def _transform_extract_bias_add(g):
# type:(NNEFGraph)->None
supported_ops = {"conv", "deconv"}
for nnefop in list(g.operations):
if nnefop.name in supported_ops and len(nnefop.inputs) >= 3:
bias = nnefop.inputs[2]
nnefop.inputs = tuple(nnefop.inputs)[:2]
if not (bias.is_constant and bias.data == [0]):
output_with_bias = nnefop.output
output_without_bias = NNEFTensor(graph=g,
name=None,
dtype=output_with_bias.dtype,
shape=output_with_bias.shape)
nnefop.outputs = output_without_bias
NNEFOperation(graph=g,
name="_bias_add",
inputs=(output_without_bias, bias),
outputs=output_with_bias)
def transform_extract_padding(g):
# type:(NNEFGraph)->None
forward_ops = {
"conv", "argmax_pool", "max_pool_with_index", "max_pool", "avg_pool"
}
backward_ops = {
"deconv", "conv_grad_filter", "max_pool_grad_with_index",
"avg_pool_grad", "max_pool_grad"
}
supported_ops = forward_ops.union(backward_ops)
for nnefop in list(g.operations):
if nnefop.name in supported_ops:
in_op_padding, separate_padding = get_paddings(nnefop)
nnefop.attribs["padding"] = in_op_padding
assert separate_padding is None if nnefop.name in backward_ops else True
if separate_padding is not None:
input = nnefop.inputs[0]
pad_op = NNEFOperation(
graph=g,
name="box",
inputs=input,
attribs=dict(size=[1] * input.rank,
border=nnefop.attribs["border"],
padding=separate_padding),
outputs=NNEFTensor(
graph=g,
name=None,
shape=[
s + p + q
for s, (p, q) in zip(input.shape, separate_padding)
],
dtype=input.dtype))
nnefop.inputs = (pad_op.output, ) + tuple(nnefop.inputs)[1:]
def create_copy_op(self, graph, input, output):
return NNEFOperation(graph=graph,
name="copy",
inputs=input,
outputs=output)
def create_transpose_op(self, graph, input, output, axes):
return NNEFOperation(graph=graph,
name="transpose",
inputs=input,
attribs=dict(axes=axes),
outputs=output)