Ting Chen
Ting Chen
It's an empirical question and you could try both and select whichever is better for you. Sometimes even with the same sizes, random crops could still help. On Fri, Oct...
The sort function was written this way so it runs on TPU. If you just use GPU, you should be able to do with it `tf.sort`, which probably doesn't have...
This is the code that would run on TPU, and we didn't experience this issue. I wonder if there's any collapsing in the input features that lead to the same...
The 6 hour benchmark was run using the command in readme for pretraining ResNet-50 on ImageNet. However, that was run a while ago before some recent code update. So it...
unfortunately i don't recall the configuration that i used a few months ago (i remember it was pretty standard but..). i will try to find some time to test it...
Do you know which line of code leads to this error? and have you set the image size, model size, etc. the same for train run and for eval run?
I think you hit an edge case/bug :) since you use image_size=32, center_crop at eval time is disabled, so the code directly reshape a raw image of size >32x32x3 into...
it's similar to moco, in our case it helps ~1% across. we do not integrate the code in this repo to avoid increasing the complexity.
https://github.com/google-research/simclr/blob/master/colabs/finetuning.ipynb can do linear evaluation (note: imagenet dataset is not available in colab)
are you restoring weights from a checkpoint? I believe TF2 code should support training with multiple gpus, but @saxenasaurabh could further confirm.