Function MovingAvgQuantize

tensorflow/contrib/quantize/python/quant_ops.py:191–320

Adds a layer that collects quantization ranges as EMAs of input ranges. MovingAvgQuantize creates variables called 'min' and 'max', representing the interval used for quantization and clamping. Args: inputs: a tensor containing values to be quantized. per_channel: (default False) a b

(inputs,
                      per_channel=False,
                      init_min=-6.0,
                      init_max=6.0,
                      ema_decay=0.999,
                      vars_collection=ops.GraphKeys.MOVING_AVERAGE_VARIABLES,
                      name_prefix='MovingAvgQuantize',
                      reuse=None,
                      is_training=True,
                      num_bits=8,
                      narrow_range=False,
                      symmetric=False,
                      use_qdq=False)

def MovingAvgQuantize(inputs,
                      per_channel=False,
                      init_min=-6.0,
                      init_max=6.0,
                      ema_decay=0.999,
                      vars_collection=ops.GraphKeys.MOVING_AVERAGE_VARIABLES,
                      name_prefix='MovingAvgQuantize',
                      reuse=None,
                      is_training=True,
                      num_bits=8,
                      narrow_range=False,
                      symmetric=False,
                      use_qdq=False):
  """Adds a layer that collects quantization ranges as EMAs of input ranges.

  MovingAvgQuantize creates variables called 'min' and 'max', representing the
  interval used for quantization and clamping.

  Args:
    inputs: a tensor containing values to be quantized.
    per_channel: (default False) a boolean specifying whether to use different
      quantization ranges per output channel.
    init_min: a float scalar, the initial value for variable min.
    init_max: a float scalar, the initial value for variable max.
    ema_decay: EMA decay parameter.
    vars_collection: (Optional) collection where to store variables for
      quantization interval ends.
    name_prefix: name_prefix for created nodes.
    reuse: whether or not the layer and its variables should be reused. To be
      able to reuse the layer scope must be given.
    is_training: Whether the op is applied to a training or eval graph.
    num_bits: Number of bits to use for quantization, must be between 2 and 8.
    narrow_range: Whether to use the narrow quantization range
      [1; 2^num_bits - 1] or wide range [0; 2^num_bits - 1].
    symmetric: If true, use symmetric quantization limits instead of training
      the minimum and maximum of each quantization range separately.
    use_qdq: Use tf.quantize_and_dequantize_v3 op instead of fake_quant_with_min_max_vars
      for quantization. The qdq op is used for scaling with no zero point.
  Returns:
    a tensor containing quantized values.
  """
  with variable_scope.variable_scope(
      None, default_name=name_prefix, values=[inputs], reuse=reuse) as scope:
    scope.set_partitioner(None)
    input_shape = inputs.get_shape()
    if per_channel:
      input_dim = len(input_shape)
      # Only support quantizing 1-, 2- and 4-dimensional tensors.
      assert input_dim in [1, 2, 4], ('Expected 1D, 2D or 4D input, was: %s in '
                                      ' scope: %s' % (input_shape, name_prefix))
      min_max_shape = [input_shape[-1]]
    else:
      min_max_shape = []

    vars_collections = [vars_collection] if vars_collection else []
    min_var = _ModelVariable(
        'min',
        shape=min_max_shape,