Brad Larson
Brad Larson
The recently released nvcaffe version 0.16.2 adds support for fp16 calculations and weight storage via the FLOAT16 type. Following the instructions here: https://docs.nvidia.com/deeplearning/dgx/caffe-user-guide/index.html#nvcaffeparm and in this presentation from GTC: http://on-demand.gputechconf.com/gtc/2017/presentation/s7218-boris-gainsberg-training-deep-networks-with-half-precision-float.pdf...
Now that this is getting close enough to the point where people can start using it in notebooks, etc., I think it might help to improve some of the default...
The Tensor initializer lets you manually specify a device for placement. For example, you should be able to explicitly place a Tensor on the first CPU using the eager mode...
As part of integrating Deep Q-Networks in [swift-models PR #617](https://github.com/tensorflow/swift-models/pull/617), the `Raw.gatherNd()` operator was needed. No extension to Tensor implemented this, so a `dimensionGathering()` function on Tensor was added [here](https://github.com/tensorflow/swift-models/blob/master/Gym/DQN/Gathering.swift)....
Some simple image classification models can trigger a segfault when using the XLA device specifically on macOS. For now, we're explicitly having them use the eager-mode device instead until this...
As pointed out in PR #579, the tf.keras implementation of `categorical_hinge()` expects values to be -1 or 1, but [will convert binary labels of 0 or 1 to -1 and...
Thanks for putting this library together, we're looking at using it in the near term to perform image loading and saving within Swift for TensorFlow's models ([pending PR for this](https://github.com/tensorflow/swift-models/pull/429)...
With Swift for TensorFlow, and the Swift data science ecosystem in general, still being fairly early-stage, many elements of Python-based TensorFlow don't have equivalents on the Swift side. This looks...
As brought up by Xander in [swift-apis issue #25](https://github.com/tensorflow/swift-apis/issues/25), our efforts on checkpointing have thus far focused on model state checkpoints, but there are other state parameters that should be...
In testing PR #671, we noticed that the GPT-2 model now exhausts all available memory on 8 GB GPUs (example: GTX 1080) for both eager mode and X10 runtimes. It...