Skip to content

Latest commit

 

History

History
1979 lines (1072 loc) · 58.2 KB

CHANGELOG.md

File metadata and controls

1979 lines (1072 loc) · 58.2 KB

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

0.63.6 - 2022-03-24

Added

Fixed

  • Fixed bug in NER ONNX prediction gaganmanku96
  • Fixed bug in NER tokenization which prevented multiprocessing being used correctly mxa4646
  • Fixed some broken links in docs jordimas

[0.63.5] - 2022-02-25

Added

  • Added support for selecting FAISS index type with RetrievalModel.

[0.63.4] - 2021-11-20

Added

  • Added support for individual training of context/query encoders in RetrievalModel.

Fixed

  • Fixed a bug for pre-tokenized input to ONNX models. whr778
  • BigBird bugfix where training dataset samples were being truncated at 510 token. whr778
  • Fixed bug when FP16 is not used with RetrievalModel. tiena2cva
  • Fixed bug in auto model for QuestionAnsweringModel. lekhang4497
  • Fixed bug where wrong predictions weren't returned in ClassificationModel jinschoi

0.63.0 - 2021-11-10

Added

  • Added support for document retrieval with the RetrievalModel class

0.62.1 - 2021-09-24

Fixed

  • Fixed bug when using onnx with ClassificationModel kongyurui
  • Fixed potential bug with NERModel when the input text is empty whr778
  • Fixed bug in sentencepiece tokenizer for some models whr778
  • Fixed issue with Seq2SeqModel showing the first training epoch as epoch 0 dopc
  • Fixed bug where eval_file was not used with ConvAIModel cahya-wirawan

Changed

  • Replaced tensorboardx imports with default torch version

0.62.0 - 2021-09-24

Added

  • Additional loss functions for ClassificationModel and NERModel. Zhylkaaa

Changed

  • Deprecated custom classification models. Zhylkaaa

[0.61.14] - 2021-09-23

Changed

  • W&B run id is now assigned as a model attribute

[0.70.0] - 2021-08

TODO

DOCS

  • use_hf_datasets
  • RAG
  • Retrieval

[0.61.13] - 2021-07-24

Added

  • Pretraining and finetuning BigBird and XLMRoBERTa LMs whr778

[0.61.10] - 2021-07-13

Added

Fixed

  • Bug fixes

[0.61.9] - 2021-06-21

Fixed

  • Updated W&B repo label

[0.61.8] - 2021-06-21

Fixed

  • Reverted changes to W&B repo label which weren't working.

0.61.7 - 2021-06-21

Changed

  • Updated W&B repo label

0.61.6 - 2021-05-28

Fixed

  • Fixed the onnx predict loop whr778

Added

  • Added NER support for BigBird, Deberta, Deberta-v2, and xlm pretrained models whr778
  • Added BigBird for regular sequence classification (not multilabel) @manueltonneau

0.61.5 - 2021-05-18

Added

  • Fixed possible bug when using HF Datasets with Seq2SeqModel
  • Added repo: simpletransformers to W&B config

0.61.4 - 2021-03-27

Added

  • Bug fixed in LanguageModelingModel which could occur when loading a GPU trained model on CPU. alvaroabascar
  • Bug fixed in NER ONNX for models with token type ids. whr778
  • Bug fixed in NER lazy loading. mhdhdri
  • Bug fixed in sliding window tie breaking calebchiam

Changed

  • Thread count no longer fixed when using ONNX rayline

[0.61.3] - 2021-03-19

Changed

  • Return full retrieved docs with RAG

Added

  • Added extra args for RAG:
    • split_text_character
    • split_text_n

0.61.0 - 2021-03-19

Added

  • Added support for RAG models (in Seq2Seq) - docs will be updated soon
  • Added support for Huggingface Datasets library for memory efficient training. Currently supports:
    • Classification (all)
    • NER
    • Language Modeling
    • Seq2Seq
    • T5
    • QA (Note that HF Datasets might not always work with QAModel)

0.60.9 - 2021-02-19

Added

  • Added XLNet support for NER

Fixed

  • Fixed bug where polynomial_decay_schedule_power value was not being set correctly

Changed

  • Switched to using FastTokenizers where possible

[0.60.8] - 2021-02-12

Fixed

  • Fixed bug in loading cached features with classification models

[0.60.7] - 2021-02-11

Changed

  • Multiprocessing during tokenization is now turned on by default. If you experience any instability, this can be turned off by setting use_multiprocessing=False

[0.60.6] - 2021-02-05

Changed

  • Multiprocessing during tokenization is now turned off by default. You can enable this by setting use_multiprocessing=True. However, the latest Pytorch versions seems to be unstable when using multiprocessing.

