def __init__(self, meta_file, checkpoint_file, frozen_file, dest_nodes = None):
super(TensorflowParser, self).__init__()
# load model files into TensorFlow graph
if meta_file:
model = TensorflowParser._load_meta(meta_file)
if checkpoint_file:
self.ckpt_data = TensorflowParser._load_weights(checkpoint_file)
self.weight_loaded = True
if dest_nodes != None:
from tensorflow.python.framework.graph_util import extract_sub_graph
model = extract_sub_graph(model, dest_nodes.split(','))
# Build network graph
self.tf_graph = TensorflowGraph(model)
self.tf_graph.build()
def __init__(self, input_args, dest_nodes = None):
super(TensorflowParser, self).__init__()
# load model files into Keras graph
from six import string_types as _string_types
if isinstance(input_args, _string_types):
model = TensorflowParser._load_meta(input_args)
elif isinstance(input_args, tuple):
model = TensorflowParser._load_meta(input_args[0])
self.ckpt_data = TensorflowParser._load_weights(input_args[1])
self.weight_loaded = True
if dest_nodes != None:
from tensorflow.python.framework.graph_util import extract_sub_graph
model = extract_sub_graph(model, dest_nodes.split(','))
# Build network graph
self.tf_graph = TensorflowGraph(model)
self.tf_graph.build()
class TensorflowParser(Parser):
skip_prefix = [
"^",
"train_op",
"save",
"gradients",
"init",
"global_step",
"distort_image",
"Adagrad",
]
skip_scope = [
"random_uniform",
"Initializer",
"optimizer",
"weight_loss",
"parallel_read",
"case"
]
skip_type = set([
"L2Loss",
"VariableV2",
"Const",
"Assign",
"RandomUniform",
"FIFOQueueV2"
])
dtype_map = {
0 : graph_pb2.DT_UNDEFINED,
1 : graph_pb2.DT_FLOAT32,
2 : graph_pb2.DT_FLOAT64,
3 : graph_pb2.DT_INT32,
4 : graph_pb2.DT_UINT8,
5 : graph_pb2.DT_INT16,
6 : graph_pb2.DT_INT8,
7 : graph_pb2.DT_STRING,
9 : graph_pb2.DT_INT64
}
@property
def src_graph(self):
return self.tf_graph
@staticmethod
def _load_meta(model_network_path):
"""Load a tensorflow meta file from disk
Parameters
----------
model_network_path: str
Path where the model network path is (protobuf meta file)
Returns
-------
model: A tensorflow protobuf file
"""
from tensorflow.core.protobuf import meta_graph_pb2
from mmdnn.conversion.common.IR.IR_graph import load_protobuf_from_file
meta_graph = meta_graph_pb2.MetaGraphDef()
load_protobuf_from_file(meta_graph, model_network_path)
graph = meta_graph.graph_def
print ("Tensorflow model file [%s] loaded successfully." % model_network_path)
return graph
@staticmethod
def _load_weights(model_weight_path):
"""Load a tensorflow checkpoint file from disk
Parameters
----------
model_weight_path: str
Path where the weight path is (checkpoint file)
Returns
-------
model: tensor name --> ndarry
"""
reader = tensorflow.train.NewCheckpointReader(model_weight_path)
var_to_shape_map = reader.get_variable_to_shape_map()
data = dict()
for name in var_to_shape_map:
tensor = reader.get_tensor(name)
data[name] = tensor
print ("Tensorflow checkpoint file [%s] loaded successfully. [%d] variables loaded." % (model_weight_path, len(data)))
return data
@staticmethod
def _get_scopes(layer_name):
return layer_name.split('/')
def _convert_reduction_operators(self, source_node, new_op = None):
IR_node = self._convert_identity_operation(source_node, 1, new_op)
# keep dims
IR_node.attr['keepdims'].b = source_node.layer.attr['keep_dims'].b
# axes
axes = self.get_parent(source_node.name, [1]).layer.attr['value'].tensor
axes = tensor_util.MakeNdarray(axes)
IR_node.attr['axes'].list.i.extend(axes)
def _convert_layers_batchnorm(self, source_node):
# name, op
IR_node = self.IR_graph.node.add()
TensorflowParser._copy_and_reop(source_node, IR_node, 'BatchNorm')
# epsilon
epsilon = self.get_parent(source_node.name, [1])
IR_node.attr['epsilon'].f = epsilon.layer.attr['value'].tensor.float_val[0]
# moving variance (var)
moving_variance = self.get_parent(source_node.name, [0, 0])
if self.weight_loaded:
self.set_weight(source_node.name, 'var', self.ckpt_data[moving_variance.name])
# gamma (scale)
gamma = self.get_son(source_node.name, [0, 0], True)
gamma = self.get_parent(gamma.name, [1, 0], True)
if gamma is None or not gamma.type.startswith('Variable'):
IR_node.attr['scale'].b = False
output_node = self.get_son(source_node.name, [0, 0, 0], True)
else:
IR_node.attr['scale'].b = True
if self.weight_loaded:
self.set_weight(source_node.name, 'scale', self.ckpt_data[gamma.name])
output_node = self.get_son(source_node.name, [0, 0, 0, 0], True)
# mean
mean = self.get_parent(output_node.name, [1, 1, 0, 0], True)
if self.weight_loaded:
self.set_weight(source_node.name, 'mean', self.ckpt_data[mean.name])
# bias
bias = self.get_parent(output_node.name, [1, 0, 0], True)
if bias is None or not bias.type.startswith('Variable'):
IR_node.attr['bias'].b = False
else:
IR_node.attr['bias'].b = True
if self.weight_loaded:
self.set_weight(source_node.name, 'bias', self.ckpt_data[bias.name])
# input node
assert output_node.type == 'Add'
input_node = self.get_parent(output_node.name, [0, 0])
IR_node.input.append(input_node.real_name)
# output node
output_node.real_name = source_node.name
def __init__(self, input_args, dest_nodes = None):
super(TensorflowParser, self).__init__()
# load model files into Keras graph
from six import string_types as _string_types
if isinstance(input_args, _string_types):
model = TensorflowParser._load_meta(input_args)
elif isinstance(input_args, tuple):
model = TensorflowParser._load_meta(input_args[0])
self.ckpt_data = TensorflowParser._load_weights(input_args[1])
self.weight_loaded = True
if dest_nodes != None:
from tensorflow.python.framework.graph_util import extract_sub_graph
model = extract_sub_graph(model, dest_nodes.split(','))
# Build network graph
self.tf_graph = TensorflowGraph(model)
self.tf_graph.build()
@classmethod
def _skip_node(cls, source_node):
if source_node.covered:
return True
for prefix in cls.skip_prefix:
if source_node.name.startswith(prefix):
return True
scopes = TensorflowParser._get_scopes(source_node.name)
for s in scopes:
if s in cls.skip_scope:
return True
return False
@staticmethod
def tensor_shape_to_list(shapes):
if isinstance(shapes, attr_value_pb2.AttrValue):
return [dim.size for dim in shapes.shape.dim]
else:
ret = []
for shape in shapes:
this_one = [dim.size for dim in shape.dim]
ret.append(this_one)
return ret
def _convert_padding(self, source_node, IR_node, kernel_size):
# TODO: Fused conv and pool with padding is different from defused operators
input_node = self.get_parent(source_node.name, [0])
