sujay-pandit / rtl2embeddings

Learning semantic representation of design for verification

This repository contains code that replicates some of the experiments presented in Learning Semantic Reprsentation to Verify Design. Note that this is an unofficial implementation that does not have anything to do with the authors of above paper.

How to

Setup dependencies

Base dependencies

This step will install commonly required dependencies for using this repository.

# Pull submodules
git submodule update --init --recursive

# Install python dependencies (tested with python 3.6)
pip install -r requirements.txt

Generate data

1. See this document to parse design and generate coverage information.
2. Perform this step to generate data for NN training and evaluation.

python nn/datagen.py
  --graph_dir $DATA_DIR/generated/d2v_data/graph \ # See ./docs/data_generation.md to generate this
  --tp_cov_dir $DATA_DIR/generated/d2v_data/tp_cov \ # See ./docs/data_generation.md to generate this
  --tf_data_dir $DATA_DIR/nn_dataset # Where to output the final data.

Train model

After all data is prepared, you can train a model.

python nn/train.py \
  --graph_dir $DATA_DIR/generated/d2v_data/graph \ # See ./docs/data_generation.md to generate this
  --tp_cov_dir $DATA_DIR/generated/d2v_data/tp_cov \ # See ./docs/data_generation.md to generate this
  --batch_size 256 \ # Adjust this if you run into an OOM error.
  --split_ratio $SPLIT \ # Comma separated numbers that add up to 1 (e.g. 0.7,0.2,0.1)
  --seed 123 \ # Random seed for reproducibility
  --append_seed_to_ckpt_dir \ # This will add seed information to the checkpoint directory name
  --ckpt_dir $DATA/ckpts/split_$SPLIT/ \  # Where to save model checkpoints
  --tf_data_dir $DATA_DIR/nn_dataset \
  --use_attention  # Set if you want to use attention in the model