[0.60.3] - 2021-02-02

Changed

  • Multiprocessing is now turned off by default for evaluation. This is to avoid potential errors when doing evaluation during training. You can enable this by setting use_multiprocessing_for_evaluation to True.

0.60.2 - 2021-02-02

Fixed

  • Fixed bug in ClassificationDataset mapmeld

0.60.1 - 2021-02-02

Added

  • Added new NER models:

    • ALBERT
    • MPNet
    • SqueezeBERT
    • MobileBERT

  • Added new QA models:

0.60.0 - 2021-02-02

Added

  • Added class weights support for Longformer classification
  • Added new classification models:
    • SqueezeBert
    • DeBERTa
    • MPNet

Changed

  • Updated ClassificationModel logic to make it easier to add new models

0.51.16 - 2021-01-29

Fixed

  • Fixed bug in LayoutLM classification

0.51.15 - 2021-01-24

Fixed

  • Fixed bug in Language Generation models mapmeld
  • Fixed bug in MBart models nilboy

0.51.14 - 2021-01-24

Fixed

  • Fixed bug introduced in 0.51.12 when using sliding window

0.51.13 - 2021-01-11

Fixed

  • Fixed bug introduced in 0.51.12 with multiclass classification

0.51.12 - 2021-01-11

Changed

  • Added Area under the ROC curve (AUROC) and the Area under the Precision-Recall curve (AUPRC) as default metrics for binary classification @manueltonneau

Fixed

  • Fixed issues with models not being set to train modd when evaluating while training nilboy

0.51.11 - 2021-01-09

Changed

  • Removed do_lower_case when using AutoTokenizer

0.51.10 - 2020-12-29

Fixed

  • Fixed bug in QuestionAnsweringModel when using cached features for evaluation
  • Fixed bugs in ConvAIModel due to compatibility issues

0.51.9 / 0.51.8 - 2020-12-29

0.51.9 / 0.51.8 - 2020-12-29

0.51.9 / 0.51.8 - 2020-12-29

Added

  • Added the special_tokens_list arg which can be used to add additional special tokens to the tokenizer karthik19967829

0.51.7 - 2020-12-29

Fixed

  • Fixed bug during predicton when sliding_window=True and output_hidden_states=True calebchiam

[0.51.6] - 2020-12-10

Added

Fixed

  • Fixed bug where T5Model would save evaluation checkpoints even when save_eval_checkpoints is False.
  • Fixed bug where args.silent was not used in NERModel. mossadhelali

Changed

  • Changed the default value of dataloader_num_workers (for Pytorch Dataloaders) to 0. This is to avoid memory leak issues with Pytorch multiprocessing with text data.

0.51.5 - 2020-12-10

Added

  • Added support for T5/mT5 models in Simple Viewer

Fixed

  • Fixed bug where QuestionAnsweringModel and Seq2SeqModel would save (although not use) cached features even when no_cache is set

0.51.3 - 2020-12-10

Fixed

0.51.2 - 2020-12-09

Fixed

  • Fixed bug in ConvAI interact_single() method

0.51.1 - 2020-12-08

Fixed

  • Fixed bug in mbart predict() function. DM2493

Added

  • Added docs for language generation and multi-modal classifcation

0.51.0 - 2020-12-05

Added

  • Added support for MT5
  • Added support for Adafactor optimizer
  • Added support for various schedulers:
    • get_constant_schedule
    • get_constant_schedule_with_warmup
    • get_linear_schedule_with_warmup
    • get_cosine_schedule_with_warmup
    • get_cosine_with_hard_restarts_schedule_with_warmup
    • get_polynomial_decay_schedule_with_warmup

Changed

  • T5Model now has a required model_type parameter ("t5" or "mt5")

Fixed

  • Fixed issue with class weights not working in ClassificationModel when using mult-GPU training

0.50.0 - 2020-12-01

Changed

  • Compatibility with Transformers 4.0.0.

0.49.4 - 2020-11-25

Added

  • Added not_saved_args to model_args. Any args specified in this set will not be saved when the model is saved.
  • RepresentationModel improvements. aesuli

0.49.3 - 2020-11-09

Changed

  • ROC and PR W&B charts are no longer generated when using sliding window to avoid an error.
  • Fixed issue with ONNX in NER gaganmanku96
  • Fixed issues with wandb sweeps khituras ohstopityou

0.49.1 - 2020-11-22

Fixed

  • Fixed issue with Marian models using deprecated function. @bestvater
  • Added custom tokenizer option and random (no-pretraining) initialization option for T5Model. sarapapi

0.49.0 - 2020-11-09

Added

Fixed

  • Fixed issue with Seq2SeqModel when the model_name contained backslashes.
  • Fixed issue with saving args when a dataset_class is specified in Seq2SeqModel.

