* Optimize Token Classification models for TPU
As per the XLA document XLA cannot handle masked indexing well. So token classification
models for BERT and others use an implementation based on `torch.where`. This implementation
works well on TPU.
ALBERT token classification model uses the masked indexing which causes performance issues
on TPU. This PR fixes this issue by following the BERT implementation.
* Same fix for ELECTRA
* Same fix for LayoutLM
* Properly use test_fetcher for examples
* Fake example modification
* Fake modeling file modification
* Clean fake modifications
* Run example tests for any modification.
* Fix issue when labels are supplied as Numpy array instead of list
* Fix issue when labels are supplied as Numpy array instead of list
* Fix same issue in the `TokenClassification` data collator
* Style pass
Update GPT Neo ONNX config to match the changes implied by the simplification of the local attention
Co-authored-by: Michael Benayoun <michael@huggingface.co>
* Add long-overdue link to the Google TRC project
* Apply suggestions from code review
Co-authored-by: Suraj Patil <surajp815@gmail.com>
Co-authored-by: Stefan Schweter <stefan@schweter.it>
* Enabling dataset iteration on pipelines.
Enabling dataset iteration on pipelines.
Unifying parameters under `set_parameters` function.
Small fix.
Last fixes after rebase
Remove print.
Fixing text2text `generate_kwargs`
No more `self.max_length`.
Fixing tf only conversational.
Consistency in start/stop index over TF/PT.
Speeding up drastically on TF (nasty bug where max_length would increase
a ton.)
Adding test for support for non fast tokenizers.
Fixign GPU usage on zero-shot.
Fix working on Tf.
Update src/transformers/pipelines/base.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Update src/transformers/pipelines/base.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Small cleanup.
Remove all asserts + simple format.
* Fixing audio-classification for large PR.
* Overly explicity null checking.
* Encapsulating GPU/CPU pytorch manipulation directly within `base.py`.
* Removed internal state for parameters of the pipeline.
Instead of overriding implicitly internal state, we moved
to real named arguments on every `preprocess`, `_forward`,
`postprocess` function.
Instead `_sanitize_parameters` will be used to split all kwargs
of both __init__ and __call__ into the 3 kinds of named parameters.
* Move import warnings.
* Small fixes.
* Quality.
* Another small fix, using the CI to debug faster.
* Last fixes.
* Last fix.
* Small cleanup of tensor moving.
* is not None.
* Adding a bunch of docs + a iteration test.
* Fixing doc style.
* KeyDataset = None guard.
* RRemoving the Cuda test for pipelines (was testing).
* Even more simple iteration test.
* Correct import .
* Long day.
* Fixes in docs.
* [WIP] migrating object detection.
* Fixed the target_size bug.
* Fixup.
* Bad variable name.
* Fixing `ensure_on_device` respects original ModelOutput.
* Moving slow tokenizer to the Trie world.
* Adding more docstrings to the Trie.
* Fixing doctest (incompatible wiht our format? )
* Update src/transformers/tokenization_utils.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* Adding a lot more comment into the internals of this algorithm.
* Cleaner doc.
* Fixing the namings.
* Update src/transformers/tokenization_utils.py
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
* quality.
* Fixing longest first match.
* Small improvements to cuts + more test + canine resistant test.
* Fixing fast test.
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
Ibraheem Moosa