def map_crop(cls, node):
offset = node.parameters.offset
if offset:
kwargs = {'offset': int(node.parameters.offset[0])}
return Node.create('crop', **kwargs)
else:
return Node.create('crop')
def map_crop(cls, node):
offset = node.parameters.offset
if offset:
kwargs = {'offset': int(node.parameters.offset[0])}
return Node.create('crop', **kwargs)
else:
return Node.create('crop')
def map_scale(cls, node):
# TODO: The gamma parameter has to be set (in node.data?) and this should work.
# Also, mean should be set to 0, and var to 1, just to be safe.
if node.data:
scale_value = float(node.parameters.filler.value)
if node.parameters.bias_term:
bias_value = float(node.parameters.bias_filler.value)
kwargs = {
'use_scale': True,
'use_bias': node.parameters.bias_term,
'gamma': scale_value,
'beta': bias_value,
'epsilon': 0
}
else:
kwargs = {
'use_scale': True,
'use_bias': node.parameters.bias_term,
'gamma': scale_value,
'epsilon': 0
}
cls._convert_output_shape(kwargs, node)
return Node.create('Affine', **kwargs)
else:
return Node.create('Mul')
def map_scale(cls, node):
# TODO: The gamma parameter has to be set (in node.data?) and this should work.
# Also, mean should be set to 0, and var to 1, just to be safe.
if node.data:
raise NotImplementedError
scale_value = float(node.parameters.filler.value)
kwargs = {
'scale': False,
'bias': False,
'gamma': scale_value,
'epsilon': 0
}
cls._convert_output_shape(kwargs, node)
return Node.create('Scale', **kwargs)
else:
return Node.create('Mul')
def map_eltwise(cls, node):
operations = {0: 'Mul', 1: 'Add', 2: 'Max'}
op_code = node.parameters.operation
try:
return Node.create(operations[op_code])
except KeyError:
raise ConversionError('Unknown elementwise operation: {}'.format(op_code))
def map_batch_norm(cls, node):
scale_offset = len(node.data) == 4
kwargs = {} if scale_offset else {'scale': False, 'bias': False}
epsilon = node.parameters.eps
kwargs['epsilon'] = epsilon
cls._convert_output_shape(kwargs, node)
return Node.create('BatchNorm', **kwargs)
def map_inner_product(cls, node):
#TODO: Axis
assert node.parameters.axis == 1
#TODO: Unbiased
shape = TensorShape()
dim = shape.dim.add()
dim.size = -1
dim = shape.dim.add()
dim.size = 1
for i in node.output_shape[1:]:
dim.size *= i
kwargs = {
'use_bias': node.parameters.bias_term,
'units': node.parameters.num_output,
'_output_shapes': [shape]
}
# check if need the Flatten layer
parent, _ = node.get_only_parent()
ret = []
# if parent.output_shape.height > 1 or parent.output_shape.width > 1:
ret.append(cls._add_flatten_layer(parent))
ret.append(Node.create('FullyConnected', **kwargs))
return ret
def map_scale(cls, node):
raise NotImplementedError
# TODO: The gamma parameter has to be set (in node.data?) and this should work.
# Also, mean should be set to 0, and var to 1, just to be safe.
scale_value = float(node.parameters.filler.value)
kwargs = {'scale' : True, 'bias' : False, 'gamma' : scale_value, 'epsilon': 0}
return Node.create('BatchNorm', **kwargs)
def map_convolution(cls, node):
parent, _ = node.get_only_parent()
kwargs = cls.get_kernel_params(node, parent.output_shape)
kwargs['kernel_shape'] = [node.kernel_parameters.k_h, node.kernel_parameters.k_w, parent.output_shape.channels, node.parameters.num_output]
kwargs['use_bias'] = node.parameters.bias_term
kwargs['group'] = node.parameters.group
return Node.create('Conv', **kwargs)
def map_convolution(cls, node):
parent, _ = node.get_only_parent()
kwargs = cls.get_kernel_params(node, parent.output_shape)
kwargs['kernel_shape'] = [node.kernel_parameters.k_h, node.kernel_parameters.k_w, parent.output_shape.channels, node.parameters.num_output]
kwargs['use_bias'] = node.parameters.bias_term
kwargs['group'] = node.parameters.group
return Node.create('Conv', **kwargs)
def map_eltwise(cls, node):
operations = {0: 'Mul', 1: 'Add', 2: 'Max'}
op_code = node.parameters.operation
try:
return Node.create(operations[op_code])
except KeyError:
raise ConversionError('Unknown elementwise operation: {}'.format(op_code))
def map_scale(cls, node):
