Sean Moriarity
Sean Moriarity
With the support of `while` loops in `Nx` and progress towards streaming data to the device, we should consider an API which allows for fusing the training loop into a...
This is important for deploying on mobile/embedded applications. Goal is to achieve feature parity with PyTorch: https://pytorch.org/docs/stable/quantization.html
Currently, XLA's GPU `while` thunk actually runs the loop on the CPU and is inefficient for training RNNs. TensorFlow implements a [CuDNNRNN](https://www.tensorflow.org/api_docs/python/tf/raw_ops/CudnnRNN) op which takes advantage of specialized GPU RNN...
On the high-level API, we can integrate named tensors by specifying the expected names on input: ```elixir Axon.input(batch: nil, channels: 3, height: 224, width: 224) ``` Then we'll need to...
Notebooks are an excellent teaching resource. Examples should be converted to: [Livebook's](https://github.com/elixir-nx/livebook) Live Markdown with accompanying explanations of the models demonstrated.
Somewhat related to #21 PyTorch Lightning supports building custom loggers for integration with third-party logging tools (like TensorBoard). We should include a similar API so training can be monitored in...
Necessary for reproducible results. Requires https://github.com/elixir-nx/nx/issues/333