def extract_convolution(self,
input_name,
output_name,
scope_id,
num_output,
kernel_size,
stride,
padding,
data_format="NCHW",
weight_format="NHWC",
axis=1,
dilation=1,
groups=1,
layer_names=["convolution", "kernel", "bias"]):
kernel, bias = self.get_weights(scope_id, layer_names)
if (weight_format == "NHWC"):
if (len(kernel.shape) == 3):
kernel = kernel.transpose([2, 1, 0])
kernel = np.expand_dims(kernel, -1)
elif (len(kernel.shape) == 4):
kernel = kernel.transpose([3, 2, 0, 1])
else:
print("[ERROR] unsupported convolution kernel size")
exit(1)
layer = caffe_net.LayerParameter(name=output_name,
type='Convolution',
bottom=[input_name],
top=[output_name])
if (bias is None):
layer.add_data(kernel)
bias_term = False
else:
layer.add_data(kernel, bias)
bias_term = True
layer.convolution_param(num_output, kernel_size, stride, padding,
bias_term, dilation, groups)
self.caffe_model.add_layer(layer)
assert (data_format == "NCHW")
if (self.data_dict[input_name] is not None):
input_data, input_shape, inv_transpose_dims = self.preprocess_nchwc8_nchw_input(
input_name, axis)
output_data = Operators.convolution(input_data, kernel, bias,
num_output, kernel_size,
stride, padding, dilation,
groups, output_name)
self.data_dict[
output_name] = output_data #self.postprocess_nchwc8_nchw_output(output_data, input_shape, inv_transpose_dims)
else:
self.data_dict[output_name] = None
return output_name
