MGeo is a multi-modal multi-task geographic language model.
We support 5 pipeline tasks and 1 pretrained model MGeo on maas.
In the same time, we propose GeoGLUE, a geographic evaluation benchmark. MGeo can be finetuned on GeoGLUE tasks.
Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/11273012
* add prov city dist feature to gis encoder
* finish mgeo fintune and pipeline
* text classification add token type id
* to_device support ModelOutput class
* update token classification model lable mask logic
1. Abstract keys of dicts needed by nlp metric classes into the init method
2. Add Preprocessor.save_pretrained to save preprocessor information
3. Abstract the config saving function, which can lead to normally saving in the direct call of from_pretrained, and the modification of cfg one by one when training.
4. Remove SbertTokenizer and VecoTokenizer, use transformers' tokenizers instead
5. Use model/preprocessor's from_pretrained in all nlp pipeline classes.
6. Add model_kwargs and preprocessor_kwargs in all nlp pipeline classes
7. Add base classes for fill-mask and text-classification preprocessor, as a demo for later changes
8. Fix user feedback: Re-train the model in continue training scenario
9. Fix user feedback: Too many checkpoint saved
10. Simplify the nlp-trainer
11. Fix user feedback: Split the default trainer's __init__ method, which makes user easier to override
12. Add safe_get to Config class
---------------------------- Another refactor from version 36 -------------------------
13. Name all nlp transformers' preprocessors from TaskNamePreprocessor to TaskNameTransformersPreprocessor, for example:
TextClassificationPreprocessor -> TextClassificationTransformersPreprocessor
14. Add a base class per task for all nlp tasks' preprocessors which has at least two sub-preprocessors
15. Add output classes of nlp models
16. Refactor the logic for token-classification
17. Fix bug: checkpoint_hook does not support pytorch_model.pt
18. Fix bug: Pipeline name does not match with task name, so inference will not succeed after training
NOTE: This is just a stop bleeding solution, the root cause is the uncertainty of the relationship between models and pipelines
Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10723513
* add save_pretrained to preprocessor
* save preprocessor config in hook
* refactor label-id mapping fetching logic
* test ok on sentence-similarity
* run on finetuning
* fix bug
* pre-commit passed
* fix bug
* Merge branch 'master' into feat/refactor_config
# Conflicts:
# modelscope/preprocessors/nlp/nlp_base.py
* add params to init
* 1. support max ckpt num 2. support ignoring others but bin file in continue training 3. add arguments to some nlp metrics
* Split trainer init impls to overridable methods
* remove some obsolete tokenizers
* unfinished
* support input params in pipeline
* fix bugs
* fix ut bug
* fix bug
* fix ut bug
* fix ut bug
* fix ut bug
* add base class for some preprocessors
* Merge commit '379867739548f394d0fa349ba07afe04adf4c8b6' into feat/refactor_config
* compatible with old code
* fix ut bug
* fix ut bugs
* fix bug
* add some comments
* fix ut bug
* add a requirement
* fix pre-commit
* Merge commit '0451b3d3cb2bebfef92ec2c227b2a3dd8d01dc6a' into feat/refactor_config
* fixbug
* Support function type in registry
* fix ut bug
* fix bug
* Merge commit '5f719e542b963f0d35457e5359df879a5eb80b82' into feat/refactor_config
# Conflicts:
# modelscope/pipelines/nlp/multilingual_word_segmentation_pipeline.py
# modelscope/pipelines/nlp/named_entity_recognition_pipeline.py
# modelscope/pipelines/nlp/word_segmentation_pipeline.py
# modelscope/utils/hub.py
* remove obsolete file
* rename init args
* rename params
* fix merge bug
* add default preprocessor config for ner-model
* move a method a util file
* remove unused config
* Fix a bug in pbar
* bestckptsaver:change default ckpt numbers to 1
* 1. Add assert to max_epoch 2. split init_dist and get_device 3. change cmp func name
* Fix bug
* fix bug
* fix bug
* unfinished refactoring
* unfinished
* uw
* uw
* uw
* uw
* Merge branch 'feat/refactor_config' into feat/refactor_trainer
# Conflicts:
# modelscope/preprocessors/nlp/document_segmentation_preprocessor.py
# modelscope/preprocessors/nlp/faq_question_answering_preprocessor.py
# modelscope/preprocessors/nlp/relation_extraction_preprocessor.py
# modelscope/preprocessors/nlp/text_generation_preprocessor.py
* uw
* uw
* unify nlp task outputs
* uw
* uw
* uw
* uw
* change the order of text cls pipeline
* refactor t5
* refactor tg task preprocessor
* fix
* unfinished
* temp
* refactor code
* unfinished
* unfinished
* unfinished
* unfinished
* uw
* Merge branch 'feat/refactor_config' into feat/refactor_trainer
* smoke test pass
* ut testing
* pre-commit passed
* Merge branch 'master' into feat/refactor_config
# Conflicts:
# modelscope/models/nlp/bert/document_segmentation.py
# modelscope/pipelines/nlp/__init__.py
# modelscope/pipelines/nlp/document_segmentation_pipeline.py
* merge master
* unifnished
* Merge branch 'feat/fix_bug_pipeline_name' into feat/refactor_config
* fix bug
* fix ut bug
* support ner batch inference
