GAMMA: Automating the HW Mapping of DNN Models on Accelerators via Genetic Algorithm
We support naive multi-objective optimization, where the user can specify up to three different objectives. If the user want single-objective optimization, simply don't specify fitness2 and fitness3.
- fitness1: The fitness objective
- fitness2: (Optional) The second objective
- fitness3: (Optional) The third objective
- config_path: Configuration path, should include arch.yaml, problem.yaml, (and sparse.yaml if sparsity is considered)
- use_sparse: Enable it to explore sparse accelerator space, otherwise explore dense accelerator space
- explore_bypass: Enable it to explore bypass buffer option
- epochs: Number of generations
- num_pops: Number of populations
- save_chkpt: To save the trace of improvement over epoch or not. Specify if the user want to save the trace.
- report_dir: The report directory for the generated map.yaml and the trace-file