input_shape = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
if source_node.get_attr('padding') == 'VALID':
dims = len(input_shape)
assign_IRnode_values(IR_node, {'auto_pad' : "VALID", 'pads' : [0, 0] * dims})
elif source_node.get_attr('padding') == 'SAME':
padding = compute_tf_same_padding(
input_shape,
kernel_size,
source_node.get_attr('strides'))
assign_IRnode_values(IR_node, {'auto_pad' : "SAME_LOWER", 'pads' : padding})
else:
assert False
def _convert_pooling(self, source_node, pool_type):
IR_node = self._convert_identity_operation(source_node, new_op='Pool')
kwargs = {}
# strides
kwargs['strides'] = source_node.get_attr('strides')
# window_shape
kwargs['kernel_shape'] = source_node.get_attr('ksize')
# pool type
kwargs['pooling_type'] = pool_type
# padding
self._convert_padding(source_node, IR_node, kwargs['kernel_shape'][1:-1])
assign_IRnode_values(IR_node, kwargs)
def gen_IR(self):
for layer in self.src_graph.topological_sort:
current_node = self.src_graph.get_node(layer)
if self._skip_node(current_node):
continue
node_type = current_node.type
if hasattr(self, "rename_" + node_type):
func = getattr(self, "rename_" + node_type)
func(current_node)
else:
self.rename_UNKNOWN(current_node)
@staticmethod
def _copy_and_reop(source_node, IR_node, new_op = None):
if new_op == None: new_op = source_node.type
IR_node.name = source_node.name
IR_node.op = new_op
kwargs = {}
if 'data_format' in source_node.layer.attr:
kwargs['data_format'] = source_node.get_attr('data_format')
if 'dtype' in source_node.layer.attr:
assert source_node.layer.attr['dtype'].type in TensorflowParser.dtype_map, 'type [{}] is unknown.'.format(source_node.layer.attr['dtype'].type)
IR_node.attr["dtype"].type = TensorflowParser.dtype_map[source_node.layer.attr['dtype'].type]
if '_output_shapes' in source_node.layer.attr:
IR_node.attr["_output_shapes"].MergeFromString(source_node.layer.attr['_output_shapes'].SerializeToString())
assign_IRnode_values(IR_node, kwargs)
def _convert_inedge(self, source_node, IR_node, start_idx = 0, end_idx = None):
if end_idx == None: end_idx = len(source_node.in_edges)
for idx in range(start_idx, end_idx):
IR_node.input.append(self.src_graph.get_node(source_node.in_edges[idx]).real_name)
def _get_bias(self, source_node, IR_node):
if not source_node.out_edges:
return
add_node = self.tf_graph.get_node(source_node.out_edges[0])
if add_node.type != "Add" and add_node.type != "BiasAdd":
return
variable = self.tf_graph.get_node(add_node.in_edges[1])
variable = self.tf_graph.get_node(variable.in_edges[0])
assert variable.layer.attr['shape'].shape.dim[0].size == IR_node.attr['kernel_shape'].list.i[-1]
if self.weight_loaded:
assert variable.name in self.ckpt_data
current_layer = self.weights[source_node.name]
current_layer['bias'] = self.ckpt_data[variable.name]
add_node.real_name = IR_node.name
add_node.covered = True
IR_node.attr['use_bias'].b = True
@staticmethod
def _copy_shape(source_node, IR_node):
assert 'shape' in source_node.layer.attr
IR_node.attr['shape'].shape.MergeFromString(source_node.layer.attr['shape'].shape.SerializeToString())
def rename_UNKNOWN(self, source_node):
if source_node.type in self.skip_type:
return
print("Tensorflow has not supported operator [%s] with name [%s]."
% (source_node.type, source_node.name))
return
def rename_Placeholder(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op='DataInput')
# shape
TensorflowParser._copy_shape(source_node, IR_node)
def rename_Conv2D(self, source_node):
"""
weights: name_weights, name_bias
"""
IR_node = self._convert_identity_operation(source_node, 1, 'Conv')
kwargs = {}
# strides
kwargs['strides'] = source_node.get_attr('strides')
# input[1] : W
# filter
W = self.tf_graph.get_node(source_node.layer.input[1])
W = self.tf_graph.get_node(W.layer.input[0]).layer
kwargs['kernel_shape'] = self.tensor_shape_to_list(W.attr['shape'])
# padding
self._convert_padding(source_node, IR_node, kwargs['kernel_shape'][:-2])
if self.weight_loaded:
self.set_weight(source_node.name, 'weights', self.ckpt_data[W.name])
assign_IRnode_values(IR_node, kwargs)
# output[0] : B
self._get_bias(source_node, IR_node)
def _convert_identity_operation(self, source_node, in_edge_count = None, new_op = None):
IR_node = self.IR_graph.node.add()
TensorflowParser._copy_and_reop(source_node, IR_node, new_op)
self._convert_inedge(source_node, IR_node, 0, in_edge_count)
return IR_node
def rename_Relu(self, source_node):
self._convert_identity_operation(source_node)
def rename_Relu6(self, source_node):
self._convert_identity_operation(source_node)
def rename_Add(self, source_node):
if not source_node.covered:
scopes = self._get_scopes(source_node.name)
if len(scopes) < 3:
self._convert_identity_operation(source_node)
elif scopes[-2] == 'dropout':
# converted [dropout]
pass
elif scopes[-2] == 'batchnorm':
# convert [tf.contrib.layers.batch_norm]
self._convert_layers_batchnorm(source_node)
else:
# normal Add
self._convert_identity_operation(source_node)
def rename_Sub(self, source_node):
self._convert_identity_operation(source_node)
def rename_Reshape(self, source_node):
IR_node = self._convert_identity_operation(source_node, 1)
kwargs = {'shape' : self.tensor_shape_to_list(source_node.get_attr('_output_shapes'))[0]}
assign_IRnode_values(IR_node, kwargs)
def rename_MatMul(self, source_node):
"""
weights: name_weights, name_bias
"""
IR_node = self._convert_identity_operation(source_node, 1)
# units
units = source_node.layer.attr['_output_shapes'].list.shape[-1].dim[-1].size
IR_node.attr['units'].i = units
# Weights
W = self.tf_graph.get_node(self.tf_graph.get_node(source_node.in_edges[1]).in_edges[0])
if self.weight_loaded:
self.set_weight(source_node.name, 'weights', self.ckpt_data[W.name])
if source_node.out_edges:
add_node = self.tf_graph.get_node(source_node.out_edges[0])
if add_node.type == 'Add':
add_node.covered = True
add_node.real_name = source_node.real_name
# FullyConnected Layer
# name, op
TensorflowParser._copy_and_reop(source_node, IR_node, 'FullyConnected')
# get Bias
B = self.tf_graph.get_node(self.tf_graph.get_node(source_node.out_edges[0]).in_edges[1]).in_edges[0]
if self.weight_loaded:
self.set_weight(source_node.name, 'bias', self.ckpt_data[B])
IR_node.attr['use_bias'].b = True
else:
raise NotImplementedError("Not implemented yet. Please submit a issue in github and provide your models for reproduce.")
