aimet_onnx.seq_mse

Top level APIs

aimet_onnx.sequential_mse.seq_mse.SequentialMse.apply_seq_mse(model, sim, params, data_loader)

It performs following steps: 1) creates seq_mse object 2) call apply_seq_algo() member function

Parameters:

  • model (ModelProto) – float model

  • sim (QuantizationSimModel) – QuantizationSimModel object

  • params (SeqMseParams) – Sequential MSE parameters

  • data_loader (Iterable) – Data loader

Sequential MSE parameters

class aimet_onnx.sequential_mse.seq_mse.SeqMseParams(num_batches, num_candidates=20, inp_symmetry='symqt', loss_fn='mse')[source]

Sequential MSE parameters

Parameters:

  • num_batches (int) – Number of batches.

  • num_candidates (int) – Number of candidates to perform grid search. Default 20.

  • inp_symmetry (str) – Input symmetry. Available options are ‘asym’, ‘symfp’ and ‘symqt’. Default ‘symqt’.

  • loss_fn (str) – Loss function. Available options are ‘mse’, ‘l1’ and ‘sqnr’. Default ‘mse’.