theano-xnor-net代码注释7 inf_layers.py

import theanoimport theano.tensor as Timport numpy as npimport lasagnefrom fxp_helper import to_fixed_point_theano# Classic batch normalization layer with fixed point simulation feature for the output dataclass BatchNormLayer(lasagne.layers.BatchNormLayer):    """Class to override the lasagne batch norm layer. This is only implemetend for inference.    This basically clubs the normalization parameters and pre-multiplies them.    """    def __init__(self, incoming, format='float', data_bits=15, int_bits=0, **kwargs):        super(BatchNormLayer, self).__init__(incoming, **kwargs)        # club mean, gamma, inv_std        sub = -self.mean.get_value() * self.inv_std.get_value() * self.gamma.get_value() + self.beta.get_value()        scale = self.gamma.get_value() * self.inv_std.get_value()        # The below 2 params are just combination of layer params used during training.        # This is just to precompute the data independent products and keep it aside. Also, the range of        # clubbed params is less compared to isolated params. This helps to reduce the # of integer bits used to represent        # them in the fixed point        #self.add_param()就是将类属性放置在self.param()类属性中,在之后可以通过.get_all_params()来调出该曾的所有存放在self.param()中的类属性        self.sub = self.add_param(sub, sub.shape, name='sub')        self.scale = self.add_param(scale, scale.shape, name='scale')        self.format = format        # FIXME: The below variables need to be theano shared variables?        self.data_bits = theano.shared(data_bits)        self.int_bits = theano.shared(int_bits)    def get_output_for(self, input, deterministic=False, **kwargs):        """Override the lasagne implementation only during the inference time.        """        #这个函数只存在于测试阶段，训练阶段还是用lasagne        assert(deterministic), 'This layer is only implemented for inference. Use direct Lasagne implementation during training'        # prepare dimshuffle pattern inserting broadcastable axes as needed        param_axes = iter(range(input.ndim - len(self.axes)))        pattern = ['x' if input_axis in self.axes                   else next(param_axes)                   for input_axis in range(input.ndim)]        scale = self.scale.dimshuffle(pattern)        sub = self.sub.dimshuffle(pattern)        if(self.format == 'fixed'):            # assuming the parameters are already simulated for respective fixed point formats.            full_precision_out = input * scale + sub            return to_fixed_point_theano(full_precision_out, self.data_bits, self.int_bits)        else:            return input * scale + sub# Classic Lasagne dense layer with fixed point simulation feature for the output dataclass DenseLayer(lasagne.layers.DenseLayer):    def __init__(self, incoming, num_units, format='float', data_bits=15, int_bits=0, **kwargs):        num_inputs = int(np.prod(incoming.output_shape[1:]))        super(DenseLayer, self).__init__(incoming, num_units,  **kwargs)        # params for fixed point simulation        self.format = format        self.data_bits = theano.shared(data_bits)        self.int_bits = theano.shared(int_bits)    def get_output_for(self, input, deterministic=True, **kwargs):        """ Dense layer with fixed point simulation option        """        if(self.format == 'fixed'):            # dot-product at full precision            fc_out = super(DenseLayer, self).get_output_for(input, **kwargs)            # reduce the precision of the output based on the data bit widths specified            fc_out = to_fixed_point_theano(fc_out, self.data_bits, self.int_bits)        else:            fc_out = super(DenseLayer, self).get_output_for(input, **kwargs)        return fc_out# Classic Lasagne Conv layer with fixed point simulation feature for the output data#该Conv2DLayer类继承lasagne模块下的Conv2DLayerclass Conv2DLayer(lasagne.layers.Conv2DLayer):    #构造函数，传入参数，如果传入的参数该参数列表没有出现，如pad与nonlinearity，该标签参数以字典的形式存入**kwargs    def __init__(self, incoming, num_filters, filter_size, format='float', data_bits=15, int_bits=0, **kwargs):        #调用父类构造函数，super函数：可以调用父类中定义的非私有函数，且防止多继承情况下的父类构造函数的重复调用        super(Conv2DLayer, self).__init__(incoming, num_filters, filter_size, **kwargs)        self.format = format        self.data_bits = data_bits        self.int_bits = int_bits    def convolve(self, input, deterministic=False, **kwargs):        feat_maps = super(Conv2DLayer, self).convolve(input, **kwargs)        if(self.format == 'fixed'):            feat_maps = to_fixed_point_theano(feat_maps, self.data_bits, self.int_bits)        else:            pass        return feat_maps