else:
# Matmul Layer
TensorflowParser._copy_and_reop(source_node, IR_node, 'FullyConnected')
assign_IRnode_values(IR_node, {'use_bias' : False})
def rename_RealDiv(self, source_node):
scopes = self._get_scopes(source_node.name)
# Deal Dropout
if scopes[-2] == 'dropout':
IR_node = self._convert_identity_operation(source_node, 1, 'Dropout')
# keep prob
if 'value' in self.tf_graph.get_node(source_node.layer.input[1]).layer.attr:
IR_node.attr['keep_prob'].f = self.tf_graph.get_node(source_node.layer.input[1]).layer.attr['value'].tensor.float_val[0]
else:
IR_node.attr['keep_prob'].f = 1.0
# Remove nodes
# Mul
mul_node = self.tf_graph.get_node(source_node.out_edges[0])
assert mul_node.type == "Mul"
mul_node.covered = True
mul_node.real_name = source_node.name
# Floor
floor_node = self.tf_graph.get_node(mul_node.in_edges[1])
assert floor_node.type == "Floor"
floor_node.covered = True
else:
assert False
def rename_Floor(self, source_node):
scopes = self._get_scopes(source_node.name)
assert scopes[-2] == 'dropout'
def rename_MaxPool(self, source_node):
self._convert_pooling(source_node, b'MAX')
def rename_AvgPool(self, source_node):
self._convert_pooling(source_node, b'AVG')
def rename_Identity(self, source_node):
source_node.real_name = self.src_graph.get_node(source_node.in_edges[0]).real_name
def rename_Squeeze(self, source_node):
IR_node = self._convert_identity_operation(source_node)
IR_node.attr['axes'].MergeFromString(source_node.layer.attr['squeeze_dims'].SerializeToString())
def rename_QueueDequeueManyV2(self, source_node):
IR_node = self._convert_identity_operation(source_node, 0, 'DataInput')
IR_node.attr['shape'].shape.MergeFromString(source_node.layer.attr['_output_shapes'].list.shape[0].SerializeToString())
IR_node.attr['shape'].shape.dim[0].size = -1
IR_node.attr['dtype'].type = self.dtype_map[source_node.layer.attr['component_types'].list.type[0]]
def rename_Pad(self, source_node):
IR_node = self._convert_identity_operation(source_node, 1, 'Pad')
kwargs = {}
kwargs['mode'] = 'constant'
kwargs['constant_values'] = 0.0
# paddings
padding = self.get_parent(source_node.name, [1]).layer.attr['value'].tensor
shapes = tensor_util.MakeNdarray(padding)
kwargs['pads'] = convert_tf_pad_to_onnx(shapes)
assign_IRnode_values(IR_node, kwargs)
def rename_Mean(self, source_node):
self._convert_reduction_operators(source_node, new_op = 'ReduceMean')
def rename_ConcatV2(self, source_node):
n = len(source_node.in_edges) - 1
IR_node = self._convert_identity_operation(source_node, n, 'Concat')
axis = self.tf_graph.get_parent(source_node.name, [n])
IR_node.attr['axis'].i = axis.layer.attr['value'].tensor.int_val[0]
def rename_DepthwiseConv2dNative(self, source_node):
IR_node = self._convert_identity_operation(source_node, 1, 'DepthwiseConv')
kwargs = {}
kwargs['strides'] = source_node.get_attr('strides')
input_node = self.src_graph.get_parent(source_node.name, [1])
kwargs['kernel_shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
self._convert_padding(source_node, IR_node, kwargs['kernel_shape'][:-2])
if self.weight_loaded:
weight = self.src_graph.get_parent(source_node.name, [1, 0])
self.set_weight(source_node.name, 'weights', self.ckpt_data[weight.name])
assign_IRnode_values(IR_node, kwargs)
def rename_FusedBatchNorm(self, source_node):
IR_node = self._convert_identity_operation(source_node, 1, 'BatchNorm')
IR_node.attr['epsilon'].f = source_node.get_attr('epsilon', 0)
# gamma (scale)
scale = self.get_parent(source_node.name, [1], True)
if scale.type == 'Const':
value = scale.get_attr('value')
shape = value.tensor_shape
assert len(shape.dim) == 1
shape = shape.dim[0].size
assert len(value.float_val) == 1
value = value.float_val[0]
if np.isclose(value, 1.0):
IR_node.attr['scale'].b = False
else:
IR_node.attr['scale'].b = True
if self.weight_loaded:
self.set_weight(source_node.name, 'scale', np.array([value] * shape))
else:
scale = self.get_parent(scale.name, [0], True)
if self.weight_loaded:
self.set_weight(source_node.name, 'scale', self.ckpt_data[scale.name])
IR_node.attr['scale'].b = True
# bias
bias = self.get_parent(source_node.name, [2, 0], True)
IR_node.attr['bias'].b = True
# Mean
mean = self.get_parent(source_node.name, [3, 0], True)
# Var
var = self.get_parent(source_node.name, [4, 0], True)
if self.weight_loaded:
self.set_weight(source_node.name, 'bias', self.ckpt_data[bias.name])
self.set_weight(source_node.name, 'mean', self.ckpt_data[mean.name])
self.set_weight(source_node.name, 'var', self.ckpt_data[var.name])
def __init__(self, frozen_file, inputshape, in_nodes, dest_nodes):
super(TensorflowParser2, self).__init__()
self.weight_loaded = True
# load model files into TensorFlow graph
with open(frozen_file, 'rb') as f:
serialized = f.read()
tensorflow.reset_default_graph()
original_gdef = tensorflow.GraphDef()
original_gdef.ParseFromString(serialized)
# model = original_gdef
from tensorflow.python.tools import strip_unused_lib
from tensorflow.python.framework import dtypes
from tensorflow.python.platform import gfile
input_node_names = in_nodes.split(',')
output_node_names = dest_nodes.split(',')
gdef = strip_unused_lib.strip_unused(
input_graph_def = original_gdef,
input_node_names = input_node_names,
output_node_names = output_node_names,
placeholder_type_enum = dtypes.float32.as_datatype_enum)
# Save it to an output file
frozen_model_file = './frozen.pb'
with gfile.GFile(frozen_model_file, "wb") as f:
f.write(gdef.SerializeToString())
with open(frozen_model_file, 'rb') as f:
serialized = f.read()
tensorflow.reset_default_graph()
model = tensorflow.GraphDef()
model.ParseFromString(serialized)
output_shape_map = dict()
input_shape_map = dict()
with tensorflow.Graph().as_default() as g:
tensorflow.import_graph_def(original_gdef, name='')
ops = g.get_operations()
N = len(ops)
p = 0
for i in range(N):
for x in ops[i].inputs:
input_shape_map[x.name] = x.get_shape()
for x in ops[i].outputs:
output_shape_map[x.name] = x.get_shape()
tensor_input = tensorflow.TensorShape([None, tensorflow.Dimension(inputshape[0]), tensorflow.Dimension(inputshape[1]), tensorflow.Dimension(inputshape[2])])
output_shape_map[input_node_names[0]] = tensor_input
self.tf_graph = TensorflowGraph(model)
for node in self.tf_graph.model.node:
if node.name == 'input' and node.op == 'Placeholder':
node.attr['shape'].list.shape.extend([output_shape_map[node.name + ':0'].as_proto()])
if (node.name + ':0') in output_shape_map:
node.attr['_output_shapes'].list.shape.extend([output_shape_map[node.name + ':0'].as_proto()])
