fast-deep-rnn 0.0.2

Deep Learning framework extension allowing more efficient backpropogation of gradient in a situation with branched computational graph structure

fast-deep-rnn

This is the Course Project for the DeepLearning University Course.

Install

Library can be installed from the PyPI via

pip install fast_deep_rnn

Structure

core module contains the original core of the framework with Tensor class implementation and the set of differentiable operations, organized in Modules.

core_v2 module contains the alternative proposed implementation, resulting in much faster gradient computing in RNN-s.

Notebook nbs/02_minimal_training.ipynb contains the simplest example of model, having exponential growth in original gradient computing, and benchmarking function to measure this growth. git tags baseline_benchmark_results and solution_benchmark_results contain corresponding benchmark results inside the notebook.

Notebook nbs/01_lstm_training.ipynb contains training of LSTM on number sorting task, which became possible only after the implemented optimization.