def _pad():
"""converting attributes for Pad operator"""
return AttrCvt(
op_name='pad',
transforms={
'pads': ('pad_width', (0, 0, 0, 0, 0, 0, 0, 0), _pad_sequence_fix),
'value':'constant_value'})
def _batch_norm():
"""converting attributes for BatchNorm operator"""
return AttrCvt(
op_name='BatchNorm',
transforms={'epsilon': 'eps'},
extras={'cudnn_off': 1},
ignores=['spatial', 'is_test', 'consumed_inputs'])
def _activation(name):
"""converting attributes for LeakyRelu operator"""
return AttrCvt(
op_name='LeakyReLU',
transforms={
'alpha':'slope'},
extras={'act_type': name})
def _global_pooling(name):
"""Requires kernel attribute which is not present in onnx currently.
So for now giving default kernel."""
return AttrCvt(
op_name='Pooling',
extras={'global_pool': True,
'kernel': (1, 1),
'pool_type': name})
def _upsample(name):
"""converting attributes for UpSampling operator"""
return AttrCvt(op_name=name,
transforms={
'height_scale': ('scale', 1, _upsample_scale_fix),
'mode':
('sample_type', 'nearest', _upsample_restrict_mode),
'width_scale': ('scale', 1, _upsample_scale_fix)
})
def _conv():
"""converting attributes for convolution operator"""
return AttrCvt(
op_name='Convolution',
transforms={
'kernel_shape': 'kernel',
'strides': 'stride',
'dilations': ('dilate', (0, 0)),
'pads': ('pad', (0, 0), _revert_caffe2_pad),
'group': ('num_group', 1)},
custom_check=_dimension_constraint())
def _pooling(name):
"""converting attributes for pooling operator"""
return AttrCvt(
op_name='Pooling',
transforms={
'kernel_shape': 'kernel',
'strides': 'stride',
'pads': 'pad'},
# pooling convention full to match caffe2
extras={'pool_type': name, 'pooling_convention':'valid'},
custom_check=_dimension_constraint())
def _conv_transpose():
"""converting attributes for deconvolution operator"""
return AttrCvt(op_name='Deconvolution',
transforms={
'kernel_shape': 'kernel',
'strides': 'stride',
'dilations': ('dilate', (0, 0)),
'pads': ('pad', (0, 0), _revert_caffe2_pad)
},
disables=['output_shape'],
custom_check=_dimension_constraint())
def _pooling(name):
"""converting attributes for pooling operator"""
return AttrCvt(
op_name='Pooling',
transforms={
'kernel_shape': 'kernel',
'strides': 'stride',
'pads': ('pad', (0, 0), _revert_caffe2_pad)
},
# pooling convention full to match caffe2
extras={
'pool_type': name,
'pooling_convention': 'full'
},
ignores=['dilations'],
custom_check=_dimension_constraint())
def _elemwise(name):
"""converting attributes for add operator"""
return AttrCvt(
op_name=_math_name_picker(name),
disables=['axis'],
ignores=['broadcast'])
def _global_pooling(name):
"""Requires kernel attribute which is not present in onnx currently.
So for now giving default kernel."""
return AttrCvt(
op_name='Pooling',
extras={'global_pool': True,
'kernel': (1, 1),
'pool_type': name})
# compatible operators that do NOT require any conversion.
_identity_list = []
# _convert_map defines maps of name to converter functor(callable)
_convert_map = {
# defs/experimental
'FC' : AttrCvt('FullyConnected', ignores=['axis', 'axis_w']),
# defs/generator
'Constant': Renamer('identity'),
'RandomUniform' : AttrCvt('random_uniform', ignores=['seed']),
'RandomNormal' : AttrCvt('random_normal', {'mean':'loc'}, ignores=['seed']),
'RandomUniformLike' : AttrCvt('random_uniform', ignores=['seed']),
'RandomNormalLike': AttrCvt('random_normal', {'mean':'loc'}, ignores=['seed']),
# defs/logical
# defs/math
'Add' : _elemwise('add'),
'Sub' : _elemwise('sub'),
'Mul' : _elemwise('mul'),
'Div' : _elemwise('div'),
def _batch_norm():
"""converting attributes for BatchNorm operator"""
return AttrCvt(op_name='BatchNorm',
transforms={'epsilon': ('eps', (1e-5), _change_eps_cudnn)},
ignores=['spatial', 'is_test', 'consumed_inputs'])
transforms={
'height_scale': ('scale', 1, _upsample_scale_fix),
'mode':
('sample_type', 'nearest', _upsample_restrict_mode),
'width_scale': ('scale', 1, _upsample_scale_fix)
})
# compatible operators that do NOT require any conversion.
_identity_list = []
# _convert_map defines maps of name to converter functor(callable)
_convert_map = {
# defs/experimental
'FC':
AttrCvt('FullyConnected', ignores=['axis', 'axis_w']),
# defs/generator
'Constant':
Renamer('identity'),
'RandomUniform':
AttrCvt('random_uniform', ignores=['seed']),
'RandomNormal':
AttrCvt('random_normal', {'mean': 'loc'}, ignores=['seed']),
'RandomUniformLike':
AttrCvt('random_uniform', ignores=['seed']),
'RandomNormalLike':
AttrCvt('random_normal', {'mean': 'loc'}, ignores=['seed']),
# defs/logical