if node.op == 'MirrorPad':
node.attr['paddings'].list.shape.extend([input_shape_map[node.name + '/paddings:0'].as_proto()])
if node.op == 'QuantizeV2':
node.attr['shape'].list.shape.extend([input_shape_map[node.name + ':0'].as_proto()])
if node.op == 'RequantizationRange':
map_key = node.name.split('eightbit')[0] + "eightbit_quantized_conv:0"
node.attr['shape'].list.shape.extend([input_shape_map[map_key].as_proto()])
if node.op == 'Requantize':
map_key = node.name.replace("requantize", "quantized_conv")+":0"
node.attr['shape'].list.shape.extend([input_shape_map[map_key].as_proto()])
if node.name == in_nodes:
node.attr['shape'].list.shape.extend([output_shape_map[node.name].as_proto()])
self.tf_graph.build()
def __init__(self,
meta_file,
checkpoint_file,
dest_nodes,
inputShape=None,
in_nodes=None):
super(TensorflowParser, self).__init__()
# load model files into TensorFlow graph
if meta_file:
model = TensorflowParser._load_meta(meta_file)
if checkpoint_file:
self.ckpt_data = TensorflowParser._load_weights(checkpoint_file)
self.weight_loaded = True
# extract subgraph using in_nodes and dest_nodes
if in_nodes != None and inputShape != None:
from tensorflow.python.tools import strip_unused_lib
from tensorflow.python.framework import dtypes
from tensorflow.python.platform import gfile
model = strip_unused_lib.strip_unused(
input_graph_def=model,
input_node_names=in_nodes,
output_node_names=dest_nodes,
placeholder_type_enum=dtypes.float32.as_datatype_enum)
input_list = [None]
for i in range(len(inputShape)):
input_list.append(tensorflow.Dimension(inputShape[i]))
tensor_input = tensorflow.TensorShape(input_list)
# Build network graph
self.tf_graph = TensorflowGraph(model)
for node in self.tf_graph.model.node:
if node.name in in_nodes:
node.attr['shape'].shape.CopyFrom(tensor_input.as_proto())
node.attr['_output_shapes'].list.shape.pop(
) #unknown_rank pop
node.attr['_output_shapes'].list.shape.extend(
[tensor_input.as_proto()])
# extract subgraph using dest_nodes
elif dest_nodes != None:
from tensorflow.python.framework.graph_util import extract_sub_graph
model = extract_sub_graph(model, dest_nodes)
# Get input node name
if not in_nodes:
in_nodes = []
for node in model.node:
if node.op == 'Placeholder':
in_nodes.append(node.name)
# Graph Transform
transforms = ["fold_constants(ignore_errors=true)"]
transformed_graph_def = TransformGraph(model, in_nodes, dest_nodes,
transforms)
in_type_list = {}
in_shape_list = {}
for n in transformed_graph_def.node:
if n.name in in_nodes:
in_type_list[n.name] = n.attr['dtype'].type
in_node_shape = n.attr['shape'].shape
in_node_shape_str = self._shapeToStr(in_node_shape)
in_shape_list[n.name] = in_node_shape_str
dtype = tensorflow.float32
with tensorflow.Graph().as_default() as g:
input_map = {}
for in_node in in_nodes:
if in_type_list[in_node] == 1 or in_type_list[in_node] == 0:
dtype = tensorflow.float32
elif in_type_list[in_node] == 3:
dtype = tensorflow.int32
elif in_type_list[in_node] == 10:
dtype = tensorflow.bool
x = tensorflow.placeholder(dtype, shape=in_shape_list[in_node])
input_map[in_node] = x
tensorflow.import_graph_def(transformed_graph_def,
name='',
input_map=input_map)
with tensorflow.Session(graph=g) as sess:
tempdir = tempfile.mkdtemp()
meta_graph_def = tensorflow.train.export_meta_graph(
filename=os.path.join(tempdir, 'my-model.meta'))
model = meta_graph_def.graph_def
shutil.rmtree(tempdir)
self.tf_graph = TensorflowGraph(model)
self.tf_graph.build()
process_graph(self.tf_graph, self.ckpt_data)
class TensorflowParser2(Parser):
# skip_prefix = [
# "^",
# "train_op",
# "save",
# "gradients",
# "init",
# "global_step",
# "distort_image",
# "Adagrad",
# ]
skip_prefix = [
]
skip_scope = [
"random_uniform",
"Initializer",
"optimizer",
"weight_loss",
"parallel_read",
"case"
]
skip_type = set([
"L2Loss",
"VariableV2",
"Const",
"Assign",
"RandomUniform",
"FIFOQueueV2",
"Assert",
"Unpack",
"NextIteration",
"TensorArrayV3",
"Range",
"TensorArrayScatterV3",
"TensorArrayReadV3",
"TensorArrayWriteV3",
# "Switch"
# "Dequantize",
# "RequantizationRange",
# "Requantize",
"ExpandDims",
"Identity",
# "Cast"
])
q_type = set([
"Dequantize",
"QuantizeV2",
"QuantizedConv2D",
"QuantizedReshape",
"RequantizationRange"
])
dtype_map = {
0 : graph_pb2.DT_UNDEFINED,
1 : graph_pb2.DT_FLOAT32,
2 : graph_pb2.DT_FLOAT64,
3 : graph_pb2.DT_INT32,
4 : graph_pb2.DT_UINT8,
5 : graph_pb2.DT_INT16,
6 : graph_pb2.DT_INT8,
7 : graph_pb2.DT_STRING,
9 : graph_pb2.DT_INT64
}
@property
def src_graph(self):
return self.tf_graph
def __init__(self, frozen_file, inputshape, in_nodes, dest_nodes):
super(TensorflowParser2, self).__init__()
self.weight_loaded = True
# load model files into TensorFlow graph
with open(frozen_file, 'rb') as f:
serialized = f.read()
tensorflow.reset_default_graph()
original_gdef = tensorflow.GraphDef()
original_gdef.ParseFromString(serialized)
# model = original_gdef
from tensorflow.python.tools import strip_unused_lib
from tensorflow.python.framework import dtypes
from tensorflow.python.platform import gfile
input_node_names = in_nodes.split(',')
output_node_names = dest_nodes.split(',')
gdef = strip_unused_lib.strip_unused(
input_graph_def = original_gdef,
input_node_names = input_node_names,
output_node_names = output_node_names,
placeholder_type_enum = dtypes.float32.as_datatype_enum)
# Save it to an output file
frozen_model_file = './frozen.pb'
with gfile.GFile(frozen_model_file, "wb") as f:
f.write(gdef.SerializeToString())
with open(frozen_model_file, 'rb') as f:
serialized = f.read()
tensorflow.reset_default_graph()
model = tensorflow.GraphDef()
model.ParseFromString(serialized)
output_shape_map = dict()
input_shape_map = dict()
with tensorflow.Graph().as_default() as g:
tensorflow.import_graph_def(original_gdef, name='')
ops = g.get_operations()
N = len(ops)
p = 0
for i in range(N):
for x in ops[i].inputs:
input_shape_map[x.name] = x.get_shape()
for x in ops[i].outputs:
output_shape_map[x.name] = x.get_shape()
tensor_input = tensorflow.TensorShape([None, tensorflow.Dimension(inputshape[0]), tensorflow.Dimension(inputshape[1]), tensorflow.Dimension(inputshape[2])])
output_shape_map[input_node_names[0]] = tensor_input
self.tf_graph = TensorflowGraph(model)
for node in self.tf_graph.model.node:
if node.name == 'input' and node.op == 'Placeholder':
node.attr['shape'].list.shape.extend([output_shape_map[node.name + ':0'].as_proto()])
if (node.name + ':0') in output_shape_map:
node.attr['_output_shapes'].list.shape.extend([output_shape_map[node.name + ':0'].as_proto()])
if node.op == 'MirrorPad':
node.attr['paddings'].list.shape.extend([input_shape_map[node.name + '/paddings:0'].as_proto()])
if node.op == 'QuantizeV2':
node.attr['shape'].list.shape.extend([input_shape_map[node.name + ':0'].as_proto()])
if node.op == 'RequantizationRange':
map_key = node.name.split('eightbit')[0] + "eightbit_quantized_conv:0"
node.attr['shape'].list.shape.extend([input_shape_map[map_key].as_proto()])
if node.op == 'Requantize':
map_key = node.name.replace("requantize", "quantized_conv")+":0"
node.attr['shape'].list.shape.extend([input_shape_map[map_key].as_proto()])
if node.name == in_nodes:
node.attr['shape'].list.shape.extend([output_shape_map[node.name].as_proto()])
self.tf_graph.build()
@staticmethod
def _get_scopes(layer_name):
return layer_name.split('/')
def check_const(self, node):
while node:
if node.type == "Const":
return node
elif node.type in self.skip_type:
node = self.get_parent(node.name, [0])
else:
print(node.layer)
assert False
def _convert_reduction_operators(self, source_node, new_op = None):
IR_node = self._convert_identity_operation(source_node, 1, new_op)
# keep dims
IR_node.attr['keepdims'].b = source_node.layer.attr['keep_dims'].b
# axes
axes = self.get_parent(source_node.name, [1]).layer.attr['value'].tensor
axes = tensor_util.MakeNdarray(axes)
IR_node.attr['axes'].list.i.extend(axes)
def _convert_layers_batchnorm(self, source_node):
# name, op
IR_node = self.IR_graph.node.add()
TensorflowParser2._copy_and_reop(source_node, IR_node, 'BatchNorm')
# epsilon
epsilon = self.get_parent(source_node.name, [1])
IR_node.attr['epsilon'].f = epsilon.layer.attr['value'].tensor.float_val[0]
# moving variance (var) /read
moving_variance = self.get_parent(source_node.name, [0])
if moving_variance.type == 'Identity':
moving_variance_read = self.src_graph.get_parent(moving_variance.name, [0])
tensor_content = moving_variance_read.get_attr('value')
moving_variance_content = tensor_util.MakeNdarray(tensor_content)
self.set_weight(source_node.name, 'var', moving_variance_content)
else:
print(moving_variance.layer)
assert False
# gamma (scale)
Rsqrt = self.get_son(source_node.name, [0], True)
if len(Rsqrt.out_edges) == 2:
IR_node.attr['scale'].b = False
output_node = self.get_son(source_node.name, [0, 0, 0], True)
Mul = self.get_son(source_node.name, [0, 1], True)
else:
IR_node.attr['scale'].b = True
son = self.get_son(source_node.name, [0, 0, 0], True)
gamma_from = self.get_parent(son.name, [1, 1], True)
gamma = self.check_const(gamma_from)
# gamma = self.get_parent(son.name, [1, 1, 0, 0, 0, 1], True)
gamma_tensor = gamma.get_attr('value')
scale = tensor_util.MakeNdarray(gamma_tensor)
self.set_weight(source_node.name, 'scale', scale)
output_node = self.get_son(source_node.name, [0, 0, 0, 0], True)
# print(output_node.layer)
Mul = self.get_son(source_node.name, [0, 0, 1], True)
# print(Mul.layer)
# beta (bias)
beta = self.get_parent(output_node.name, [1, 0, 0], True).get_attr('value')
bias = tensor_util.MakeNdarray(beta) #(96,)
IR_node.attr['bias'].b = True
self.set_weight(source_node.name, 'bias', bias)
# moving mean (mean)
moving_mean = self.get_parent(Mul.name, [0, 0]).get_attr('value')
mean = tensor_util.MakeNdarray(moving_mean)
self.set_weight(source_node.name, 'mean', mean)
# input node
assert output_node.type == 'Add'
input_node = self.get_parent(output_node.name, [0, 0])
IR_node.input.append(input_node.real_name)
# output node
output_node.real_name = source_node.name
def _convert_layers_instancenorm(self, source_node):
IR_node = self.IR_graph.node.add()
TensorflowParser2._copy_and_reop(source_node, IR_node, 'InstanceNorm')
# epsilon
epsilon = self.get_parent(source_node.name, [1])
epsilon_value = epsilon.get_attr('value').float_val[0]
IR_node.attr['epsilon'].f = epsilon_value
# beta
output_node = self.get_son(source_node.name, [0, 0, 0, 0], True)
beta = self.get_parent(output_node.name, [1, 0, 0, 0, 0, 1], True)
beta_tensor = beta.get_attr('value')
beta = tensor_util.MakeNdarray(beta_tensor)
self.set_weight(source_node.name, 'bias', beta)
# gamma (scale)
IR_node.attr['scale'].b = True
son = self.get_son(source_node.name, [0, 0, 0], True)
gamma = self.get_parent(son.name, [1, 1, 0, 0, 0, 1], True)
gamma_tensor = gamma.get_attr('value')
scale = tensor_util.MakeNdarray(gamma_tensor)
self.set_weight(source_node.name, 'scale', scale)
# output_node = self.get_son(source_node.name, [0, 0, 0, 0], True)
assert output_node.type == 'Add'
input_node = self.get_parent(output_node.name, [0, 0])
IR_node.input.append(input_node.real_name)
output_node.real_name = source_node.name
# assert False
@classmethod
def _skip_node(cls, source_node):
if source_node.covered:
return True
for prefix in cls.skip_prefix:
if source_node.name.startswith(prefix):
return True
scopes = TensorflowParser2._get_scopes(source_node.name)
for s in scopes:
if s in cls.skip_scope:
return True
return False
def gen_IR(self):
for layer in self.src_graph.topological_sort:
current_node = self.src_graph.get_node(layer)
if self._skip_node(current_node):
continue
node_type = current_node.type
if hasattr(self, "rename_" + node_type):
func = getattr(self, "rename_" + node_type)
func(current_node)
else:
self.rename_UNKNOWN(current_node)
@staticmethod
def tensor_shape_to_list(shapes):
if isinstance(shapes, attr_value_pb2.AttrValue):
return [dim.size for dim in shapes.shape.dim]
else:
ret = []
for shape in shapes:
this_one = [dim.size for dim in shape.dim]
ret.append(this_one)
return ret
@staticmethod
def _copy_and_reop(source_node, IR_node, new_op = None):
if new_op == None: new_op = source_node.type
IR_node.name = source_node.name
IR_node.op = new_op
kwargs = {}
if 'data_format' in source_node.layer.attr:
kwargs['data_format'] = source_node.get_attr('data_format')
if 'T' in source_node.layer.attr:
if source_node.type not in TensorflowParser2.q_type:
if source_node.type == 'Enter':
IR_node.attr["dtype"].type = TensorflowParser2.dtype_map[6]
else:
assert source_node.layer.attr['T'].type in TensorflowParser2.dtype_map, 'type [{}] is unknown.'.format(source_node.layer.attr['dtype'].type)
IR_node.attr["dtype"].type = TensorflowParser2.dtype_map[source_node.layer.attr['T'].type]
else:
IR_node.attr["dtype"].type = TensorflowParser2.dtype_map[6]
if '_output_shapes' in source_node.layer.attr:
IR_node.attr["_output_shapes"].MergeFromString(source_node.layer.attr['_output_shapes'].SerializeToString())
