aimet_onnx.apply_seq_mse¶

Top level APIs

aimet_onnx.apply_seq_mse(sim, inputs, num_candidates=20)[source]¶

Sequentially optimizes the QuantizationSimModel’s weight encodings to reduce MSE loss at layer outputs.

Parameters:

sim (QuantizationSimModel) – Calibrated QuantizationSimModel instance to optimize
inputs (Collection[Dict[str, np.ndarray]]) – The set of input samples to use during optimization
num_candidates (int) – Number of encoding candidates to sweep for each weight. Decreasing this can reduce runtime but may lead to lower accuracy.