TODO.md

TODO list

Overall

• Add more options with Jinja ? For example --without-tensorflow would remove all references to TensorFlow. Not easy to test.
• Add other templates ? Such as unsupervised templates.
• Fix all TODOs
• Add a note in the README.md about torch packages management (cf. https://stackoverflow.com/questions/57689387/equivalent-for-find-links-in-setup-py)
• Improve 'packages' option in all setup.py
• Fix all linters

All templates

• Review how some models' parameters are managed. For example, the models' learning rate should probably be an init argument.
• Are we checking all parameters the same way ?
• Add ROC curves in plots metrics.
• Add a "model" that would aggregate several models (weak learners) into a meta model.
• We should probably remove nb_iter_keras stuff. It is not used.
• Review how the get_classes_from_proba and inverse_transform functions work! It is not very understandable!
• Add used preprocessing as models' attributes.
• Remove / fix some # type: ignore.
• Rework models folder hierarchy. We should add a directory per library (keras, sklearn, ...), and exploratory subdirectories.
• Some 0_....py scripts are not tested (in functional tests).

Template - NLP

• Should we keep the with_new_embedding argument ? It is not used.
• We should probably mutualise all keras models' predict_proba functions in model_keras.py. Only ModelTfidfDense is different, it should overload this method.
• Rework model_rules.py.
• Download flaubert_small_cased before using tests (in Actions)
• Installation error with python setup.py develop when requirements.txt is not called first -> error: requests 2.28.0 is installed but requests<2.25.1,>=2.23.0 is required by {'words-n-fun'}
• Shouldn't transformers be saved into XXX-data folder ? Like detectron models for the vision template ? (If changed, also change tutorial)

Template - Numerical

Template - Computer Vision

• Many unit test to add:
  • read_folder
  • read_folder_object_detection
  • rebuild_metadata_object_detection
  • rebuild_metadata_classification
  • Many utils_object_detector functions
  • utils_faster_rcnn tests
  • etc.
• Implement predict for object detection tasks
• Check if we can remove all OOM skips (e.g. in functional tests)