Closed RainHxj closed 2 years ago
@RainHxj Yes, we use 8 gpus in our experiments. The global batch size is 32.
Thanks for your reply. The text encoder is frozen in the training phase, which is implemented by lr=0.0. Compared with no grad, does this method affect the training speed?
Although the text encoder is fixed during training, we still need to compute its gradients to update the learnable context (the input of the text encoder). Therefore, the torch.no_grad()
mode cannot be used to accelerate the training.
Thanks.
Does the red box denote using a fixed prompt, such as "a photo of"?
We use learnable context in this case (Eq. 3 in our paper). Using the hard-craft prompt (a photo of class name
) will lead to a slightly worse result.
Thanks for your reply.
For ADE20K dataset, the batchsize is 32 (4*8gpus) ?