Changed

  • The Electra implementation used with ClassificationModel is now fully compatible with Hugging Face.

0.48.15 - 2020-10-22

Fixed

  • Updated some tokenizer arguments to the new argument names. macabdul9
  • Learning rate is now obtained from the get_last_lr() method. sarapapi

0.48.14 - 2020-10-12

Fixed

  • Fixed predict() function issue when using sliding_window.
  • Fixed issues with simple-viewer (streamlit compatibility issues)

0.48.13 - 2020-10-12

Fixed

  • Fixed issues with using mixed precision training with LanguageModelingModel.

0.48.12 - 2020-10-12

Fixed

0.48.11 - 2020-10-11

Changed

  • The train_model() method now returns training details. Specifically; global_step: Number of global steps trained training_details: Average training loss if evaluate_during_training is False or full training progress scores if evaluate_during_training is True

0.48.10 - 2020-10-11

Added

  • Added support for special tokens with Seq2SeqModel. Zhylkaaa

0.48.9 - 2020-10-07

Changed

  • Moved model.train() inside train() method.

0.48.8 - 2020-10-07

Added

0.48.7 - 2020-10-03

Added

  • Added support for multilabel classification with the CamemBERT model. @adrienrenaud

Changed

  • Output arrays in classification evaluate/predict now avoids np.append(). This should be more time and memory efficient.

0.48.6 - 2020-09-26

Added

  • Added layoutlm model for NER

Fixed

  • Potential fix for inconsistent eval_loss calculation

0.48.5 - 2020-09-17

Added

  • Added convert_to_onnx function to the following models:
    • ClassificationModel
    • NERModel
  • Converted ONNX models can be loaded (requires specifying onnx: True in model_args) and used for prediction.
  • Added fp16 support for evaluation and prediction (requires Pytorch >= 1.6) for the following models:
    • ClassificationModel
    • ConvAI
    • MultiModalClassificationModel
    • NERModel
    • QuestionAnsweringModel
    • Seq2Seq
    • T5Model
  • Added multigpu prediction/eval in
    • ClassificationModel
    • ConvAI
    • MultiModalClassificationModel
    • NERModel
    • QuestionAnsweringModel
    • Seq2Seq
    • T5Model

Fixed

  • Thread count can now be specified for MultiLabelClassificationModel.

0.48.4 - 2020-09-23

Fixed

  • Fixed compatibility issue with transformers 3.2. (BertPreTrainedModel was being imported from an incompatible path)

[0.48.3] - 2020-09-08

  • Version numbering issue fixed.

[0.48.2] - 2020-09-08

Fixed

  • Fixed missing padding_strategy argument in squad_convert_example_to_features()

0.48.1 - 2020-09-08

Fixed

  • Bug when using sliding window with multiclass classification
  • Bug in ConvAI where model was being accessed before being created

0.48.0 - 2020-09-06

Added

  • Added dynamic quantization support for all models.
  • Added ConvAI docs to documentation website. @pablonm3

[0.47.6] - 2020-09-01

Fixed

  • Fixed missing padding_strategy argument in squad_convert_example_to_features() cahya-wirawan

[0.47.5] - 2020-09-01

Added

  • Added dynamic quantization, thread_count arg, and avoids padding during inference for Classification models. karthik19967829

Fixed

  • Bug fix which fixes reprocessing data after reading from cache in Seq2SeqDataset and SimpleSummarizationDataset @Zhylkaaa

0.47.4 - 2020-08-29

Fixed

  • Bug fix in MultilabelClassificationModel when using sentence pairs.

0.47.3 - 2020-08-19

Fixed

  • Bug fix in ConvAI Sxela

0.47.0 - 2020-08-09

Added

  • Added support for testing models through a Streamlit app. Use the command `simple-viewer". Currently supports:

See docs for details.

0.46.5 - 2020-08-05

Changed

  • Python version requirement changed back to 3.6 for Colab support.
  • Miscellaneous bug fixes in 0.46.3 and 0.46.4

0.46.2 - 2020-08-01

Fixed

  • Fixed unreachable condition in Electra language modeling.

0.46.1 - 2020-08-01

Fixed

  • Bug in ConvAI models where cache_dir was not being created.

0.46.0 - 2020-08-01

Changed

0.45.5 - 2020-07-29

Fixed

  • Bug fixed in loading classiication models with a labels_map where labels are ints.

0.45.4 - 2020-07-28

Fixed

  • Bug fixed in lazy loading classification tasks where lazy_text_column=0 caused an error.

0.45.2 - 2020-07-25

Added

  • Added dataloader_num_workers to ModelArgs for specifying the number of processes to be used with a Pytorch dataloader.

