flexflow/keras/layers/convolutional.py [25:171]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - class Conv2D(Layer): __slots__ = ['in_channels', 'out_channels', 'kernel_size', 'stride', \ 'padding', 'groups', 'activation', 'use_bias', 'kernel_initializer', \ 'bias_initializer'] def __init__(self, filters, input_shape=None, kernel_size=0, strides=(1, 1), padding="valid", data_format=None, dilation_rate=(1, 1), groups=1, activation=None, use_bias=True, kernel_initializer='glorot_uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs): if data_format == 'channels_last': assert 0, "data_format channels_last is not supported" if dilation_rate != (1,1): assert 0, "dilation_rate is not supported" if kernel_regularizer != None: assert 0, "kernel_regularizer is not supported" if bias_regularizer != None: assert 0, "bias_regularizer is not supported" if activity_regularizer != None: assert 0, "activity_regularizer is not supported" if kernel_constraint != None: assert 0, "kernel_constraint is not supported" if bias_constraint != None: assert 0, "bias_constraint is not supported" super(Conv2D, self).__init__("conv2d", "Conv2D", **kwargs) if kernel_initializer == "glorot_uniform": self.kernel_initializer = DefaultInitializer() elif isinstance(kernel_initializer, Initializer) == True: self.kernel_initializer = kernel_initializer else: assert 0, "[Dense]: unknown kernel_initializer" if bias_initializer == "zeros": self.bias_initializer = DefaultInitializer() elif isinstance(bias_initializer, Initializer) == True: self.bias_initializer = bias_initializer else: assert 0, "[Dense]: unknown bias_initializer" self.in_channels = 0 self.out_channels = filters assert len(kernel_size)==2, "wrong dim of kernel_size" self.kernel_size = kernel_size assert len(strides)==2, "wrong dim of stride" self.stride = strides if padding == "valid": self.padding = (0, 0) elif padding == "same": self.padding = "same" elif (isinstance(padding, list) or isinstance(padding, tuple)): assert len(padding)==2, "[Conv2D]: wrong dim of padding" self.padding = tuple(padding) else: assert 0, "[Conv2D]: check padding" self.groups = groups if (activation == None): self.activation = ff.ActiMode.AC_MODE_NONE elif(activation =="relu"): self.activation = ff.ActiMode.AC_MODE_RELU else: assert 0, "activation is not supported" if input_shape != None: if len(input_shape) == 4: self.in_channels = input_shape[1] self.input_shape = (input_shape[0], input_shape[1], input_shape[2], input_shape[3]) elif len(input_shape) == 3: self.in_channels = input_shape[0] self.input_shape = (0, input_shape[0], input_shape[1], input_shape[2]) self.use_bias = use_bias def verify_meta_data(self): assert self.input_shape != (0, 0, 0, 0), "[Conv2D]: input shape is wrong" assert self.output_shape != (0, 0, 0, 0), "[Conv2D]: output shape is wrong" assert self.in_channels != 0, "[Conv2D]: in channels is wrong" assert self.out_channels != 0, "[Conv2D]: out channels is wrong" def get_summary(self): summary = "%s%s\t\t%s\t%s\n"%(self._get_summary_name(), self.output_shape, self.input_shape, self._get_summary_connected_to()) return summary def get_weights(self, ffmodel): return self._get_weights(ffmodel) def set_weights(self, ffmodel, kernel, bias): self._set_weights(ffmodel, kernel, bias) def __call__(self, input_tensor): return self._connect_layer_1_input_1_output(input_tensor) def _calculate_inout_shape(self, input_tensor): assert input_tensor.num_dims == 4, "[Conv2D]: shape of input tensor is wrong" input_b = input_tensor.batch_shape[0] input_d = input_tensor.batch_shape[1] input_h = input_tensor.batch_shape[2] input_w = input_tensor.batch_shape[3] assert input_h != 0, "wrong input_h" assert input_w != 0, "wrong input_w" assert input_d != 0, "wrong input_d" #calculate padding for same if (self.padding == 'same'): if (input_h % self.stride[0] == 0): padding_h = max(self.kernel_size[0] - self.stride[0], 0) else: padding_h = max(self.kernel_size[0] - (input_h % self.stride[0]), 0) if (input_w % self.stride[1] == 0): padding_w = max(self.kernel_size[1] - self.stride[1], 0) else: padding_w = max(self.kernel_size[1] - (input_w % self.stride[1]), 0) self.padding = (padding_h//2, padding_w//2) fflogger.debug("conv2d same padding %s" %(str(self.padding))) self.input_shape = (input_b, input_d, input_w, input_h) self.in_channels = input_d output_h = 1 + math.floor((input_h + 2 * self.padding[0] - self.kernel_size[0]) / self.stride[0]) output_w = 1 + math.floor((input_w + 2 * self.padding[1] - self.kernel_size[1]) / self.stride[1]) output_d = self.out_channels self.output_shape = (input_b, output_d, output_h, output_w) fflogger.debug("conv2d input %s, output %s" %( str(self.input_shape), str(self.output_shape))) def _verify_inout_tensor_shape(self, input_tensor, output_tensor): assert input_tensor.num_dims == 4, "[Conv2D]: check input tensor dims" assert input_tensor.batch_shape[1] == self.input_shape[1] assert input_tensor.batch_shape[2] == self.input_shape[2] assert input_tensor.batch_shape[3] == self.input_shape[3] assert output_tensor.num_dims == 4, "[Conv2D]: check output tensor dims" assert output_tensor.batch_shape[1] == self.output_shape[1] assert output_tensor.batch_shape[2] == self.output_shape[2] assert output_tensor.batch_shape[3] == self.output_shape[3] def _reset_layer(self): self.in_channels = 0 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - python/flexflow/keras/layers/convolutional.py [25:171]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - class Conv2D(Layer): __slots__ = ['in_channels', 'out_channels', 'kernel_size', 'stride', \ 'padding', 'groups', 'activation', 'use_bias', 'kernel_initializer', \ 'bias_initializer'] def __init__(self, filters, input_shape=None, kernel_size=0, strides=(1, 1), padding="valid", data_format=None, dilation_rate=(1, 1), groups=1, activation=None, use_bias=True, kernel_initializer='glorot_uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs): if data_format == 'channels_last': assert 0, "data_format channels_last is not supported" if dilation_rate != (1,1): assert 0, "dilation_rate is not supported" if kernel_regularizer != None: assert 0, "kernel_regularizer is not supported" if bias_regularizer != None: assert 0, "bias_regularizer is not supported" if activity_regularizer != None: assert 0, "activity_regularizer is not supported" if kernel_constraint != None: assert 0, "kernel_constraint is not supported" if bias_constraint != None: assert 0, "bias_constraint is not supported" super(Conv2D, self).__init__("conv2d", "Conv2D", **kwargs) if kernel_initializer == "glorot_uniform": self.kernel_initializer = DefaultInitializer() elif isinstance(kernel_initializer, Initializer) == True: self.kernel_initializer = kernel_initializer else: assert 0, "[Dense]: unknown kernel_initializer" if bias_initializer == "zeros": self.bias_initializer = DefaultInitializer() elif isinstance(bias_initializer, Initializer) == True: self.bias_initializer = bias_initializer else: assert 0, "[Dense]: unknown bias_initializer" self.in_channels = 0 self.out_channels = filters assert len(kernel_size)==2, "wrong dim of kernel_size" self.kernel_size = kernel_size assert len(strides)==2, "wrong dim of stride" self.stride = strides if padding == "valid": self.padding = (0, 0) elif padding == "same": self.padding = "same" elif (isinstance(padding, list) or isinstance(padding, tuple)): assert len(padding)==2, "[Conv2D]: wrong dim of padding" self.padding = tuple(padding) else: assert 0, "[Conv2D]: check padding" self.groups = groups if (activation == None): self.activation = ff.ActiMode.AC_MODE_NONE elif(activation =="relu"): self.activation = ff.ActiMode.AC_MODE_RELU else: assert 0, "activation is not supported" if input_shape != None: if len(input_shape) == 4: self.in_channels = input_shape[1] self.input_shape = (input_shape[0], input_shape[1], input_shape[2], input_shape[3]) elif len(input_shape) == 3: self.in_channels = input_shape[0] self.input_shape = (0, input_shape[0], input_shape[1], input_shape[2]) self.use_bias = use_bias def verify_meta_data(self): assert self.input_shape != (0, 0, 0, 0), "[Conv2D]: input shape is wrong" assert self.output_shape != (0, 0, 0, 0), "[Conv2D]: output shape is wrong" assert self.in_channels != 0, "[Conv2D]: in channels is wrong" assert self.out_channels != 0, "[Conv2D]: out channels is wrong" def get_summary(self): summary = "%s%s\t\t%s\t%s\n"%(self._get_summary_name(), self.output_shape, self.input_shape, self._get_summary_connected_to()) return summary def get_weights(self, ffmodel): return self._get_weights(ffmodel) def set_weights(self, ffmodel, kernel, bias): self._set_weights(ffmodel, kernel, bias) def __call__(self, input_tensor): return self._connect_layer_1_input_1_output(input_tensor) def _calculate_inout_shape(self, input_tensor): assert input_tensor.num_dims == 4, "[Conv2D]: shape of input tensor is wrong" input_b = input_tensor.batch_shape[0] input_d = input_tensor.batch_shape[1] input_h = input_tensor.batch_shape[2] input_w = input_tensor.batch_shape[3] assert input_h != 0, "wrong input_h" assert input_w != 0, "wrong input_w" assert input_d != 0, "wrong input_d" #calculate padding for same if (self.padding == 'same'): if (input_h % self.stride[0] == 0): padding_h = max(self.kernel_size[0] - self.stride[0], 0) else: padding_h = max(self.kernel_size[0] - (input_h % self.stride[0]), 0) if (input_w % self.stride[1] == 0): padding_w = max(self.kernel_size[1] - self.stride[1], 0) else: padding_w = max(self.kernel_size[1] - (input_w % self.stride[1]), 0) self.padding = (padding_h//2, padding_w//2) fflogger.debug("conv2d same padding %s" %(str(self.padding))) self.input_shape = (input_b, input_d, input_w, input_h) self.in_channels = input_d output_h = 1 + math.floor((input_h + 2 * self.padding[0] - self.kernel_size[0]) / self.stride[0]) output_w = 1 + math.floor((input_w + 2 * self.padding[1] - self.kernel_size[1]) / self.stride[1]) output_d = self.out_channels self.output_shape = (input_b, output_d, output_h, output_w) fflogger.debug("conv2d input %s, output %s" %( str(self.input_shape), str(self.output_shape))) def _verify_inout_tensor_shape(self, input_tensor, output_tensor): assert input_tensor.num_dims == 4, "[Conv2D]: check input tensor dims" assert input_tensor.batch_shape[1] == self.input_shape[1] assert input_tensor.batch_shape[2] == self.input_shape[2] assert input_tensor.batch_shape[3] == self.input_shape[3] assert output_tensor.num_dims == 4, "[Conv2D]: check output tensor dims" assert output_tensor.batch_shape[1] == self.output_shape[1] assert output_tensor.batch_shape[2] == self.output_shape[2] assert output_tensor.batch_shape[3] == self.output_shape[3] def _reset_layer(self): self.in_channels = 0 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -