| import gradio as gr | |
| import pandas as pd | |
| import circuit_transformer as ct | |
| circuit_transformer = ct.CircuitTransformer() | |
| aiger_str = """aag 33 8 0 2 25 | |
| 2\n4\n6\n8\n10\n12\n14\n16\n58\n67 | |
| 18 13 16\n20 19 7\n22 21 15\n24 3 9\n26 25 11 | |
| 28 27 17\n30 3 6\n32 29 31\n34 29 32\n36 23 35 | |
| 38 7 36\n40 10 29\n42 41 32\n44 13 15\n46 42 45 | |
| 48 47 21\n50 39 49\n52 4 45\n54 25 53\n56 54 5 | |
| 58 51 57\n60 45 12\n62 18 61\n64 63 19\n66 48 64 | |
| """ | |
| def predict(input_text, mcts_playouts): | |
| if input_text[-1] != '\n': | |
| input_text += '\n' | |
| try: | |
| aig, info = ct.read_aiger(aiger_str=input_text) | |
| except Exception as e: | |
| raise gr.Error("Aiger format error!") | |
| num_inputs, num_outputs = info[1], info[3] | |
| if not isinstance(aig, list): | |
| aig = [aig] | |
| if num_inputs > 8: | |
| raise gr.Error("Number of inputs should <= 8") | |
| if num_outputs > 2: | |
| raise gr.Error("Number of outputs should <= 2") | |
| mcts_playouts = int(mcts_playouts) | |
| optimized_aigs = circuit_transformer.optimize( | |
| aigs=[aig], | |
| num_mcts_steps=1 if mcts_playouts > 0 else 0, | |
| num_mcts_playouts_per_step=mcts_playouts | |
| ) | |
| optimized_aig = optimized_aigs[0] | |
| input_tt = ct.compute_input_tt(num_inputs) | |
| aig_tts = ct.compute_tts(aig, input_tt=input_tt) | |
| aig_tts_str = "\n".join([tt.to01() for tt in aig_tts]) | |
| optimized_aig_tts = ct.compute_tts(optimized_aig, input_tt=input_tt) | |
| optimized_aig_tts_str = "\n".join([tt.to01() for tt in optimized_aig_tts]) | |
| num_ands = pd.DataFrame(data={ | |
| "circuit": ["Original", "Optimized"], | |
| "size": [ct.count_num_ands(aig), ct.count_num_ands(optimized_aig)] | |
| }) | |
| ct.plot_network(aig, view=False, filename="aig.png") | |
| ct.plot_network(optimized_aig, view=False, filename="optimized_aig.png") | |
| return ct.write_aiger(optimized_aig), num_ands, "aig.png", "optimized_aig.png", aig_tts_str, optimized_aig_tts_str, str(ct.cec(aig, optimized_aig)) | |
| gradio_app = gr.Interface( | |
| predict, | |
| inputs=[gr.Code(value=aiger_str, label="Input logic circuit (in Aiger format, #(inputs) <= 8, #(outputs) <= 2)"), | |
| gr.Slider(minimum=0, maximum=8, value=0, step=1, label="Monte-Carlo tree search steps")], | |
| outputs=[gr.Code(label="Output logic circuit (size-optimized, in Aiger format)"), | |
| gr.BarPlot(label="Circuit size comparison", x="circuit", y="size", sort="-x"), | |
| gr.Image(label="Original logic circuit (And-Inverter Graph)"), | |
| gr.Image(label="Optimized logic circuit (And-Inverter Graph)"), | |
| gr.Text(label="Truth table of the original circuit"), | |
| gr.Text(label="Truth table of the optimized circuit"), | |
| gr.Text(label="Equivalence"),], | |
| title="Circuit Transformer for Size Minimization of Logic Circuits", | |
| description="""This is a demo to show how a [Circuit Transformer](https://openreview.net/forum?id=kpnW12Lm9p) | |
| minimize the size of a logic circuit by next token prediction, while strictly preserving logical equivalence. | |
| [[GitHub Repo]](https://github.com/snowkylin/circuit-transformer)""" | |
| ) | |
| if __name__ == "__main__": | |
| gradio_app.launch() |