def tf_create_attrs(node, input_feature_channel, output_feature_channel):
data_format = node.pb.attr["data_format"]
dilations = tf_int_list(node.pb.attr["dilations"].list)
if len(dilations) == 0:
dilations = None
attrs = {
'type': 'Convolution',
'auto_pad':
convert_tf_padding_to_str(node.pb.attr['padding'].s.decode()),
'bias_addable': True,
'bias_term': False,
'dilation': dilations,
'stride': tf_int_list(node.pb.attr["strides"].list),
'channel_dims': tf_data_format_channel(data_format),
'batch_dims': tf_data_format_batch(data_format),
'input_feature_channel': input_feature_channel,
'output_feature_channel': output_feature_channel,
'layout': data_format.s.decode(),
# get_group and get_output_feature_dim are special attrs that stores lambdas ( lambda node, kernel_shape:...)
# this attrs calls in infer function to calculate output feature dimension and group attr
'get_group': None, # lambda should return group attr for given node
'get_output_feature_dim':
None, # lamda should return output feature dimension
}
return attrs
def tf_squeeze_ext(pb):
return {
# TODO handle case when squeeze_dims are not in pb
'type': 'Reshape',
'squeeze_dims': tf_int_list(pb.attr['squeeze_dims'].list),
'infer': tf_squeeze_infer
}
def extract(cls, node):
attrs = {
'spatial_dims':
int64_array([1, 2]),
'sizes':
tf_int_list(node.pb.attr['ksizes'].list),
'strides':
tf_int_list(node.pb.attr['strides'].list),
'rates':
tf_int_list(node.pb.attr['rates'].list),
'auto_pad':
convert_tf_padding_to_str(node.pb.attr['padding'].s.decode()),
}
ExtractImagePatches.update_node_stat(node, attrs)
return cls.enabled
def create_pooling_attrs(node, pool_method):
data_format = node.pb.attr["data_format"]
attrs = {
'auto_pad': convert_tf_padding_to_str(node.pb.attr['padding'].s.decode()),
'window': tf_int_list(node.pb.attr['ksize'].list),
'spatial_dims': tf_data_format_spatial(data_format),
'pad': None, # will be inferred when input shape is known
'stride': tf_int_list(node.pb.attr['strides'].list),
'pad_spatial_shape': None,
'output_spatial_shape': None,
'pool_method': pool_method,
'layout': data_format.s.decode(),
'exclude_pad': True,
}
return attrs
def tf_create_attrs(node, input_feature_channel, output_feature_channel):
data_format = node.pb.attr["data_format"]
return {
'auto_pad': convert_tf_padding_to_str(node.pb.attr['padding']),
'bias_addable': True,
'bias_term': False,
'spatial_dims': tf_data_format_spatial(data_format),
'channel_dims': tf_data_format_channel(data_format),
'batch_dims': tf_data_format_batch(data_format),
'pad': None, # will be inferred when input shape is known
'pad_spatial_shape': None,
'output_spatial_shape': None,
'output_shape': None,
'output': None,
'stride': tf_int_list(node.pb.attr["strides"].list),
'type': None, # don't set type until we are sure it is really translated to correct IR; see infer function
'group': None,
'layout': data_format.s.decode(),
'input_feature_channel': input_feature_channel,
'output_feature_channel': output_feature_channel,
}
def extract(cls, node: Node):
Squeeze.update_node_stat(
node,
{'squeeze_dims': tf_int_list(node.pb.attr['squeeze_dims'].list)})
return cls.enabled