Changed

  • Bumped required transformers version to 3.0.2

0.45.0 - 2020-07-19

Added

  • Added Text Representation Generation (RepresentationModel). @pablonm3

0.44.0 - 2020-07-05

Added

  • Lazy loading support added for QuestionAnsweringModel.

0.43.6 - 2020-07-05

Fixed

  • Bug fixed in Seq2Seq tasks.
  • Bug fixed in NERModel where the classification report was missing in checkpoints.
  • Bug fixed in ELECTRA.
  • Bug fixed in Seq2Seq generic encoder-decoder model.
  • Bug fixed in Seq2Seq tasks.
  • Bug fixed in regression prediction.
  • Bug fixed in loading multiclass classification models when num_labels aren't specified.

0.43.0 - 2020-07-05

Added

  • Added support for custom parameter groups.

Fixed

  • ELECTRA pretraining no longer replaces 10% of masked inputs with random tokens. @dev-sngwn

0.42.0 - 2020-07-05

Added

  • Added better support for Chinese text in Language Modeling. @taranais
  • Added mobilebert for classification, NER, QA, and Seq2Seq. @flozi00

Fixed

  • Fixed incorrect logic when using early_stopping_metric_minimize. @djstrong
  • Fixed issue with cache directory being created even when no_cache is set. @henninglebbaeus

Changed

  • Running loss is now shown next to the tqdm bar (with the tqdm bar description)
  • Removed tokenizers and transformers version pins (added earlier to avoid compatibility issues)

0.41.2 - 2020-07-03

Fixed

  • Fixed bugs with args not being passed correctly to wandb in the following models:
    • MultiModalClassificationModel
    • ConvAIModel
    • Seq2SeqModel
    • T5Model
  • Fixed bugs in Seq2SeqModel and T5Model when not using use_multiprocessed_decoding.

Changed

  • Set use_multiprocessed_decoding=False as default for Seq2Seq models to avoid a bug.

[0.41.1] - 2020-07-02

Fixed

  • Fixed bug where the returned value from MultiModalClassificationModel.evaluate_model() was incorrect.

0.41.0 - 2020-07-02

Added

Changed

  • Added lazy_loading attibute to ClassificationArgs which now controls whether lazy loading is used.
  • Replaced lazy_loading_header_row attribute in ClassificationArgs with lazy_loading_start_line.
  • Unnecessary Docs spacing removed @bryant1410
  • Set required tokenizer version to 0.7 until breaking changes are resolved.

0.40.2 - 2020-06-25

Fixed

  • Fixed bug in Multi-Modal classification when using evaluate_during_training.

0.40.1 - 2020-06-25

Added

  • Added interact_single() method to ConvAIModel. This accepts a message and conversation history (and an optional personality). @Amit80007

Fixed

  • Fixed bug in multi modal classification @tekkon

Changed

0.40.0 - 2020-06-23

Added

  • All Simple Transformers models can now be used with W&B sweeps.
  • eval_model() now logs to wandb in the following models (can be turned off with wandb_log=False)
    • ClassificationModel
    • NERModel
  • Model args can now be specified through the relevant dataclass. (E.g. ClassificationArgs, T5Args, etc.)
  • All model args are now saved and loaded automatically with Simple Transformers models.
  • Multiprocessed decoding support added to Seq2Seq models
  • Custom label support for Classification tasks (except multilabel).
  • NER labels can be given as NERArgs.labels_list (persists through args saving)

Changed

  • Changed NERModel.eval_model() to return usable model_outputs
  • Improved tqdm descriptions in progress bars
  • ConvAIModel arg no_sample renamed to do_sample for consistency

0.34.4 - 2020-06-17

Added

  • Added num_return_sequences, top_k, and top_p args for Seq2SeqModel.

Fixed

Changed

[0.34.3] - 2020-06-13

Fixed

  • Fixed bug in question answering when not using multiprocessing for feature conversion.

[0.34.2] - 2020-06-13

Fixed

  • Fixed bug in sentence-pair task feature conversion.

0.34.1 - 2020-06-12

Fixed

  • Fixed bug in multi-modal classification due to compatibility issues with breaking changes in transformers==2.11.0.

0.34.0 - 2020-06-09

Added

  • Added distributed training support for language model training. @cahya-wirawan
  • Added multiprocessed decoding support for T5 models.

0.33.2 - 2020-06-08

Fixed

  • Fixed bug in adding prefix space. Included longformer in list of models where prefix spaces are added. @guy-mor

0.33.1 - 2020-06-08

Changed

  • Changed the tokenization logic of RoBERTa (and other models using GPT-2 tokenizer) so that a prefix space will be added to input sentences.

