- 27 Nov, 2020 1 commit
-
-
Kristian Holsheimer authored
* [FlaxBert] Fix non-broadcastable attention mask for batched forward-passes * [FlaxRoberta] Fix non-broadcastable attention mask * Use jax.numpy instead of ordinary numpy (otherwise not jit-able) * Partially revert "Use jax.numpy ..." * Add tests for batched forward passes * Avoid unnecessary OOMs due to preallocation of GPU memory by XLA * Auto-fix style * Re-enable GPU memory preallocation but with mem fraction < 1/paralleism
-
- 17 Nov, 2020 1 commit
-
-
Sylvain Gugger authored
* Put models in subfolders * Styling * Fix imports in tests * More fixes in test imports * Sneaky hidden imports * Fix imports in doc files * More sneaky imports * Finish fixing tests * Fix examples * Fix path for copies * More fixes for examples * Fix dummy files * More fixes for example * More model import fixes * Is this why you're unhappy GitHub? * Fix imports in conver command
-
- 16 Nov, 2020 1 commit
-
-
Sylvain Gugger authored
* Use the CI to identify failing tests * Remove from all examples and tests * More default switch * Fixes * More test fixes * More fixes * Last fixes hopefully * Use the CI to identify failing tests * Remove from all examples and tests * More default switch * Fixes * More test fixes * More fixes * Last fixes hopefully * Run on the real suite * Fix slow tests
-
- 19 Oct, 2020 1 commit
-
-
Funtowicz Morgan authored
* WIP flax bert * Initial commit Bert Jax/Flax implementation. * Embeddings working and equivalent to PyTorch. * Move embeddings in its own module BertEmbeddings * Added jax.jit annotation on forward call * BertEncoder on par with PyTorch ! :D * Add BertPooler on par with PyTorch !! * Working Jax+Flax implementation of BertModel with < 1e-5 differences on the last layer. * Fix pooled output to take only the first token of the sequence. * Refactoring to use BertConfig from transformers. * Renamed FXBertModel to FlaxBertModel * Model is now initialized in FlaxBertModel constructor and reused. * WIP JaxPreTrainedModel * Cleaning up the code of FlaxBertModel * Added ability to load Flax model saved through save_pretrained() * Added ability to convert Pytorch Bert model to FlaxBert * FlaxBert can now load every Pytorch Bert model with on-the-fly conversion * Fix hardcoded shape values in conversion scripts. * Improve the way we handle LayerNorm conversion from PyTorch to Flax. * Added positional embeddings as parameter of BertModel with default to np.arange. * Let's roll FlaxRoberta ! * Fix missing position_ids parameters on predict for Bert * Flax backend now supports batched inputs Signed-off-by:
Morgan Funtowicz <morgan@huggingface.co> * Make it possible to load msgpacked model on convert from pytorch in last resort. Signed-off-by:
Stefan Schweter <stefan@schweter.it> * Attempt to make style * Attempt to make style in tests too * Added jax & jaxlib to the flax optional dependencies. * Attempt to fix flake8 warnings ... * Redo black again and again * When black and flake8 fight each other for a space ...
馃挜 馃挜 馃挜 * Try removing trailing comma to make both black and flake happy! * Fix invalid is_<framework>_available call, thanks @LysandreJik馃帀 * Fix another invalid import in flax_roberta test * Bump and pin flax release to 0.1.0. * Make flake8 happy, remove unused jax import * Change the type of the catch for msgpack. * Remove unused import. * Put seed as optional constructor parameter. * trigger ci again * Fix too much parameters in BertAttention. * Formatting. * Simplify Flax unittests to avoid machine crashes. * Fix invalid number of arguments when raising issue for an unknown model. * Address @bastings comment in PR, moving jax.jit decorated outside of __call__ * Fix incorrect path to require_flax/require_pytorch functions. Signed-off-by:Morgan Funtowicz <funtowiczmo@gmail.com> * Attempt to make style. Signed-off-by:
馃幎 Signed-off-by:Marc van Zee <marcvanzee@gmail.com> * Bump flax to it's latest version Co-authored-by:
Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
-
