tensorlayer3/tensorlayer/layers/convolution/quan_conv.py

#! /usr/bin/python
# -*- coding: utf-8 -*-

import tensorlayer as tl
from tensorlayer import logging
from tensorlayer.layers.core import Module
from tensorlayer.layers.utils import (quantize_active_overflow, quantize_weight_overflow)

__all__ = ['QuanConv2d']


class QuanConv2d(Module):
    """The :class:`QuanConv2d` class is a quantized convolutional layer without BN, which weights are 'bitW' bits and the output of the previous layer
    are 'bitA' bits while inferencing.
    Note that, the bias vector would not be binarized.

    Parameters
    ----------
    With TensorLayer

    n_filter : int
        The number of filters.
    filter_size : tuple of int
        The filter size (height, width).
    strides : tuple of int
        The sliding window strides of corresponding input dimensions.
        It must be in the same order as the ``shape`` parameter.
    act : activation function
        The activation function of this layer.
    padding : str
        The padding algorithm type: "SAME" or "VALID".
    bitW : int
        The bits of this layer's parameter
    bitA : int
        The bits of the output of previous layer
    use_gemm : boolean
        If True, use gemm instead of ``tf.matmul`` for inference.
        TODO: support gemm
    data_format : str
        "channels_last" (NHWC, default) or "channels_first" (NCHW).
    dilation_rate : tuple of int
        Specifying the dilation rate to use for dilated convolution.
    W_init : initializer
        The initializer for the the weight matrix.
    b_init : initializer or None
        The initializer for the the bias vector. If None, skip biases.
    in_channels : int
        The number of in channels.
    name : None or str
        A unique layer name.

    Examples
    ---------
    With TensorLayer

    >>> net = tl.layers.Input([8, 12, 12, 64], name='input')
    >>> quanconv2d = tl.layers.QuanConv2d(
    ...     n_filter=32, filter_size=(5, 5), strides=(1, 1), act=tl.ReLU, padding='SAME', name='quancnn2d'
    ... )(net)
    >>> print(quanconv2d)
    >>> output shape : (8, 12, 12, 32)

    """

    def __init__(
        self,
        bitW=8,
        bitA=8,
        n_filter=32,
        filter_size=(3, 3),
        strides=(1, 1),
        act=None,
        padding='SAME',
        use_gemm=False,
        data_format="channels_last",
        dilation_rate=(1, 1),
        W_init=tl.initializers.truncated_normal(stddev=0.02),
        b_init=tl.initializers.constant(value=0.0),
        in_channels=None,
        name=None  # 'quan_cnn2d',
    ):
        super().__init__(name, act=act)
        self.bitW = bitW
        self.bitA = bitA
        self.n_filter = n_filter
        self.filter_size = filter_size
        self.strides = self._strides = strides
        self.padding = padding
        self.use_gemm = use_gemm
        self.data_format = data_format
        self.dilation_rate = self._dilation_rate = dilation_rate
        self.W_init = W_init
        self.b_init = b_init
        self.in_channels = in_channels

        if self.in_channels:
            self.build(None)
            self._built = True

        logging.info(
            "QuanConv2d %s: n_filter: %d filter_size: %s strides: %s pad: %s act: %s" % (
                self.name, n_filter, str(filter_size), str(strides), padding,
                self.act.__class__.__name__ if self.act is not None else 'No Activation'
            )
        )

        if self.use_gemm:
            raise Exception("TODO. The current version use tf.matmul for inferencing.")

        if len(self.strides) != 2:
            raise ValueError("len(strides) should be 2.")

    def __repr__(self):
        actstr = self.act.__name__ if self.act is not None else 'No Activation'
        s = (
            '{classname}(in_channels={in_channels}, out_channels={n_filter}, kernel_size={filter_size}'
            ', strides={strides}, padding={padding}'
        )
        if self.dilation_rate != (1, ) * len(self.dilation_rate):
            s += ', dilation={dilation_rate}'
        if self.b_init is None:
            s += ', bias=False'
        s += (', ' + actstr)
        if self.name is not None:
            s += ', name=\'{name}\''
        s += ')'
        return s.format(classname=self.__class__.__name__, **self.__dict__)

    def build(self, inputs_shape):
        if self.data_format == 'channels_last':
            self.data_format = 'NHWC'
            if self.in_channels is None:
                self.in_channels = inputs_shape[-1]
            self._strides = [1, self._strides[0], self._strides[1], 1]
            self._dilation_rate = [1, self._dilation_rate[0], self._dilation_rate[1], 1]
        elif self.data_format == 'channels_first':
            self.data_format = 'NCHW'
            if self.in_channels is None:
                self.in_channels = inputs_shape[1]
            self._strides = [1, 1, self._strides[0], self._strides[1]]
            self._dilation_rate = [1, 1, self._dilation_rate[0], self._dilation_rate[1]]
        else:
            raise Exception("data_format should be either channels_last or channels_first")

        self.filter_shape = (self.filter_size[0], self.filter_size[1], self.in_channels, self.n_filter)

        self.W = self._get_weights("filters", shape=self.filter_shape, init=self.W_init)
        if self.b_init:
            self.b = self._get_weights("biases", shape=(self.n_filter, ), init=self.b_init)
            self.bias_add = tl.ops.BiasAdd(data_format=self.data_format)

        self.conv2d = tl.ops.Conv2D(
            strides=self.strides, padding=self.padding, data_format=self.data_format, dilations=self._dilation_rate
        )

    def forward(self, inputs):
        if self._forward_state == False:
            if self._built == False:
                self.build(tl.get_tensor_shape(inputs))
                self._built = True
            self._forward_state = True

        inputs = quantize_active_overflow(inputs, self.bitA)

        W_ = quantize_weight_overflow(self.W, self.bitW)

        outputs = self.conv2d(inputs, W_)

        if self.b_init:
            outputs = self.bias_add(outputs, self.b)
        if self.act:
            outputs = self.act(outputs)

        return outputs