* fix ut bug
* fix bug
* support batch inference on three nlp tasks
* unfinished
* fix bug
* fix bug
* Merge branch 'master' into feat/refactor_config
# Conflicts:
# modelscope/models/base/base_model.py
# modelscope/pipelines/nlp/conversational_text_to_sql_pipeline.py
# modelscope/pipelines/nlp/dialog_intent_prediction_pipeline.py
# modelscope/pipelines/nlp/dialog_modeling_pipeline.py
# modelscope/pipelines/nlp/dialog_state_tracking_pipeline.py
# modelscope/pipelines/nlp/document_segmentation_pipeline.py
# modelscope/pipelines/nlp/faq_question_answering_pipeline.py
# modelscope/pipelines/nlp/feature_extraction_pipeline.py
# modelscope/pipelines/nlp/fill_mask_pipeline.py
# modelscope/pipelines/nlp/information_extraction_pipeline.py
# modelscope/pipelines/nlp/named_entity_recognition_pipeline.py
# modelscope/pipelines/nlp/sentence_embedding_pipeline.py
# modelscope/pipelines/nlp/summarization_pipeline.py
# modelscope/pipelines/nlp/table_question_answering_pipeline.py
# modelscope/pipelines/nlp/text2text_generation_pipeline.py
# modelscope/pipelines/nlp/text_classification_pipeline.py
# modelscope/pipelines/nlp/text_error_correction_pipeline.py
# modelscope/pipelines/nlp/text_generation_pipeline.py
# modelscope/pipelines/nlp/text_ranking_pipeline.py
# modelscope/pipelines/nlp/token_classification_pipeline.py
# modelscope/pipelines/nlp/word_segmentation_pipeline.py
# modelscope/pipelines/nlp/zero_shot_classification_pipeline.py
# modelscope/trainers/nlp_trainer.py
* pre-commit passed
* fix bug
* Merge branch 'master' into feat/refactor_config
# Conflicts:
# modelscope/preprocessors/__init__.py
* fix bug
* fix bug
* fix bug
* fix bug
* fix bug
* fixbug
* pre-commit passed
* fix bug
* fixbug
* fix bug
* fix bug
* fix bug
* fix bug
* self review done
* fixbug
* fix bug
* fix bug
* fix bugs
* remove sub-token offset mapping
* fix name bug
* add some tests
* 1. support batch inference of text-generation,text2text-generation,token-classification,text-classification 2. add corresponding UTs
* add old logic back
* tmp save
* add tokenize by words logic back
* move outputs file back
* revert veco token-classification back
* fix typo
* Fix description
* Merge commit '4dd99b8f6e4e7aefe047c68a1bedd95d3ec596d6' into feat/refactor_config
* Merge branch 'master' into feat/refactor_config
# Conflicts:
# modelscope/pipelines/builder.py
1. Add F1 score to sequence classification metric
2. Fix a bug that the evaluate method in trainer does not support a pure pytorch_model.bin
3. Fix a bug in evaluation of veco trainer
4. Add some tips if lr_scheduler in the trainer needs a higher version torch
5. Add some comments
Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10532230
Features:
1. Refactor the directory structure of nlp models. All model files are placed into either the model folder or the task_model folder
2. Refactor all the comments to google style
3. Add detail comments to important tasks and nlp models, to list the description of the model, and its preprocessor&trainer
4. Model Exporting now supports a direct all to TorchModelExporter(no need to derive from it)
5. Refactor model save_pretrained method to support direct running(independent from trainer)
6. Remove the judgement of Model in the pipeline base class, to support outer register models running in our pipelines
7. Nlp trainer now has a NLPTrainingArguments class , user can pass arguments into the dataclass, and use it as a normal cfg_modify_fn, to simplify the operation of modify cfg.
8. Merge the BACKBONES and the MODELS, so user can get a backbone with the Model.from_pretrained call
9. Model.from_pretrained now support a task argument, so user can use a backbone and load it with a specific task class.
10. Support Preprocessor.from_pretrained method
11. Add standard return classes to important nlp tasks, so some of the pipelines and the models are independent now, the return values of the models will always be tensors, and the pipelines will take care of the conversion to numpy and the following stuffs.
12. Split the file of the nlp preprocessors, to make the dir structure more clear.
Bugs Fixing:
1. Fix a bug that lr_scheduler can be called earlier than the optimizer's step
2. Fix a bug that the direct call of Pipelines (not from pipeline(xxx)) throws error
3. Fix a bug that the trainer will not call the correct TaskDataset class
4. Fix a bug that the internal loading of dataset will throws error in the trainer class
Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10490585
1. Fix bugs in daily test
2. Fix a bug that the updating of lr is before the first time of updating of optimizer
TODO this will still cause warnings when GA is above 1
3. Remove the judgement of mode in text-classification's preprocessor to fit the base trainer(Bug)
Update some regression bins to fit the preprocessor
4. Update the regression tool to let outer code modify atol and rtol
5. Add the default metric for text-classification task
6. Remove the useless ckpt conversion method in bert to avoid the requirement of tf when loading modeling_bert
Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10430764