if 'paddings' in source_node.layer.attr:
IR_node.attr["paddings"].MergeFromString(source_node.layer.attr['paddings'].SerializeToString())
assign_IRnode_values(IR_node, kwargs)
def _convert_inedge(self, source_node, IR_node, start_idx = 0, end_idx = None):
if end_idx == None: end_idx = len(source_node.in_edges)
for idx in range(start_idx, end_idx):
IR_node.input.append(self.src_graph.get_node(source_node.in_edges[idx]).real_name)
@staticmethod
def _copy_shape(source_node, IR_node):
assert 'shape' in source_node.layer.attr
if source_node.layer.attr['shape'].list.shape:
IR_node.attr['shape'].shape.MergeFromString(source_node.layer.attr['shape'].list.shape[0].SerializeToString())
else:
IR_node.attr['shape'].shape.MergeFromString(source_node.layer.attr['shape'].shape.SerializeToString())
# print(IR_node)
def rename_UNKNOWN(self, source_node):
if source_node.type in self.skip_type:
return
print("Tensorflow has not supported operator [%s] with name [%s]."
% (source_node.type, source_node.name))
assert False
def rename_NoOp(self, source_node):
return
def _convert_padding(self, source_node, IR_node, kernel_size):
# TODO: Fused conv and pool with padding is different from defused operators
input_node = self.get_parent(source_node.name, [0])
input_shape = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
if source_node.get_attr('padding') == 'VALID':
dims = len(input_shape)
assign_IRnode_values(IR_node, {'auto_pad' : "VALID", 'pads' : [0, 0] * dims})
elif source_node.get_attr('padding') == 'SAME':
padding = compute_tf_same_padding(
input_shape,
kernel_size,
source_node.get_attr('strides'))
assign_IRnode_values(IR_node, {'auto_pad' : "SAME_LOWER", 'pads' : padding})
else:
assert False
def _get_bias(self, source_node, IR_node):
if not source_node.out_edges:
return
add_node = self.tf_graph.get_node(source_node.out_edges[0])
if add_node.type != "Add" and add_node.type != "BiasAdd":
return
variable = self.tf_graph.get_node(add_node.in_edges[1]) #add_bias node
if variable.type == 'Identity':
variable = self.tf_graph.get_node(variable.in_edges[0])
bias_value = variable.get_attr('value')
bias = tensor_util.MakeNdarray(bias_value)
# assert variable.get_attr('_output_shapes')[0].dim[0].size == IR_node.attr['kernel_shape'].list.i[-1]
add_node.real_name = IR_node.name
add_node.covered = True
IR_node.attr['use_bias'].b = True
current_layer = self.weights[source_node.name]
current_layer['bias'] = bias
def _convert_pooling(self, source_node, pool_type):
IR_node = self._convert_identity_operation(source_node, new_op='Pool')
kwargs = {}
# strides
kwargs['strides'] = source_node.get_attr('strides')
# window_shape
kwargs['kernel_shape'] = source_node.get_attr('ksize')
# pool type
kwargs['pooling_type'] = pool_type
# padding
self._convert_padding(source_node, IR_node, kwargs['kernel_shape'][1:-1])
assign_IRnode_values(IR_node, kwargs)
def _convert_identity_operation(self, source_node, start_idx = 0, end_idx = None, new_op = None):
IR_node = self.IR_graph.node.add()
TensorflowParser2._copy_and_reop(source_node, IR_node, new_op)
self._convert_inedge(source_node, IR_node, start_idx, end_idx)
return IR_node
def rename_Relu6(self, source_node):
self._convert_identity_operation(source_node, new_op = 'Relu6')
def rename_DepthwiseConv2dNative(self, source_node):
IR_node = self._convert_identity_operation(source_node, end_idx = 1, new_op = 'DepthwiseConv')
kwargs = {}
kwargs['strides'] = source_node.get_attr('strides')
input_node = self.src_graph.get_parent(source_node.name, [1])
kwargs['kernel_shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
self._convert_padding(source_node, IR_node, kwargs['kernel_shape'][:-2])
weight = self.src_graph.get_parent(source_node.name, [1, 0]).get_attr('value')
weight_content = tensor_util.MakeNdarray(weight)
self.set_weight(source_node.name, 'weights', weight_content)
assign_IRnode_values(IR_node, kwargs)
def rename_BatchNormWithGlobalNormalization(self, source_node):
IR_node = self._convert_identity_operation(source_node, start_idx=0, end_idx=1, new_op='BatchNorm')
# epsilon
IR_node.attr['epsilon'].f = source_node.get_attr('variance_epsilon')
# moving variance (var) /read
moving_variance = self.get_parent(source_node.name, [2])
tensor_variance = moving_variance.get_attr('value')
moving_variance_content = tensor_util.MakeNdarray(tensor_variance)
self.set_weight(source_node.name, 'var', moving_variance_content)
# gamma (scale)
gamma = self.get_parent(source_node.name, [4])
gamma_value = gamma.get_attr('value')
gamma = tensor_util.MakeNdarray(gamma_value)
self.set_weight(source_node.name, 'scale', gamma)
# beta (bias)
beta = self.get_parent(source_node.name, [3])
beta_value = beta.get_attr('value')
beta = tensor_util.MakeNdarray(beta_value)
self.set_weight(source_node.name, 'bias', beta)
# moving mean (mean)
mean = self.get_parent(source_node.name, [1])
mean_value = mean.get_attr('value')
mean = tensor_util.MakeNdarray(mean_value)
self.set_weight(source_node.name, 'mean', mean)
def rename_Placeholder(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op='DataInput')
TensorflowParser2._copy_shape(source_node, IR_node)
print(IR_node)
IR_node.attr['shape'].shape.dim[0].size = -1
IR_node.attr['_output_shapes'].list.shape[0].dim[0].size = -1
def rename_CheckNumerics(self, source_node):
return
def rename_Reshape(self, source_node):
IR_node = self._convert_identity_operation(source_node, end_idx = 1)
kwargs = {'shape' : self.tensor_shape_to_list(source_node.get_attr('_output_shapes'))[0]}
assign_IRnode_values(IR_node, kwargs)
def rename_MirrorPad(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'MirrorPad')
input_node = self.src_graph.get_parent(source_node.name, [1])
tensor_content = tensor_util.MakeNdarray(input_node.get_attr('value')).reshape(-1)
# print(tensor_content)
kwargs = {}
kwargs['mode'] = source_node.get_attr('mode')
kwargs['pads'] = tensor_content.tolist()
assign_IRnode_values(IR_node, kwargs)
# print(IR_node)
def rename_Min(self, source_node):
# print(source_node.layer)
IR_node = self._convert_identity_operation(source_node, start_idx=0, end_idx=1, new_op = 'Min')
kwargs = {}
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs['shape_0'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
input_node = self.src_graph.get_parent(source_node.name, [1])
kwargs['shape_1'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
# assert False
def rename_Max(self, source_node):
