Comar
Comar
At first, I exclude the images specified in list_low_colorful.csv in training as I mentioned in the original paper. On the other hand, cplx_img_ids.txt is the indices of validation images, so...
> > > generate paired prompts and then use larger batch size (perhaps by accumulation) > > > > > > What's the appropriate batch size, from your experience? >...
Thank you for your works! I will check the additional code and merge :) Now I'm personally busy, so it needs time.
Additionally, I also used python environment with yml file. And, I met heavy warning like below. ``` Some weights of the model checkpoint at openai/clip-vit-large-patch14 were not used when initializing...
Only a solution I know so far is to change optimizer SGD, which enables training with 24GB memory ... as you know, this is not good solution.
Thank you for your interest on our project. The script 'colorize.real.sh' is set to use a batch size of 1, so changing the batch size will not solve the problem....
This is really wierd. A possible scenario I can imagine is that there is an other process which is using a huge parts of the GPU memory. Did you check...
I'm sorry it is really hard to know what the problem is. The next steps I can recommend to you are as follow - Check the version of Pytorch. v1.10.0...
I found solution. For some reason, checkpoint wrapper in BasicTransformerBlock results in the aforementioned problem. By removing the wrapper, I solved the problem.
Maybe next week, I will provide the code .