manisnesan/fastchai

Repository capturing deep learning & nlp experiments using fastai & pytorch

FastChai

Outline

• Computer Vision
• Natural Language Processing
  • Sequence Classification
  • Question Answering
• MLOps
  • ML Monitoring

Experiments using fastai, transformers, evidently for ml-monitoring

Computer Vision

ImageWoof

• Baseline using fastai defaults with Resnet34
• Fit OneCycle using Resnet34 with LearningRate(LR) Finder, Discriminative LR and Mixed Precision
• MixUp, a type of DataAugmentation
• Labelsmoothing as regularization
• Test Time Augmentation
• ranger as an optimizer, a mix of LookAhead and RAdam instead of the default AdamW
• Native Mixed Precision

MNIST

• Resnet from scratch

Paddy - Road to the top series

• Baseline
• Small models

Natural Language Processing

HASOC - Hate Speech and Offensive Content [en]

• ULMFiT Approach using Spacy with Gradual Unfreezing, Weighted Cross Entropy + Label Smoothing for handling imbalanced data

Diaster Tweet [en]

• Demonstrates how to use Cleanlab to detect and remove noisy labels and train a classifier using ULMFiT Approach (Accuracy change : 0.79 -> 0.85)

Jigsaw Toxicity Classification [en]

• Demonstrates how to use blurr, library that integrates HF transformers with fastai.

IMDB Movie Review Sentiment Classification[en]

• Another example demonstrating how to use blurr MidLevel API and DistilRoberta for Sequence Classification.

MLOps

ML Monitoring

• Data Drift & Model Classification Performance for Iris dataset