aisingapore / sgnlp Goto Github PK

Include Demo API from the refactored ASGCN from #41

To allow the containers to start up quicker, models and files were downloaded and cached during build time.

Recent changes in the huggingface transformers package has broken this functionality:

• Released in v4.22.0
• Issue

Possible choices moving forward:

• Write a simple caching utility function
• Stick to versions of transformers before 4.22.0

See issues reported here,
https://epoch.aisingapore.org/community/sg-nlp/how-do-we-train-senticgcn-models-with-our-dataset/

The documentation only listed the trainer code for executing the trainer function but did not indicate that the config file should be pass via the --config flag.

Update documentation to indicate how to pass the config file to the trainer.

• Using multiple workers of the default type 'sync' in gunicorn is not working on Kubernetes
• Workers constantly terminated due to signal 9
• Try gevent to see if it works out

• Deploy to GKE

• Create a function or class that handles common setup between the demo apis

Refactor Sentic-GCN code base from https://github.com/BinLiang-NLP/Sentic-GCN to SGNLP

This task is for migrating all pretrained weights from the sgnlp GCP storage bucket to use the sgnlp-models storage bucket.

• Documentation
• Tests
• Model card

BiERU: Bidirectional Emotional Recurrent Unit for Conversational Sentimental Analysis by Li et al. 2021

https://arxiv.org/abs/2006.00492

• Original LIF code uses allennlp framework: https://github.com/nusnlp/LIF
• Earlier attempts of reimplementing the model in pytorch led to worse results
• For now, we will retain the original code and framework used to train it
• Add a class with a from_pretrained method such that usage experience will be similar to other models

• Remove relevant code in package
• Remove relevant documentation

Paper: https://aclanthology.org/2020.emnlp-main.108.pdf

Prof: Jiang Jing from SMU

Repo: GitHub - jefferyYu/DualHierarchicalTransformer: Predicting Stance and Rumor Veracity via Dual Hierarchical Transformer

• Move tests to github actions

refactor RST model to add to sgnlp

valset_ratio needs to be >=0 and <1 in dataclass.py

The documentation's initial page does not elaborate much on what exists in the documentation.

Further elaboration on what is available can help users get started.

• Add elaboration on side tabs (quickstart, advanced tutorial, api reference)

The current implementation of preprocessor matches a single aspect index for the purpose of matching postprocessor output. The aspect index field for process_input payload should be expended to handle aspects with multiple indexes.

Latest SenticGCN implementation for the Dev branch.
In the dataclass.py, post_init method in SenticGCNTrainArgs, there are the following assertions,

assert self.repeats > 1, "Repeats value must be at least 1."
assert self.patience > 1, "Patience value must be at least 1." 

The comparison operator should be >= instead.

• Trained segmenter model have similarly good metrics but performs very differently
• Check commits between segmenter-experiment-4 and 5

• Add tests
• Fix bug that happens on certain inputs
• Ensure training is stable and replicable
• Refine documentation

• Create a base logging file to utils folder
• Replace all logging using the new logger

Write a contributing guide

• Add additional tutorial on how to use UFD to train and evaluate on custom dataset
• Bug fix for UFD parse_args_and_load_config util function

• Create a module named sgnlp.base
• Add abstractmethods for preprocess, save, load
• Add batch iteration to parent __call__
• Parent __call__ should return a dictionary

• Change demo api output from a fraction to percentage
• Change module file names to lowercase
• Remove dependency on the datasets package