# print(source_node.layer)
IR_node = self._convert_identity_operation(source_node, start_idx=0, end_idx=1, new_op = 'Max')
kwargs = {}
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs['shape_0'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
input_node = self.src_graph.get_parent(source_node.name, [1])
kwargs['shape_1'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
# assert False
def rename_Mul(self, source_node):
input_node_0 = self.src_graph.get_parent(source_node.name, [0])
# mean/read
if input_node_0.type == 'Identity':
input_node_0_read = self.src_graph.get_parent(input_node_0.name, [0])
tensor_content = input_node_0_read.get_attr('value')
tensor_content = tensor_util.MakeNdarray(tensor_content)
self.set_weight(source_node.name, 'weights', tensor_content)
IR_node = self._convert_identity_operation(source_node, start_idx = 1)
else:
IR_node = self._convert_identity_operation(source_node)
def rename_Add(self, source_node):
scopes = self._get_scopes(source_node.name)
# print(scopes)
if scopes[-2] == 'batchnorm':
if scopes[-3] == 'BatchNorm':
self._convert_layers_batchnorm(source_node)
elif scopes[-3] == 'InstanceNorm':
self._convert_layers_instancenorm(source_node)
else:
print(scopes)
IR_node = self._convert_identity_operation(source_node, new_op = "Add")
def rename_Sub(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = "Sub")
def rename_Sum(self, source_node):
IR_node = self._convert_identity_operation(source_node, start_idx=0, end_idx=1, new_op = 'Sum')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs = {}
kwargs['cal_shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
input_node_indices = self.src_graph.get_parent(source_node.name, [1])
indice_value = input_node_indices.get_attr('value')
if indice_value.tensor_content:
shapes = tensor_util.MakeNdarray(indice_value)
c = shapes.tolist()
kwargs['sum_indices'] = c
else:
kwargs['sum_indices'] = input_node_indices.get_attr('value').int_val[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Rsqrt(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = "Rsqrt")
kwargs = {}
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Square(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'Square')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs = {}
kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Sigmoid(self, source_node):
IR_node = self._convert_identity_operation(source_node)
kwargs = {}
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Reciprocal(self, source_node):
IR_node = self._convert_identity_operation(source_node, start_idx=0, end_idx=1, new_op = 'Reciprocal')
def rename_Minimum(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'Minimum')
def rename_Maximum(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'Maximum')
def rename_RealDiv(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'RealDiv')
def rename_Enter(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'Enter')
def rename_Switch(self, source_node):
source_node.real_name = self.src_graph.get_node(source_node.in_edges[0]).real_name
def rename_Exp(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'Exp')
def rename_ResizeBilinear(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'ResizeBilinear')
def rename_Cast(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'Cast')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs = {}
kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Prod(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'Prod')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs = {}
kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
input_node_const = self.src_graph.get_parent(source_node.name, [1])
kwargs['const'] = input_node_const.get_attr('value').int_val[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Shape(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'Shape')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs = {}
kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Squeeze(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'Squeeze')
def rename_Pack(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'Pack')
def rename_Gather(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'Gather')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs = {}
# kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
input_node_indices = self.src_graph.get_parent(source_node.name, [1])
indice_value = input_node_indices.get_attr('value')
shapes = tensor_util.MakeNdarray(indice_value)
c = shapes.tolist()
kwargs['gather_indices'] = c
assign_IRnode_values(IR_node, kwargs)
def rename_StridedSlice(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'StridedSlice')
kwargs = {}
input_node = self.src_graph.get_parent(source_node.name, [0])
# kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
input_node_const0 = self.src_graph.get_parent(source_node.name, [1])
input_node_const1 = self.src_graph.get_parent(source_node.name, [2])
input_node_const2 = self.src_graph.get_parent(source_node.name, [3])
if input_node_const0.get_attr('value').int_val:
kwargs['const0'] = input_node_const0.get_attr('value').int_val[0]
if input_node_const1.get_attr('value').int_val:
kwargs['const1'] = input_node_const1.get_attr('value').int_val[0]
if input_node_const2.get_attr('value').int_val:
kwargs['const2'] = input_node_const2.get_attr('value').int_val[0]
assign_IRnode_values(IR_node, kwargs)
def rename_ExpandDims(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'ExpandDims')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs = {}
kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
input_node_indices = self.src_graph.get_parent(source_node.name, [1])
kwargs['exp_dim'] = input_node_indices.get_attr('value').int_val[0]
assign_IRnode_values(IR_node, kwargs)
def rename_ResizeNearestNeighbor(self, source_node):
IR_node = self._convert_identity_operation(source_node)
kwargs = {}
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
input_node_size = self.src_graph.get_parent(source_node.name, [1])
kwargs['size'] = self.tensor_shape_to_list(input_node_size.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Conv2D(self, source_node):
IR_node = self._convert_identity_operation(source_node, end_idx=1, new_op = 'Conv')
kwargs = {}
kwargs['strides'] = source_node.get_attr('strides')
kwargs['padding'] = source_node.get_attr('padding')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
# weights
input_node_weight = self.src_graph.get_parent(source_node.name, [1])
if input_node_weight.type == 'Const':