0.33.0 - 2020-06-08

Added

  • Added Longformer model support for;
    • Classification
    • NER
    • Seq2Seq tasks. @flozi00

0.32.3 - 2020-06-04

Fixed

  • Fixed compatibility issues with breaking changes in transformers==2.11.0. @fcggamou

0.32.1 - 2020-06-01

Fixed

  • Fixed bug when using output_hidden_states with ClassificationModel. @jpotniec

0.32.0 - 2020-06-01

Added

  • Added Lazy Loading support for classification tasks (except multi-label). (Docs)

0.31.0 - 2020-05-30

Added

  • Added Longformer model support for Language Modeling.

0.30.0 - 2020-05-27

Added

  • Added XLM-RoBERTa support for question answering tasks.
  • Added save_optimizer_and_scheduler (default 1) to global_args which controls whether optimizer and scheduler is saved along with the model. Disabling significantly reduces the disk space used by saved models.

Fixed

  • Bug in XLM tokenizer when preprocessing QA datasets.
  • QuestionAnsweringModel.predict(n_best_size=n) now correctly returns n answers per question (along with n probabilities).

BREAKING CHANGE

  • QuestionAnsweringModel.predict() now returns two lists (a list of dicts with question ids mapped to answers and a list of dicts with question ids mapped to the answer probabilities).

0.29.0 - 2020-05-24

Fixed

  • Fixed issues with training ELECTRA language models from scratch. @aced125 @Laksh1997
  • Fixed bug in save_discriminator() method.

Changed

  • The parallel process count is now limited to 61 by default on Windows systems. @leungi

0.28.10 - 2020-05-23

Added

  • Added more generation/decoding parameters for T5 models.

Fixed

  • Fixed bug with cached features not being used with T5 models.

0.28.9 - 2020-05-19

Fixed

  • Fixed bug where final model was not being saved automatically.

0.28.8 - 2020-05-19

Fixed

  • Fixed bug where some models were not using multiprocessing_chunksize argument.

0.28.7 - 2020-05-19

Fixed

  • Fixed bug in NERModel.predict() method when split_on_space=False. @alexysdussier

0.28.6 - 2020-05-19

Added

  • Added multiprocessing support for Question Answering tasks for substantial performance boost where CPU-bound tasks (E.g. prediction especially with long contexts)
  • Added multiprocessing_chunksize (default 500) to global_args for finer control over chunking. Usually, the optimal value will be (roughly) number of examples / process count.

0.28.5 - 2020-05-18

Added

  • Added no_save option to global_args. Setting this to True will prevent models from being saved to disk.
  • Added minimal training script for Seq2Seq models in the examples directory.

0.28.4 - 2020-05-15

Fixed

  • Fixed potential bugs in loading weights when fine-tuning an ELECTRA language model. Fine-Tuning an ELECTRA language model now requires both model_name and model_type to be set to electra.

0.28.3 - 2020-05-15

Changed

  • Updated Seq2SeqModel to use MarianTokenizer with MarianMT models. @flozi00

0.28.2 - 2020-05-14

Fixed

  • Bug fix for generic Seq2SeqModel

0.28.1 - 2020-05-14

Fixed

  • Bug when training language models from scratch

0.28.0 - 2020-05-11

Added

  • Sequence-to-Sequence task support added. This includes the following models:
    • BART
    • Marian
    • Generic Encoder-Decoder
  • The args dict of a task-specific Simple Transformers model is now saved along with the model. When loading the model, these values will be read and used. Any new args passed into the model initialization will override the loaded values.

0.27.3 - 2020-05-10

Added

  • Support for AutoModel in NER, QA, and LanguageModeling. @flozi00

Fixed

  • Now predict function from NER_Model returns a value model_outputs that contains: A Python list of lists with dicts containing each word mapped to its list with raw model output. @flaviussn

Changed

  • Pillow import is now optional. It only needs to be installed if MultiModal models are used.

0.27.2 - 2020-05-08

Fixed

  • Fixed T5 lm_labels not being masked properly

Changed

  • Torchvision import is now optional. It only needs to be installed if MultiModal models are used.

0.27.1 - 2020-05-05

Fixed

  • Fixed issue with custom evaluation metrics not being handled correctly in MultiLabelClassificationModel. @galtay

0.27.0 - 2020-05-05

Added

  • Added support for T5 Model.
  • Added do_sample arg to language generation.
  • NERModel.predict() now accepts a split_on_space optional argument. If set to False, to_predict must be a a list of lists, with the inner list being a list of strings consisting of the split sequences. The outer list is the list of sequences to predict on.

Changed

  • eval_df argument in NERModel.train_model() renamed to eval_data to better reflect the input format. Added Deprecation Warning.

