swift-models icon indicating copy to clipboard operation
swift-models copied to clipboard

Models and examples built with Swift for TensorFlow

Results 46 swift-models issues
Sort by recently updated
recently updated
newest added

Review the scope of `WordSeg` so that internal components are not exposed unnecessarily, especially evolution code.

`Sequential{N}` typealiases were added in https://github.com/tensorflow/swift-apis/pull/934: ```swift import TensorFlow // Before: typealias Model = Sequential // After: typealias Model = Sequential3 ``` The tuple-like typealiases are easier to use. Long...

help wanted
good first issue

The CoLA example does not currently have a clear separation between model and dataset. Instantiating an instance of `CoLA` involves [providing a pre-trained BERT model](https://github.com/tensorflow/swift-models/blob/master/Examples/BERT-CoLA/main.swift#L38-L42). This could be clearer by...

help wanted

Verify that [all BERT variants](https://github.com/tensorflow/swift-models/blob/master/Models/Text/BERT.swift#L361-L386) converge and match a reference implementation.

help wanted
good first issue

The CoLA dataset uses [custom extraction functions](https://github.com/tensorflow/swift-models/blob/master/Datasets/CoLA/DataUtilities.swift#L328-L349) to download data files. Use tools common to other datasets in swift-models instead. /cc @BradLarson @sgugger

help wanted
good first issue

Before more transformer-based language models are added to this repo, let's pay down some debt on the ones we have: - [ ] #315 Various BERT cleanup - [x] #480...

help wanted

Imported from JIRA https://bugs.swift.org/browse/TF-1207 Type: | Bug | Priority: | Minor -- | -- | -- | -- Reporter: | Takehiro Araki | Assignee: | Unassigned Resolution: | Unresolved |...

When running the GPTInferenceUI on Windows, it seems that the model shape mismatch is back: ``` Fatal error: Incompatible shapes: [1,43,768] vs. [1,43]: file S:\4\tensorflow-swift-apis\Sources\TensorFlow\bindings\EagerExecution.swift, line 301 ``` As soon...

Imported from JIRA https://bugs.swift.org/browse/TF-643 Type: | Bug | Priority: | Minor -- | -- | -- | -- Reporter: | Brennan Saeta | Assignee: | Unassigned Resolution: | Unresolved |...

Currently, model loading is being done by explicitly writing the name of the tensor, for example, [BERT](https://github.com/tensorflow/swift-models/blob/ead9a321f0a41b169a5ece1a12d008160c10c383/Models/Text/BERT.swift#L795), and VGG in #334. While VGG only has 16 layers, when going forward...