Welcome to the rnn_benchmarks repository! We offer:
- A training speed comparison of different LSTM implementations across deep learning frameworks
- Common input sizes, network configurations and cost functions from automatic speech recognition
- Best-practice scripts to learn coding up a network, optimizers, loss functions etc.
- Arxiv paper: LSTM Benchmarks for Deep Learning Frameworks
- LSTM benchmarks between PyTorch 0.4, TensorFlow 1.8, Keras 2.1.6 and latest Lasagne
Go to the folder 'main' and execute the 'main.py' script in the corresponding benchmark folder. Before running 'main.py', you need to give the paths to the python environment that contain the corresponding framework. The 'main.py' script creates a 'commands.sh' script that will execute the benchmarks. The measured execution times will be written to 'results/results.csv'. The toy data and default parameters are provided by 'support.py', to make sure every script uses the same hyperparameters.