raise NotImplementedError
# TODO: The gamma parameter has to be set (in node.data?) and this should work.
# Also, mean should be set to 0, and var to 1, just to be safe.
scale_value = float(node.parameters.filler.value)
kwargs = {'scale' : True, 'bias' : False, 'gamma' : scale_value, 'epsilon': 0}
return Node.create('BatchNorm', **kwargs)
def map_p_re_lu(cls, node):
# print(node.parameters)
# assert False
kwargs = {}
# kwargs['gamma'] = 0.25
cls._convert_output_shape(kwargs, node)
return Node.create('PRelu', **kwargs)
def map_deconvolution(cls, node):
raise NotImplementedError()
parent, _ = node.get_only_parent()
kwargs = cls.get_kernel_params(node, parent.output_shape)
kwargs['kernel_shape'] = [node.kernel_parameters.k_h, node.kernel_parameters.k_w, parent.output_shape.channels, node.parameters.num_output]
kwargs['group'] = node.parameters.group
return Node.create('deconv', **kwargs)
def map_deconvolution(cls, node):
raise NotImplementedError()
parent, _ = node.get_only_parent()
kwargs = cls.get_kernel_params(node, parent.output_shape)
kwargs['kernel_shape'] = [node.kernel_parameters.k_h, node.kernel_parameters.k_w, parent.output_shape.channels, node.parameters.num_output]
kwargs['group'] = node.parameters.group
return Node.create('deconv', **kwargs)
def map_eltwise(cls, node):
operations = {0: 'Mul', 1: 'Add', 2: 'Max'}
op_code = node.parameters.operation
try:
kwargs = {}
cls._convert_output_shape(kwargs, node)
return Node.create(operations[op_code], **kwargs)
except KeyError:
raise ConversionError(
'Unknown elementwise operation: {}'.format(op_code))
def map_convolution(cls, node):
kwargs, padding = cls.get_kernel_params(node)
parent, _ = node.get_only_parent()
kwargs['filter'] = [
node.kernel_parameters.k_h, node.kernel_parameters.k_w,
parent.output_shape.channels, node.parameters.num_output
]
kwargs['use_bias'] = node.parameters.bias_term
group = node.parameters.group
if group != 1:
kwargs['group'] = group
if padding['paddings'] != None:
return [
Node.create('Pad', **padding),
Node.create('Convolution', **kwargs)
]
else:
return Node.create('Convolution', **kwargs)
def map_p_re_lu(cls, node):
# print(node.parameters)
# assert False
try:
scale_value = float(node.parameters.filler.value)
kwargs = {'gamma': scale_value}
except ConversionError:
kwargs = {'gamma': 0.25}
cls._convert_output_shape(kwargs, node)
return Node.create('PRelu', **kwargs)
def map_lrn(cls, node):
params = node.parameters
assert params.local_size % 2 == 1
kwargs = {
'size': int((params.local_size + 1) / 2),
'alpha': params.alpha,
'beta': params.beta,
'k': params.k
}
cls._convert_output_shape(kwargs, node)
return Node.create('LRN', **kwargs)
def _add_flatten_layer(cls, node):
shape = TensorShape()
dim = shape.dim.add()
dim.size = -1
dim = shape.dim.add()
dim.size = 1
for i in node.output_shape[1:]:
dim.size *= i
kwargs = {'_output_shapes': [shape]}
return Node.create('Flatten', **kwargs)
def _add_flatten_layer(cls, node):
shape = TensorShape()
dim = shape.dim.add()
dim.size = -1
dim = shape.dim.add()
dim.size = 1
for i in node.output_shape[1:]:
dim.size *= i
kwargs = {'_output_shapes' : [shape]}
return Node.create('Flatten', **kwargs)
def map_crop(cls, node):
kwargs = {}
cls._convert_output_shape(kwargs, node)
offset = node.parameters.offset
if offset:
if len(offset) == 1:
kwargs['border'] = [offset[0], 0, offset[0], 0]
else:
kwargs['border'] = [offset[0], 0, offset[1], 0]
return Node.create('Crop', **kwargs)
def map_pooling(cls, node):
parent, _ = node.get_only_parent()
kwargs = cls.get_kernel_params(node, parent.output_shape)
if node.parameters.pool == 0:
kwargs['pooling_type'] = 'MAX'
elif node.parameters.pool == 1:
kwargs['pooling_type'] = 'AVG'
else:
# Stochastic pooling, for instance.
raise ConversionError('Unsupported pooling type.')
