfigured the dataset somehow such that train_size=0. I'd set num_workers=0 and try adding some breakpoints into single_image_dataset.py to try and figure out what is going on.
It has also been some time since I've trained a GAN on DLAS. If you're interested in actually getting this to work, I would highly recommend reverting to the commit where I added the ESRGAN recipe: bbc677d
To be clear: I'm happy to answer questions. I just will not be fixing any bugs that get reported. Happy to review PRs though.
No worries, with Tortoise taking off the way it is, I'm sure you've been a lot busier than expected. I'm still new to all this in general so I'll try to find an easier spot to jump in.