About

I primarily used TensorFlow for large-scale recommendation tasks when in big company, but PyTorch could be more efficient for smaller tasks in a smaller company.

This directory aims to train a Click-Through Rate (CTR) model using PyTorch. It's a simple example, seeking to keep everything minimal.

Tips:

1. Apply polars instead of pandas to process data. I first implemented the pandas version of FeatureTransformer, but it's very slow when datasize> 20 millon.
2. Apply parquets instead of pickle files to save data samples. It's really saved my memory.

Features:

• Fast FeatureTransformer for DNN (tested on >100millon rows on single machine). Which support:
  • Categorical: automatic vocabulary extraction, low-frequency filtering, dynamic embedding, hash embedding
  • Numerical: standard/0-1 normalization, automatic discretization, automatic update of statistical number for standard/0-1 normalization when multiple times of fitting.
  • Variable-length sequence feature: there's order in the sequence, if want to padding when feature transformer, please put the latest data before the oldest data as it may pads at the end of the sequence by default.
  • More feature configurations...
• Implemented a common Trainer for training pytorch models, and save/load the results
• Basic FastAPI for Model API Serving

Install

pip install git+https://github.com/xiahouzuoxin/torchctr

And follow the example to use.

Model API Serving

An simple serving without latency considered.

1. [Optional] According to your model and data processing, maybe need create a new ServingModel like BaseServingModel

2. Set up the service:

   • Debuging: Given service name and model path from command line

     cd $torchctr_root
     python -m torchctr.serving.serve --name [name] --path [path/to/model or path/to/ckpt] --serving_class BaseServingModel
     

   • Production: write the command line parameters to serving_models variable in torchctr/serving/serve.py

3. Test the service: reference test_predict in example