info-ground
info-ground copied to clipboard
BigRedT
Hi, can I ask some question about your paper ？ (●'◡'●)
Thank you very much for an extraordinary job！
I'm very interested in your work and I hope to follow in your footsteps (●'◡'●)
Can your work run in many gpu ?
Hi @jun0wanan,
In current form, our code base only supports single GPU training. Part of the challenge in supporting multi-gpu training is the contrastive loss which requires one caption to be compared to all other images in the mini-batch to compute the loss. Note that this is different from typical classification tasks where an image and its label are sufficient to compute loss for that sample and therefore it is easy to partition the batch and place each partition on a separate GPU.
One solution that might work reasonably well is to only use images placed on the same GPU as negatives for the contrastive loss. For example, if the batch size is 100 and you have 4 GPUs, each GPU handles a subset of size 25. So instead of using 99 images as negatives for each caption, you would be using 24.
Hope this helps!
Hi @jun0wanan,
In current form, our code base only supports single GPU training. Part of the challenge in supporting multi-gpu training is the contrastive loss which requires one caption to be compared to all other images in the mini-batch to compute the loss. Note that this is different from typical classification tasks where an image and its label are sufficient to compute loss for that sample and therefore it is easy to partition the batch and place each partition on a separate GPU.
One solution that might work reasonably well is to only use images placed on the same GPU as negatives for the contrastive loss. For example, if the batch size is 100 and you have 4 GPUs, each GPU handles a subset of size 25. So instead of using 99 images as negatives for each caption, you would be using 24.
Hope this helps!
hi，author~ I find a .py have a little error? Maybe it is that I didn't understand your setting..
In https://github.com/BigRedT/info-ground/blob/master/exp/ground/run/eval_flickr_phrase_loc_model_selection.py
model_nums = find_all_model_numbers(exp_const.model_dir) for num in model_nums: continue if num <= 3000: continue model_const.model_num = n
continue why ?
Hi @jun0wanan,
In current form, our code base only supports single GPU training. Part of the challenge in supporting multi-gpu training is the contrastive loss which requires one caption to be compared to all other images in the mini-batch to compute the loss. Note that this is different from typical classification tasks where an image and its label are sufficient to compute loss for that sample and therefore it is easy to partition the batch and place each partition on a separate GPU.
One solution that might work reasonably well is to only use images placed on the same GPU as negatives for the contrastive loss. For example, if the batch size is 100 and you have 4 GPUs, each GPU handles a subset of size 25. So instead of using 99 images as negatives for each caption, you would be using 24.
Hope this helps!
hi ，and I find my model will decrease a lot:
