swift-models

swift-models copied to clipboard

The DQN implementation within the Gym examples additionally needs Matplotlib and Numpy to work correctly, and we should create a section in the Readme there that describes this model and...

BradLarson

[BERT-CoLA] Data source resolution

2

The CoLA dataset is reliant on a [firebase url with included token](https://github.com/tensorflow/swift-models/blob/master/Datasets/CoLA/CoLA.swift#L70-L71). Identify the source and determine whether we should host these files from GCS instead. /cc @BradLarson @eaplatanios

texasmichelle

Convert image classification and generative examples to use X10 by default

5

With the release of the 0.9 toolchain containing X10, we now have a higher-performance alternative to the default eager execution mode. In the interest of having swift-models demonstrate best practices,...

BradLarson

Currently, `Alphabet`'s [`dictionary`](https://github.com/tensorflow/swift-models/blob/master/Support/Text/WordSeg/Alphabet.swift#L25) maps from `String` rather than `Character` to support tokens of length > 1 character. Using `Character` instead of `String` would work if we used special Unicode characters...

texasmichelle

Currently, the `WordSeg` dataset uses characters from all datasets to create an instance of `Alphabet`, but only the training set should be used. This potentially involves architectural changes to `CharacterSequence`,...

texasmichelle

Resolve occasional crash in Examples/GPT2-Inference

4

Examples/GPT2-Inference occasionally crashes unexpectedly. Figure out why!

texasmichelle

[TextUnsupervised] Add encoded version of WikiText103

1

We recently generated encoded version of WikiText2 and got both that and original uncoded files into GCS bucket. (https://github.com/tensorflow/swift-models/pull/490). To keep aligned, we wanted to do the same for WikiText103.

xihui-wu

All of our model inference unit tests use the default eager mode for execution. With X10 now available as an additional backend, we might want to add a parallel set...

BradLarson

Is there any documentation on how to use the CheckpointWriter API?

1

Is there any resource/notebook on Using CheckpointWriter api inside a training loop for any model in swift?

sibtain-raxa

There's a very cool post [here](https://blaisecruz.com/pytorch-on-tpus/) (HT @BradLarson) with benchmarks for PyTorch XLA. It would be great to see how x10 compares by adding a DistilGPT2 variant and `DistilGPT2-WikiText2` example.

texasmichelle