cls._convert_output_shape(kwargs, node)
return Node.create('Pool', **kwargs)
def map_deconvolution(cls, node):
parent, _ = node.get_only_parent()
kwargs = cls.get_kernel_params(node, parent.output_shape)
kwargs['kernel_shape'] = [node.kernel_parameters.k_h, node.kernel_parameters.k_w, node.parameters.num_output, parent.output_shape.channels]
kwargs['use_bias'] = node.parameters.bias_term
if node.parameters.dilation:
dilation = node.parameters.dilation[0]
if dilation != 1:
kwargs['dilations'] = [1, dilation, dilation, 1]
kwargs['group'] = node.parameters.group
return Node.create('ConvTranspose', **kwargs)
def map_deconvolution(cls, node):
parent, _ = node.get_only_parent()
kwargs = cls.get_kernel_params(node, parent.output_shape)
kwargs['kernel_shape'] = [node.kernel_parameters.k_h, node.kernel_parameters.k_w, node.parameters.num_output, parent.output_shape.channels]
kwargs['use_bias'] = node.parameters.bias_term
if node.parameters.dilation:
dilation = node.parameters.dilation[0]
if dilation != 1:
kwargs['dilations'] = [1, dilation, dilation, 1]
kwargs['group'] = node.parameters.group
return Node.create('ConvTranspose', **kwargs)
def map_pooling(cls, node):
parent, _ = node.get_only_parent()
kwargs = cls.get_kernel_params(node, parent.output_shape)
if node.parameters.pool == 0:
kwargs['pooling_type'] = 'MAX'
elif node.parameters.pool == 1:
kwargs['pooling_type'] = 'AVG'
else:
# Stochastic pooling, for instance.
raise ConversionError('Unsupported pooling type.')
cls._convert_output_shape(kwargs, node)
return Node.create('Pool', **kwargs)
def map_pooling(cls, node):
kwargs, padding = cls.get_kernel_params(node)
if node.parameters.pool == 0:
kwargs['pooling_type'] = 'MAX'
elif node.parameters.pool == 1:
kwargs['pooling_type'] = 'AVG'
else:
# Stochastic pooling, for instance.
raise ConversionError('Unsupported pooling type.')
kwargs['window_shape'] = [
1, node.kernel_parameters.k_h, node.kernel_parameters.k_w, 1
]
cls._convert_output_shape(kwargs, node)
if padding['paddings'] != None:
return [
Node.create('Pad', **padding),
Node.create('Pool', **kwargs)
]
else:
return Node.create('Pool', **kwargs)
def map_unpooling(cls, node):
kwargs = {}
kwargs['kernel_shape'] = [
1, node.kernel_parameters.k_h, node.kernel_parameters.k_w, 1
]
kwargs['pads'] = [
0, node.kernel_parameters.p_h, node.kernel_parameters.p_w, 0
]
kwargs['strides'] = [
1, node.kernel_parameters.s_h, node.kernel_parameters.s_w, 1
]
cls._convert_output_shape(kwargs, node)
return Node.create('Unpool', **kwargs)
def map_inner_product(cls, node):
#TODO: Axis
assert node.parameters.axis == 1
#TODO: Unbiased
kwargs = {'use_bias' : node.parameters.bias_term, 'units' : node.parameters.num_output}
# check if need the Flatten layer
parent, _ = node.get_only_parent()
ret = []
# if parent.output_shape.height > 1 or parent.output_shape.width > 1:
ret.append(cls._add_flatten_layer(parent))
ret.append(Node.create('FullyConnected', **kwargs))
return ret
def map_inner_product(cls, node):
#TODO: Axis
assert node.parameters.axis == 1
#TODO: Unbiased
kwargs = {'use_bias' : node.parameters.bias_term, 'units' : node.parameters.num_output}
# check if need the Flatten layer
parent, _ = node.get_only_parent()
ret = []
# if parent.output_shape.height > 1 or parent.output_shape.width > 1:
ret.append(cls._add_flatten_layer(parent))
ret.append(Node.create('FullyConnected', **kwargs))
return ret
