How nyu-mlda/abc-rl works
ABC-RL trains and evaluates reinforcement-learning agents that optimize logic-synthesis sequences for benchmark AIG circuits, targeting area-delay improvement over a resyn2 baseline. The user supplies a benchmark split and technology library; the training entry point coordinates circuit environments, MCTS experience collection, replay, and policy updates. The repository has ICLR GCN and DAGNN model variants. The README also documents pretrained evaluation and synergy/embedding analyses. Some wiring is supported only by the SOURCE INDEX; unsampled implementation details are not inferred.
Main entry points: src/main.py ICLR runner · src/main.py DAGNN runner · LogicSynthesisPolicy.py DAGNN policy · evaluate_synergy.py ICLR evaluation
Components 31
Training and search
- ICLR runner
ABC-RL_ICLR/src/main.py - DAGNN runner
ABC-RL_DAGNN/src/main.py - ICLR MCTS
ABC-RL_ICLR/src/mcts.py - DAGNN MCTS
ABC-RL_DAGNN/src/mcts.py - ICLR replay memory
ABC-RL_ICLR/src/replay_memory.py - DAGNN replay memory
ABC-RL_DAGNN/src/replay_memory.py - ICLR trainer
ABC-RL_ICLR/src/trainer.py - DAGNN trainer
ABC-RL_DAGNN/src/trainer.py
Circuit environment
- ICLR synthesis env
ABC-RL_ICLR/src/LogicSynthesisEnv.py - DAGNN synthesis env
ABC-RL_DAGNN/src/LogicSynthesisEnv.py - ICLR static env
ABC-RL_ICLR/src/static_env.py - DAGNN static env
ABC-RL_DAGNN/src/static_env.py - DAG utilities
ABC-RL_DAGNN/src/dag_utils.py
Graph policies
- ICLR GCN policy
ABC-RL_ICLR/src/LogicSynthesisPolicy.py - DAGNN policy
ABC-RL_DAGNN/src/LogicSynthesisPolicy.py - GCN convolution
ABC-RL_DAGNN/src/gcn_conv.py - GAT convolution
ABC-RL_DAGNN/src/gat_conv.py - DeepSets convolution
ABC-RL_DAGNN/src/deepset_conv.py - Gated-sum convolution
ABC-RL_DAGNN/src/gated_sum_conv.py - MLP layers
ABC-RL_DAGNN/src/mlp.py
Evaluation and analysis
- FR logits metrics
ABC-RL_DAGNN/src/basic_FR_logits.py - FR logits metrics
ABC-RL_ICLR/src/basic_FR_logits.py - Energy scoring
ABC-RL_ICLR/src/energy.py - Softmax scoring
ABC-RL_ICLR/src/softmax_response.py - ICLR evaluation
ABC-RL_ICLR/src/evaluate_synergy.py - DAGNN evaluation
ABC-RL_DAGNN/src/evaluate_pretrained.py - Embedding analysis
ABC-RL_DAGNN/src/generateTSNEplot.py
Other components
- Experimenter
- AIG benchmarks
arithmetic - ABC synthesis
- Tech library
lib
Connections 33
- Experimenter → ICLR runner: starts training
README.md - Experimenter → DAGNN runner: starts training
README.md - AIG benchmarks → ICLR runner: provides circuits
README.md - AIG benchmarks → DAGNN runner: provides circuits
README.md - Tech library → ICLR runner: configures synthesis
README.md - Tech library → DAGNN runner: configures synthesis
README.md - ICLR runner → ICLR trainer: creates trainer
ABC-RL_ICLR/src/main.py - ICLR runner → ICLR replay memory: creates memory
ABC-RL_ICLR/src/main.py - ICLR runner → ICLR synthesis env: runs episodes
ABC-RL_ICLR/src/main.py - ICLR runner → ICLR MCTS: uses search
ABC-RL_ICLR/src/main.py - DAGNN runner → DAGNN trainer: creates trainer
ABC-RL_DAGNN/src/main.py - DAGNN runner → DAGNN replay memory: creates memory
ABC-RL_DAGNN/src/main.py - DAGNN runner → DAGNN synthesis env: runs episodes
- DAGNN runner → DAGNN MCTS: uses search
- ICLR trainer → ICLR GCN policy: constructs policy
ABC-RL_ICLR/src/trainer.py - DAGNN trainer → DAGNN policy: constructs policy
ABC-RL_DAGNN/src/trainer.py - ICLR synthesis env → ABC synthesis: executes synthesis
ABC-RL_ICLR/src/LogicSynthesisEnv.py - ICLR synthesis env → ICLR static env: imports dynamics
ABC-RL_ICLR/src/LogicSynthesisEnv.py - DAGNN synthesis env → DAGNN static env: imports dynamics
ABC-RL_DAGNN/src/LogicSynthesisEnv.py - DAGNN synthesis env → DAG utilities: uses graph utilities
ABC-RL_DAGNN/src/LogicSynthesisEnv.py - DAGNN policy → DAG utilities: uses graph utilities
ABC-RL_DAGNN/src/LogicSynthesisPolicy.py - DAGNN policy → GCN convolution: uses convolution
ABC-RL_DAGNN/src/LogicSynthesisPolicy.py - DAGNN policy → GAT convolution: uses convolution
ABC-RL_DAGNN/src/LogicSynthesisPolicy.py - DAGNN policy → DeepSets convolution: uses convolution
ABC-RL_DAGNN/src/LogicSynthesisPolicy.py - DAGNN policy → Gated-sum convolution: uses convolution
ABC-RL_DAGNN/src/LogicSynthesisPolicy.py - DeepSets convolution → MLP layers: uses layers
ABC-RL_DAGNN/src/deepset_conv.py - ICLR evaluation → ICLR runner: evaluates agent
ABC-RL_ICLR/src/evaluate_synergy.py - DAGNN evaluation → ICLR runner: evaluates trained agent
ABC-RL_DAGNN/src/evaluate_pretrained.py - Embedding analysis → ICLR GCN policy: extracts embeddings
ABC-RL_DAGNN/src/generateTSNEplot.py - FR logits metrics → ICLR evaluation: scores logits
- FR logits metrics → ICLR evaluation: scores logits
- Energy scoring → ICLR evaluation: scores logits
- Softmax scoring → ICLR evaluation: scores logits