Skip to content

Latest commit

 

History

History
69 lines (57 loc) · 3.73 KB

index.md

File metadata and controls

69 lines (57 loc) · 3.73 KB
Raw

Welcome to pytorch-lifestream docs

Library content

Here is a brief overview of library with links to the detailed descriptions.

Library modules:

  • ptls.preprocessing - transforms data to ptls-compatible format with pandas or pyspark. Categorical encoding, datetime transformation, numerical feature preprocessing.

  • ptls.data_load - all that you need for prepare your data to training and validation.

    • ptls.data_load.datasets - PyTorch Dataset API implementation for data access.
    • ptls.data_load.iterable_processing - generator-style filters for data transformation.
    • ptls.data_load.augmentations - functions for data augmentation.

  • ptls.frames - tools for training encoders with popular frameworks like CoLES, SimCLR, CPC, VICReg, ...

    • ptls.frames.coles - Contrastive leaning on sub-sequences.
    • ptls.frames.cpc - Contrastive learning for future event state prediction.
    • ptls.frames.bert - methods, inspired by NLP and transformer models.
    • ptls.framed.supervised - modules fo supervised training.
    • ptls.frames.inference - inference module.

  • ptls.nn - layers for model creation:

    • ptls.nn.trx_encoder - layers to produce the representation for a single transactions.
    • ptls.nn.seq_encoder - layers for sequence processing, like RNN of Transformer.
    • ptls.nn.pb - PaddedBatch compatible layers, similar to torch.nn modules, but works with ptls-data.
    • ptls.nn.head - composite layers for final embedding transformation.
    • ptls.nn.seq_step.py - change the sequence along the time axis.
    • ptls.nn.binarization, ptls.nn.normalization - other groups of layers.

How to guide

  1. Prepare your data.
    • Use Pyspark in local or cluster mode for big dataset and Pandas for small.
    • Split data into required parts (train, valid, test, ...).
    • Use ptls.preprocessing for simple data preparation.
    • Transform features to compatible format using Pyspark or Pandas functions. You can also use ptls.data_load.preprocessing for common data transformation patterns.
    • Split sequences to ptls-data format with ptls.data_load.split_tools. Save prepared data into Parquet format or keep it in memory (Pickle also works).
    • Use one of the available ptls.data_load.datasets to define input for the models.
  2. Choose framework for encoder train.
    • There are both supervised of unsupervised frameworks in ptls.frames.
    • Keep in mind that each framework requires its own batch format. Tools for batch collate can be found in the selected framework package.
  3. Build encoder.
    • All parts are available in ptls.nn.
    • You can also use pretrained layers.
  4. Train your encoder with selected framework and pytorch_lightning.
    • Provide data with one of the DataLoaders that is compatible with selected framework.
    • Monitor the progress on tensorboard.
    • Optionally tune hyperparameters.
  5. Save trained encoder for future use.
    • You can use it as single solution (e.g. get class label probabilities).
    • Or it can be a pretrained part of other neural network.
  6. Use encoder in your project.
    • Run predict for your data and get logits, probas, scores or embeddings.
    • Use ptls.data_load and ptls.data_load.datasets tools to keep your data transformation and collect batches for inference.

How to create your own components

It is possible create specific component for every library modules. Here are the links to the detailed description:

  • ptls.data_load - new data processing tools and datasets for new types of data. Link TBD
  • ptls.frames - new modules which train network for specific problems. Link TBD
  • New layers for ptls.nn. Link TBD