Age | Commit message (Collapse) | Author | |
---|---|---|---|
2022-09-30 | Update Hub client to 0.10 | ZJaume | |
2022-09-23 | Add test suite (#21) | Jaume Zaragoza | |
* Basic test for full model training * Extend full train test * Add train lite test * Ensure reproducibility of frequence noise * Unit test for noise generation * Add Tokenizer class test * Remove old test corpus file * Add classifier tests Download files on pytest setup to the test dir to avoid downloading it every time. Test normal, calibrated and raw modes. * Download models only in classifier test * Delete args object to avoid interference between tests | |||
2022-09-15 | Set SentencePiece seed | ZJaume | |
2022-09-15 | Flatten full model directory (again) | ZJaume | |
2022-09-15 | Fix transformer training | ZJaume | |
Ignore synthetic noise tag when loading data. Don't return tuples in datagen for transformer Fix TokenAndPositionEmbeddings call | |||
2022-09-05 | Look for model and vocab subdirectories when loading | ZJaume | |
2022-09-05 | Fix typo when writing version to metadata | ZJaume | |
2022-08-26 | Remove unused import in __init__.py | ZJaume | |
2022-08-24 | Restore classifier layer loading for old models | ZJaume | |
2022-08-23 | Write current package version to metadata | ZJaume | |
2022-08-23 | Move package version to setup.cfg | ZJaume | |
2022-08-19 | Don't load Hardrules objects if disabled | ZJaume | |
2022-08-17 | typo | Marta Bañón | |
2022-08-17 | typo | Marta Bañón | |
2022-08-09 | Restore retrocompatibility with older models | ZJaume | |
2022-08-09 | Fix loading local models | ZJaume | |
2022-08-09 | Speed improvements using padding longest and no max_length | ZJaume | |
2022-07-27 | Be more informative when model not found | ZJaume | |
2022-07-27 | Remove unset variable | ZJaume | |
2022-07-27 | Download models from HFHub if possible | ZJaume | |
2022-07-27 | Introduce model names | ZJaume | |
2022-07-27 | Remove vocab and model file config values, use only dir | ZJaume | |
2022-07-27 | Flatten the full model directory | ZJaume | |
2022-07-27 | Overwrite XLMRConfig class | ZJaume | |
This makes the classes more compatible with HF API and to be able to load them later more easily. | |||
2022-07-27 | Update version | ZJaume | |
2022-07-25 | Redirect predict progbar to stder in debug mode | ZJaume | |
2022-07-11 | Fix imports in classifier and train | ZJaume | |
2022-07-11 | Introduce BICLEANER_AI_THREADS to control number of threads | ZJaume | |
`--processes` option is now deprecated in favor of the environment variable. The TensorFlow threads is now set prior to initialization to avoid errors. | |||
2022-07-05 | Redirect all Keras progbars to stderr | ZJaume | |
Seems that Keras developers won't accept writing progbars to stderr, see [here](https://github.com/keras-team/keras/pull/12019) | |||
2022-07-05 | Set inter/intra_op parallelism to 0 by default | ZJaume | |
There's no need of setting it to max cpu, TF will set an optimal value | |||
2022-07-05 | Change XLMR subclass name to pass HF signature check | ZJaume | |
2022-07-05 | Metrics rename `update_state` | ZJaume | |
2022-06-16 | Force no verbosity in predict by default | ZJaume | |
2022-03-09 | Add Keras backend alias to custom_objs when loading lite models | ZJaume | |
2022-03-02 | Use empty string for user_defined symbols as default | ZJaume | |
None was being included in the vocab as a piece | |||
2022-03-01 | Explicit input shape when calibrating with TF 2.8 | ZJaume | |
2022-02-17 | Headers support | Cristian García Romero | |
2022-02-17 | Update to Hardrules 2.0 | ZJaume | |
2021-12-15 | Encode sentences during batching and unify Generator class | ZJaume | |
Huge memory savings due to vectorized and padded arrays don't stay in memory and are processed when needed. Also, speed penalty is negligible because workers process batches in parallel. | |||
2021-11-23 | Hide Transformers and Tensorflow logging messages on executable scripts | ZJaume | |
2021-11-15 | Merge branch 'dev' | ZJaume | |
2021-11-10 | Avoid write metadata file too early | ZJaume | |
2021-11-09 | Avoid generating empty sentences in synthetic noise | ZJaume | |
2021-11-08 | Remove useless logging info message | ZJaume | |
2021-11-05 | Update HF Transformers, no longer needed single GPU for prediction | ZJaume | |
2021-11-05 | Restore starting capital letter in frequency noise | ZJaume | |
2021-11-05 | Fix writing metadata when no lm is trained | ZJaume | |
2021-11-04 | Add save valid and use 'validation' naming instead of 'development | ZJaume | |
2021-11-04 | Add MCC as validation metric in XLMR, use default names for metrics | ZJaume | |
2021-07-07 | Load weights instead of full model when loading fails due to bad marshal ↵ | ZJaume | |
data (model saved with different Python version) |