def convert_multilinear_upsample(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['method'] != 'symmetric':
if converter.enable_imprecise_image_resize:
print("Warning: method={} is unsupported in multilinear_upsample, "
"using symmetric, because enable_imprecise_image_resize was True"
.format(nnef_op.attribs["method"]))
else:
assert False, "Error: method={} is unsupported in multilinear_upsample. " \
"Use enable_imprecise_image_resize=True, to suppress this error." \
.format(nnef_op.attribs["method"])
if nnef_op.attribs['border'] != 'replicate':
if converter.enable_imprecise_image_resize:
print("Warning: border={} is unsupported in multilinear_upsample, "
"using replicate, because enable_imprecise_image_resize was True"
.format(nnef_op.attribs["border"]))
else:
assert False, "Error: border={} is unsupported in multilinear_upsample. " \
"Use enable_imprecise_image_resize=True, to suppress this error." \
.format(nnef_op.attribs["border"])
scales = [float(f) for f in [1, 1] + nnef_op.attribs['factor']]
ONNXOperation(graph=onnx_graph,
name='Upsample',
inputs=(input, converter.constant_1d_tensor(graph=onnx_graph, list_=scales, dtype='FLOAT')),
attribs=dict(mode='linear'),
outputs=output)
def convert_tile(converter, nnef_op, onnx_graph):
# type: (Converter, NNEFOperation, ONNXGraph)->None
input, output = converter.converted_tensors((nnef_op.input, nnef_op.output))
ONNXOperation(graph=onnx_graph,
name='Tile',
inputs=(input,
converter.constant_1d_tensor(graph=onnx_graph,
list_=nnef_op.attribs['repeats'],
dtype='INT64')),
outputs=output)
def convert_nearest_upsample(converter, nnef_op, onnx_graph):
# type: (Converter, NNEFOperation, ONNXGraph)->None
input, output = converter.converted_tensors((nnef_op.input, nnef_op.output))
scales = [float(f) for f in [1, 1] + nnef_op.attribs['factor']]
ONNXOperation(graph=onnx_graph,
name='Upsample',
inputs=(input, converter.constant_1d_tensor(onnx_graph, scales, 'FLOAT')),
attribs=dict(mode='nearest'),
outputs=output)
def convert_reshape(converter, nnef_op, onnx_graph):
# type: (Converter, NNEFOperation, ONNXGraph)->None
input, output = converter.converted_tensors((nnef_op.input, nnef_op.output))
shape = nnef_op.attribs['shape']
axis_start = nnef_op.attribs['axis_start']
axis_count = nnef_op.attribs['axis_count']
if axis_count == -1:
axis_count = input.rank - axis_start
onnx_shape = input.shape[:axis_start] + shape + input.shape[axis_start + axis_count:]
ONNXOperation(graph=onnx_graph,
name='Reshape',
inputs=(input, converter.constant_1d_tensor(graph=onnx_graph,
list_=onnx_shape,
dtype='INT64')),
outputs=output)