def map_data(cls, node):
# TODO: We need to identify whether this is 4D image data, otherwise we shouldn't change the dimension order
shape = TensorShape()
dim = shape.dim.add()
dim.size = -1
for i in node.output_shape[2:]:
dim = shape.dim.add()
dim.size = i
dim = shape.dim.add()
dim.size = node.output_shape.channels
kwargs = {'shape': shape} # Ignore the dimension of batch size
cls._convert_output_shape(kwargs, node)
return Node.create('DataInput', **kwargs)
def map_data(cls, node):
# TODO: We need to identify whether this is 4D image data, otherwise we shouldn't change the dimension order
shape = TensorShape()
dim = shape.dim.add()
dim.size = -1
for i in node.output_shape[2:]:
dim = shape.dim.add()
dim.size = i
dim = shape.dim.add()
dim.size = node.output_shape.channels
kwargs = {'shape': shape} # Ignore the dimension of batch size
cls._convert_output_shape(kwargs, node)
return Node.create('DataInput', **kwargs)
def map_relu(cls, node):
kwargs = {}
cls._convert_output_shape(kwargs, node)
return Node.create('Relu', **kwargs)
def map_reshape(cls, node):
kwargs = {'shape' : [dim for dim in node.output_shape]}
cls._convert_output_shape(kwargs, node)
return Node.create('Reshape', **kwargs)
def map_flatten(cls, node):
return Node.create('Flatten')
def map_relu(cls, node):
kwargs = {}
cls._convert_output_shape(kwargs, node)
return Node.create('Relu', **kwargs)
def map_dropout(cls, node):
kwargs = {'keep_prob': node.parameters.dropout_ratio}
cls._convert_output_shape(kwargs, node)
return Node.create('Dropout', **kwargs)
def map_abs_val(cls, node):
return Node.create('Abs')
def map_sigmoid(cls, node):
return Node.create('Sigmoid')
def map_abs_val(cls, node):
return Node.create('Abs')
def map_sigmoid(cls, node):
return Node.create('Sigmoid')
def map_tanh(cls, node):
return Node.create('Tanh')
def map_softmax(cls, node):
kwargs = {}
cls._convert_output_shape(kwargs, node)
return Node.create('Softmax', **kwargs)
def map_softmax(cls, node):
kwargs = {}
cls._convert_output_shape(kwargs, node)
return Node.create('Softmax', **kwargs)
def map_concat(cls, node):
kwargs = {'axis': (2, 3, 1, 0)[node.parameters.axis]}
cls._convert_output_shape(kwargs, node)
return Node.create('Concat', **kwargs)
def map_lrn(cls, node):
params = node.parameters
assert params.local_size % 2 == 1
kwargs = {'size': int((params.local_size + 1) / 2), 'alpha': params.alpha, 'beta': params.beta, 'k' : params.k}
cls._convert_output_shape(kwargs, node)
return Node.create('LRN', **kwargs)
def map_batch_norm(cls, node):
kwargs = {'scale': len(node.data) >= 3, 'bias': len(node.data) == 4}
epsilon = node.parameters.eps
kwargs['epsilon'] = epsilon
cls._convert_output_shape(kwargs, node)
return Node.create('BatchNorm', **kwargs)
def map_concat(cls, node):
kwargs = {'axis': (2, 3, 1, 0)[node.parameters.axis]}
cls._convert_output_shape(kwargs, node)
return Node.create('Concat', **kwargs)
def map_tanh(cls, node):
return Node.create('Tanh')
def map_dropout(cls, node):
kwargs = {'keep_prob': node.parameters.dropout_ratio}
cls._convert_output_shape(kwargs, node)
return Node.create('Dropout', **kwargs)
def map_batch_norm(cls, node):
kwargs = {'scale' : len(node.data) >= 3, 'bias' : len(node.data) == 4}
epsilon = node.parameters.eps
kwargs['epsilon'] = epsilon
cls._convert_output_shape(kwargs, node)
return Node.create('BatchNorm', **kwargs)
def map_softmax(cls, node):
return Node.create('Softmax')