I have PyTorch tensors in the GPU, which I would like to convert to Tensorflow 1.x tensors without ever leaving the GPU.
I don't care about passing gradients back and forth. I just want the tensor to stay in the GPU. I don't care if my PyTorch tensor is copied to a TF tensor or modified in place.
I am aware that I can convert to numpy on the CPU and reload onto the GPU. However, I must stay on GPU for speed.