0.26.1 - 2020-04-27

Fixed

  • Specifying verbose=False in LanguageGenerationModel.generate() method now correctly silences logger output.

0.26.0 - 2020-04-25

Added

  • Added Electra model support for sequence classification (binary, multiclass, multilabel)
  • Added Electra model support for question answering
  • Added Roberta model support for question answering

Changed

  • Reduced logger messages during question answering evaluation

0.25.0 - 2020-04-24

Added

  • Added support for Language Generation tasks.

0.24.9 - 2020-04-22

Added

  • Added support for custom metrics with QuestionAnsweringModel.

Fixed

  • Fixed issue with passing proxies to ConvAI models. @Pradhy729

0.24.8 - 2020-04-13

Fixed

  • Fixed incorrect indexes when extracting hidden layer outputs and embedding outputs with ClassificationModel.predict() method.

0.24.7 - 2020-04-13

Added

  • Added option to get hidden layer outputs and embedding outputs with ClassificationModel.predict() method.
    • Setting config: {"output_hidden_states": True} will automatically return all embedding outputs and hidden layer outputs.

Changed

  • global_args now has a config dictionary which can be used to override default values in the confg class.
    • This can be used with ClassificationModel, MultiLabelClassificationModel, NERModel, QuestionAnsweringModel, and LanguageModelingModel

0.24.6 - 2020-04-12

Added

  • Added support for ELECTRA based NER models.

0.24.5 - 2020-04-11

Fixed

  • Fixed bug in LanguageModelingModel when loading from a training checkpoint.

[0.24.4] - 2020-04-10

Fixed

  • Fixed bug in LanguageModelingModel initialization with a trained tokenizer.

Added

  • Added support for passing proxy information with ConvAI model.

0.24.3 - 2020-04-10

Fixed

  • Fixed potential bug in NERModel predict() method when using custom labels.
  • Fixed typo in the NERModel description in the readme.

0.24.2 - 2020-04-09

Fixed

  • Fixed issues with vocab_size not being set properly in ELECTRA models.

0.24.1 - 2020-04-09

Fixed

  • Fixed bugs in minimal examples for language modeling.

Changed

  • Added vocab_size back to default args dict for clarity. (vocab_size is None by default)
  • Changed error message when training a new tokenizer with incorrect parameters for clarity.

0.24.0 - 2020-04-09

Added

  • Added ELECTRA pretraining support.
  • Added better support for configuring model architectures when training language models from scratch.
    • Any options which should be overriden from the default config can now be specified in the args dict. (config key)

Changed

  • Default entry for vocab_size removed from args for LanguageModelingModel as it differs for different model types.
    • vocab_size must now be specified whenever a new tokenizer is to be trained.

Fixed

  • Fixed bugs when training BERT (with word piece tokenization) language models from scratch.
  • Fixed incorrect special tokens being used with BERT models when training a new tokenizer.
  • Fixed potential bugs with BERT tokenizer training.

0.23.3 - 2020-04-05

Fixed

  • Fixed bug in QuestionAnsweringModel where the save_model() method wasn't being called properly.
  • Fixed bug in calculating global step when resuming training.

0.23.2 - 2020-04-02

Fixed

  • Prevent padding tokens being added when using openai-gpt and gpt2 models for language modeling.

0.23.1 - 2020-03-30

Fixed

  • Fixed bug in binary classification evaluation when data only contains one label.
  • Fixed typo in readme.

Changed

  • Cache dir is no longer created when no_cache is used.

0.23.0 - 2020-03-30

Added

  • Added support for training custom tokenizers.
  • Added improved support for training language models from scratch.
  • Added improved support for resuming training in classification, NER, and QnA tasks.

[0.22.1] - 2020-03-19

Added

  • Added support for XLMRoberta for multi-label tasks.

0.22.0 - 2020-03-14

Added

  • Added support for language model training (from scratch or fine-tuning).
  • Added option to choose which metric should be used for early stopping.

Changed

  • Switched to using the logging module over print for everything except running loss. (QuestionAnsweringModel - @travis-harper)
  • Replaced more concatenated string paths with os.path.join() when creating training_progress_scores.csv.

0.21.5 - 2020-03-12

Changed

  • Replaced concatenated string paths with os.path.join() when creating training_progress_scores.csv. @sarthakTUM

0.21.4 - 2020-03-12

Fixed

  • Fixed issue with cached eval features being used even when using predict() in ClassificationModel and NERModel.

0.21.3 - 2020-03-03

Added

  • Added classification report for NER for per-tag scores. @seo-95

0.21.2 - 2020-03-01

Fixed

  • Fixed bug with empty answers in QuestionAnsweringModel. @jacky18008

0.21.1 - 2020-02-29

Fixed

  • Fixed bug in ConvAIModel where reprocess_input_data and use_cached_eval_features args were ignored.

