neural_compressor.quantization

Neural Compressor Quantization API.

Functions

fit(model, conf[, calib_dataloader, calib_func, ...])

Quantize the model with a given configure.

Module Contents

neural_compressor.quantization.fit(model, conf, calib_dataloader=None, calib_func=None, eval_func=None, eval_dataloader=None, eval_metric=None, **kwargs)

Quantize the model with a given configure.

Parameters:

• model (torch.nn.Module) – For Tensorflow model, it could be a path to frozen pb,loaded graph_def object or a path to ckpt/savedmodel folder. For PyTorch model, it’s torch.nn.model instance. For MXNet model, it’s mxnet.symbol.Symbol or gluon.HybirdBlock instance.

• conf (PostTrainingQuantConfig) – The class of PostTrainingQuantConfig containing accuracy goal, tuning objective and preferred calibration & quantization tuning space etc.

• calib_dataloader (generator) – Data loader for calibration, mandatory for post-training quantization. It is iterable and should yield a tuple (input, label) for calibration dataset containing label, or yield (input, _) for label-free calibration dataset. The input could be a object, list, tuple or dict, depending on user implementation, as well as it can be taken as model input.

• calib_func (function, optional) – Calibration function for post-training static quantization. It is optional. This function takes “model” as input parameter and executes entire inference process.

• eval_func (function, optional) –

  The evaluation function provided by user. This function takes model as parameter, and evaluation dataset and metrics should be encapsulated in this function implementation and outputs a higher-is-better accuracy scalar value. The pseudo code should be something like: def eval_func(model):

  input, label = dataloader() output = model(input) accuracy = metric(output, label) return accuracy.

  The user only needs to set eval_func or eval_dataloader and eval_metric which is an alternative option to tune the model accuracy.

• eval_dataloader (generator, optional) – Data loader for evaluation. It is iterable and should yield a tuple of (input, label). The input could be a object, list, tuple or dict, depending on user implementation, as well as it can be taken as model input. The label should be able to take as input of supported metrics. If this parameter is not None, user needs to specify pre-defined evaluation metrics through configuration file and should set “eval_func” parameter as None. Tuner will combine model, eval_dataloader and pre-defined metrics to run evaluation process.

• eval_metric (dict or obj) –

  Set metric class or a dict of built-in metric configures,

  and neural_compressor will initialize this class when evaluation.

  1. neural_compressor have many built-in metrics, user can pass a metric configure dict to tell neural compressor what metric will be use. You also can set multi-metrics to evaluate the performance of a specific model.

     Single metric:

     {topk: 1}

     Multi-metrics:

     {topk: 1,

     MSE: {compare_label: False}, weight: [0.5, 0.5], higher_is_better: [True, False]

     }

• metrics (For the built-in)

• link (please refer to below)

• https –

  //github.com/intel/neural-compressor/blob/master/docs/source/metric.md#supported-built-in-metric-matrix.

  2. User also can get the built-in metrics by neural_compressor.Metric:

     Metric(name=”topk”, k=1)

  3. User also can set specific metric through this api. The metric class should take the outputs of the model or postprocess(if have) as inputs, neural_compressor built-in metric always take (predictions, labels) as inputs for update, and user_metric.metric_cls should be sub_class of neural_compressor.metric.BaseMetric.

Example:

# Quantization code for PTQ
from neural_compressor import PostTrainingQuantConfig
from neural_compressor import quantization
def eval_func(model):
    for input, label in dataloader:
        output = model(input)
        metric.update(output, label)
    accuracy = metric.result()
    return accuracy

conf = PostTrainingQuantConfig()
q_model = quantization.fit(model_origin,
                           conf,
                           calib_dataloader=dataloader,
                           eval_func=eval_func)

# Saved quantized model in ./saved folder
q_model.save("./saved")