I'm not an expert on this and would love to hear from one, but as far as I know there's simply more work to be done on AMD's end, to provide fast implementations of all the operations TensorFlow, PyTorch, and others need. ROCm hasn't reached parity with CUDA.

Theano used to work with OpenCL and therefore AMD, but the support was never that great, and I don't know that many people used it. And Theano is now EOL.