0.21.0 - 2020-02-29

Added

  • Added support for training Conversational AI models.
  • Added cuda_device parameter to MultiLabelClassificationModel.

Fixed

  • Fixed bug in MultiModalClassificationModel when num_labels is not given.

0.20.3 - 2020-02-22

Changed

  • reprocess_input_data changed to True by default.
  • use_cached_eval_features changed to False by default.

0.20.2 - 2020-02-22

Fixed

  • Fixed issue with early stopping not working with Question Answering.

0.20.1 - 2020-02-22

Fixed

  • Fixed issue with predict() function using cached features.

0.20.0 - 2020-02-21

Added

  • Added support for Multi Modal Classification tasks.

Fixed

  • Fixed missing variable wandb_available in Multilabel Classification.

0.19.9 - 2020-02-18

Fixed

  • Fixed missing variable wandb_available in Multilabel Classification.

0.19.8 - 2020-02-14

Fixed

  • Fixed missing variable wandb_available in Multilabel Classification.

0.19.7 - 2020-02-11

Changed

  • Removed wandb as a dependency. Installing wandb in now optional.

0.19.6 - 2020-02-11

Added

  • Added support for multilabel classification with FlauBERT.@adrienrenaud

0.19.5 - 2020-02-11

Added

  • Added support for FlauBERT with classification tasks (except multi-label).@adrienrenaud

0.19.4 - 2020-02-04

Fixed

  • Fixed error that occured when args is not given when creating a Model.

0.19.3 - 2020-02-03

Added

  • Added manual_seed to global_args . Can be used when training needs to be reproducible.

0.19.2 - 2020-01-31

Added

  • Added early stopping support for NER and Question Answering tasks.

Fixed

  • Fixed issue with nested file paths not being created.
  • wandb_kwargs not being used with NER and Question Answering.

0.19.1 - 2020-01-27

Fixed

  • Fixed issue with evaluation at the end of epochs not being considered for best model.

0.19.0 - 2020-01-26

Added

  • Added early stopping support for Classification tasks.
    • Set use_early_stopping to True to enable.
  • The best model will now be saved to {output_dir}/best_model/ when evaluate_during_training is used.
  • Added evaluate_during_training_verbose to args dict to control whether evaluation during training outputs are printed to console.
  • Added all-contributors to README to recognize contributors.

Changed

  • Evaluation during training no longer displays progress bars.
  • Evaluation during training no longer prints results to console by default.
  • Moved model/results saving logic to _save_model for readability and maintainability.
  • Updated README.

0.18.12 - 2020-01-25

Fixed

  • Added missing extra SEP token in RoBERTa, CamemBERT, and XLMRoBERTA in sentence pair tasks.

0.18.11 - 2020-01-21

Added

  • Added no_cache option to global_args which disables caching (saving and loading) of features to/from disk.

0.18.10 - 2020-01-20

Added

  • Added Makefile with tests dependency installation, test code, formatter and types.
  • Added setup.cfg file with Make configuration
  • Added some tests for the functionality

Changed

  • Files linted using flake8
  • Files formatted using black
  • Test tested with pytest
  • Unused variables deleted

0.18.9 - 2020-01-20

Fixed

  • Fixed bug with importing certain pre-trained models in MultiLabelClassificationModel .

0.18.8 - 2020-01-20

Added

  • Added **kwargs to the init methods of ClassificationModel , MultiLabelClassificationModel , QuestionAnsweringModel , and NERModel . These will be passed to the from_pretrained() method of the underlying model class.

0.18.6 - 2020-01-18

Changed

  • Reverted change made in 0.18.4 (Model checkpoint is no longer saved at the end of the last epoch as this is the same model saved in ouput_dir at the end of training).

Model checkpoint is now saved for all epochs again.

0.18.5 - 2020-01-18

Fixed

  • Fixed bug when using sliding_window .

0.18.4 - 2020-01-17

Fixed

  • Typo in classification_utils.py .

Changed

  • Model checkpoint is no longer saved at the end of the last epoch as this is the same model saved in ouput_dir at the end of training.

0.18.3 - 2020-01-15

Fixed

  • Potential bugfix for CamemBERT models which were giving identical outputs to all inputs.

0.18.2 - 2020-01-15

Added

  • Added option to turn off model saving at the end of every epoch with save_model_every_epoch .

Fixed

  • Fixed bug with missing tensorboard_folder key in certain situations.

Changed

  • Moved args items common to all classes to one place ( config/global_args.py ) for maintainability. Does not make any usage changes.

