def convert_box(converter, nnef_op, onnx_graph):
# type: (Converter, NNEFOperation, ONNXGraph)->None
input, output = converter.converted_tensors((nnef_op.input, nnef_op.output))
if nnef_op.attribs['size'] == [1] * input.rank:
onnx_op = ONNXOperation(graph=onnx_graph,
name='Pad',
inputs=input,
attribs=dict(mode=converter.onnx_pad_mode(nnef_op.attribs['border']),
pads=converter.onnx_pads(nnef_op.attribs['padding'])),
outputs=output)
if onnx_op.attribs['mode'] == 'constant':
onnx_op.attribs['value'] = 0.0
return
if nnef_op.attribs['normalize']:
partial_convert_pool(converter, nnef_op, onnx_graph,
target_name='AveragePool', input=input, outputs=(output,))
else:
temporary = ONNXTensor(graph=onnx_graph, shape=list(output.shape), dtype=output.dtype)
partial_convert_pool(converter, nnef_op, onnx_graph,
target_name='AveragePool', input=input, outputs=(temporary,), force_constant=True)
ONNXOperation(
graph=onnx_graph,
name='Mul',
inputs=(temporary,
converter.constant_0d_tensor(onnx_graph, float(np.product(nnef_op.attribs['size'])), 'FLOAT')),
outputs=output)
def convert_pad(converter, nnef_op, onnx_graph):
# type: (Converter, NNEFOperation, ONNXGraph)->None
input, output = converter.converted_tensors((nnef_op.input, nnef_op.output))
onnx_op = ONNXOperation(graph=onnx_graph,
name='Pad',
inputs=input,
attribs=dict(mode=converter.onnx_pad_mode(nnef_op.attribs['border']),
pads=converter.onnx_pads(nnef_op.attribs['padding'])),
outputs=output)
if onnx_op.attribs['mode'] == 'constant':
onnx_op.attribs['value'] = 0.0