PyTorch model tracing tutorial

• Using JIT compile support since PyTorch 1.0.0

Requirements

• python == 3.7.0
• torch == 1.0.0
• numpy == 1.16.0

Preparation

1. Upgrade PyTorch

pip install --upgrade pip
pip install --upgrade torch
pip install --upgrade numpy

Tracing

refer trace_benchmark_DNN/benchmark.py or trace_trained_RNN/trace.py

1. training a model in PyTorch
2. create a "tracer" which is an input tensor of the model
3. trace model by passing through the tracer to the model with torch.jit.trace(your_model, tracer)

overall process is executed by followings

git clone https://github.com/hyoungseok/jitTutorial.git
cd jitTutorial/trace_trained_RNN
python main.py

the traced file traced_model.pth will be created under the jitTutorial/trace_trained_RNN

Contributor

• YongRae Jo (dreamgonfly@gmail.com)

Author

• Alex Kim (hyoungseok.k@gmail.com)