0.18.1 - 2020-01-15

Fixed

  • Fixed bug with missing regression key when using MultiLabelClassification.

0.18.0 - 2020-01-15

Added

  • Sentence pair tasks are now supported.
  • Regression tasks are now supported.
  • use_cached_eval_features to args . Evaluation during training will now use cached features by default. Set to False if features should be reprocessed.

Changed

  • Checkpoints saved at the end of an epoch now follow the `checkpoint-{global_step}-epoch-{epoch_number} format.

0.17.1 - 2020-01-14

Fixed

  • Fixed wandb_kwargs key missing in args bug.

0.17.0 - 2020-01-14

Added

  • Added new model XLM-RoBERTa. Can now be used with ClassificationModel and NERModel .

0.16.6 - 2020-01-13

Added

  • Added evaluation scores from end-of-epoch evaluation to training_progress_scores.csv .

Fixed

  • Typos in README.md .

0.16.5 - 2020-01-09

Fixed

  • Reverted missed logging commands to print statements.

0.16.4 - 2020-01-09

Changed

  • Removed logging import.

0.16.3 - 2020-01-09

Fixed

  • Reverted to using print instead of logging as logging seems to be causing issues.

0.16.2 - 2020-01-08

Changed

  • Changed print statements to logging.

0.16.1 - 2020-01-07

Added

  • Added wandb_kwargs to args which can be used to specify keyword arguments to wandb.init() method.

0.16.0 - 2020-01-07

Added

  • Added support for training visualization using the W&B framework.
  • Added save_eval_checkpoints attribute to args which controls whether or not a model checkpoint will be saved with every evaluation.

0.15.7 - 2020-01-05

Added

  • Added **kwargs for different accuracy measures during multilabel training.

0.15.6 - 2020-01-05

Added

  • Added train_loss to training_progress_scores.csv (which contains the evaluation results of all checkpoints) in the output directory.

0.15.5 - 2020-01-05

Added

  • Using evaluate_during_training now generates training_progress_scores.csv (which contains the evaluation results of all checkpoints) in the output directory.

0.15.4 - 2019-12-31

Fixed

  • Fixed bug in QuestonAnsweringModel when using evaluate_during_training .

0.15.3 - 2019-12-31

Fixed

  • Fixed bug in MultiLabelClassificationModel due to tensorboard_dir being missing in parameter dictionary.

Changed

  • Renamed tensorboard_folder to tensorboard_dir for consistency.

0.19.8 - 2020-02-14

Fixed

  • Fixed missing variable wandb_available in Multilabel Classification.

Added

  • Added tensorboard_folder to parameter dictionary which can be used to specify the directory in which the tensorboard files will be stored.

0.15.1 - 2019-12-27

Added

  • Added **kwargs to support different accuracy measures at training time.

0.15.0 - 2019-12-24

Added

  • Added evaluate_during_training_steps parameter that specifies when evaluation should be performed during training.

Changed

  • A model checkpoint will be created for each evaluation during training and the evaluation results will be saved along with the model.

0.14.0 - 2019-12-24

Added

  • Added option to specify a GPU to be used when multiple GPUs are available. E.g.: cuda_device=1
  • Added do_lower_case argument for uncased models.

Fixed

  • Fixed possible bug with output directory not being created before evaluation is run when using evaluate_during_training .

0.13.4 - 2019-12-21

Fixed

  • Fixed bug with when using eval_during_training with QuestionAnswering model.

0.13.3 - 2019-12-21

Fixed

  • Fixed bug with loading Multilabel classification models.
  • Fixed formatting in README.md.

0.13.2 - 2019-12-20

Fixed

  • Fixed formatting in README.md.

0.13.1 - 2019-12-20

Fixed

  • Bug in Multilabel Classification due to missing entries in default args dict.

0.13.0 - 2019-12-19

Added

  • Sliding window feature for Binary and Multiclass Classification tasks.

0.12.0 - 2019-12-19

Added

  • Minimal examples have been added to the examples directory as Python scripts.

Changed

  • Readme updated to include the addition of examples.

0.11.2 - 2019-12-18

Fixed

  • Evaluation during training fixed for multilabel classification.

0.11.1 - 2019-12-18

Fixed

  • Broken multiprocessing support for NER tasks fixed.

0.11.0 - 2019-12-15

Added

  • CamemBERT can now be used with NERModel.

Changed

  • Readme changed to include CamemBERT for NER.

0.10.8 - 2019-12-15

Added

  • DistilBERT can now be used with NERModel.

Changed

  • Readme changed to include DistilBERT for NER.

0.10.7 - 2019-12-15

Added

  • This CHANGELOG file to hopefully serve as an evolving example of a standardized open source project CHANGELOG.