tensor_content = input_node_weight.get_attr('value')
else:
input_node_weight_read = self.src_graph.get_parent(input_node_weight.name, [0])
tensor_content = input_node_weight_read.get_attr('value')
W = tensor_util.MakeNdarray(tensor_content)
kwargs['kernel_shape'] = self.tensor_shape_to_list(input_node_weight.get_attr('_output_shapes'))[0]
self.set_weight(source_node.name, 'weights', W)
self._convert_padding(source_node, IR_node, kwargs['kernel_shape'][:-2])
assign_IRnode_values(IR_node, kwargs)
self._get_bias(source_node, IR_node)
def rename_Relu(self, source_node):
IR_node = self._convert_identity_operation(source_node)
kwargs = {'shape' : self.tensor_shape_to_list(source_node.get_attr('_output_shapes'))[0]}
assign_IRnode_values(IR_node, kwargs)
def rename_MaxPool(self, source_node):
self._convert_pooling(source_node, b'MAX')
def rename_AvgPool(self, source_node):
self._convert_pooling(source_node, b'AVG')
def rename_LRN(self, source_node):
IR_node = self._convert_identity_operation(source_node)
# alpha
IR_node.attr["alpha"].f = float(source_node.get_attr("alpha", "0.0001"))
# beta
IR_node.attr["beta"].f = float(source_node.get_attr("beta", "0.75"))
IR_node.attr["size"].i = source_node.get_attr("depth_radius")
IR_node.attr["bias"].f = float(source_node.get_attr("bias"))
def rename_Concat(self, source_node):
n = len(source_node.in_edges)
IR_node = self._convert_identity_operation(source_node, start_idx=1, end_idx=n, new_op='Concat')
axis = self.tf_graph.get_parent(source_node.name, [0])
IR_node.attr["axis"].i = axis.get_attr('value').int_val[0]
def rename_ConcatV2(self, source_node):
n = len(source_node.in_edges)
IR_node = self._convert_identity_operation(source_node, start_idx=0, end_idx=n-1, new_op='Concat')
axis = self.tf_graph.get_parent(source_node.name, [n-1])
IR_node.attr["axis"].i = axis.get_attr('value').int_val[0]
def rename_MatMul(self, source_node):
IR_node = self._convert_identity_operation(source_node, end_idx=1)
input_weight_node = self.src_graph.get_parent(source_node.name, [1])
weightnode = self.check_const(input_weight_node)
weight_value = weightnode.get_attr('value')
weight = tensor_util.MakeNdarray(weight_value)
self.set_weight(source_node.name, 'weights', weight)
units = source_node.layer.attr['_output_shapes'].list.shape[-1].dim[-1].size
IR_node.attr['units'].i = units
if source_node.out_edges and self.tf_graph.get_node(source_node.out_edges[0]).type == 'BiasAdd':
add_node = self.tf_graph.get_node(source_node.out_edges[0])
add_node.covered = True
add_node.real_name = source_node.real_name
TensorflowParser2._copy_and_reop(source_node, IR_node, 'FullyConnected')
variable = self.tf_graph.get_node(add_node.in_edges[1]) #add_bias node
biasnode = self.check_const(variable)
bias_value = biasnode.get_attr('value')
bias = tensor_util.MakeNdarray(bias_value)
self.set_weight(source_node.name, 'bias', bias)
IR_node.attr['use_bias'].b = True
def rename_Softmax(self, source_node):
IR_node = self._convert_identity_operation(source_node)
kwargs = {'shape' : self.tensor_shape_to_list(source_node.get_attr('_output_shapes'))[0]}
IR_node.attr["dim"].i = 1
assign_IRnode_values(IR_node, kwargs)
def rename_Identity(self, source_node):
source_node.real_name = self.src_graph.get_node(source_node.in_edges[0]).real_name
def rename_QuantizeV2(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'QuantizeV2')
TensorflowParser2._copy_shape(source_node, IR_node)
def rename_QuantizedRelu(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = "QuantizedRelu")
kwargs = {'shape' : self.tensor_shape_to_list(source_node.get_attr('_output_shapes'))[0]}
assign_IRnode_values(IR_node, kwargs)
def rename_QuantizedReshape(self, source_node):
IR_node = self._convert_identity_operation(source_node, end_idx = 1)
kwargs = {'shape' : self.tensor_shape_to_list(source_node.get_attr('_output_shapes'))[0]}
assign_IRnode_values(IR_node, kwargs)
def rename_QuantizedConv2D(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'QConv')
kwargs = {}
kwargs['strides'] = source_node.get_attr('strides')
kwargs['padding'] = source_node.get_attr('padding')
# weights
input_node = self.src_graph.get_parent(source_node.name, [1])
tensor_content = input_node.get_attr('value')
W = tensor_util.MakeNdarray(tensor_content)
W = W.astype(np.uint8)
kwargs['kernel_shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
input_node_minw = self.src_graph.get_parent(source_node.name, [4])
min_W = input_node_minw.get_attr('value').float_val[0]
input_node_maxw = self.src_graph.get_parent(source_node.name, [5])
max_W = input_node_maxw.get_attr('value').float_val[0]
if source_node.get_attr('Tfilter') == tensorflow.quint8:
W = ((max_W - min_W)/255.0) * W + min_W
else:
assert False, ('Only uint8 weights handled currently by the converter')
self.set_weight(source_node.name, 'kernel_weights', W)
assign_IRnode_values(IR_node, kwargs)
def rename_Dequantize(self, source_node):
IR_node = self._convert_identity_operation(source_node,start_idx=0, end_idx= 1, new_op = 'Dequantize')
kwargs = {}
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Requantize(self, source_node):
print(source_node.layer)
input_node = self.get_parent(source_node.name, [0])
son_node = self.get_son(source_node.name, [0])
son_node.real_name = source_node.name
def rename_RequantizationRange(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op = 'RequantizationRange')
TensorflowParser2._copy_shape(source_node, IR_node)
class TensorflowParser(Parser):
skip_prefix = [
"^",
"train_op",
"save",
"gradients",
"global_step",
"distort_image",
"Adagrad",
]
skip_scope = [
"random_uniform", "Initializer", "optimizer", "weight_loss",
"parallel_read", "case"
]
skip_type = set([
"L2Loss", "VariableV2", "Const", "Assign", "RandomUniform",
"FIFOQueueV2"
])
dtype_map = {
0: graph_pb2.DT_UNDEFINED,
1: graph_pb2.DT_FLOAT32,
2: graph_pb2.DT_FLOAT64,
3: graph_pb2.DT_INT32,
4: graph_pb2.DT_UINT8,
5: graph_pb2.DT_INT16,
6: graph_pb2.DT_INT8,
7: graph_pb2.DT_STRING,
9: graph_pb2.DT_INT64,
10: graph_pb2.DT_BOOL,
19: graph_pb2.DT_FLOAT16
}
@property
def src_graph(self):
return self.tf_graph
@staticmethod
def _shapeToStr(shapes):
return [dim.size if dim.size > 0 else 1 for dim in shapes.dim]
@staticmethod
def _load_meta(model_network_path):
"""Load a tensorflow meta file from disk
Parameters
----------
model_network_path: str
Path where the model network path is (protobuf meta file)
Returns
-------
model: A tensorflow protobuf file
"""
from tensorflow.core.protobuf import meta_graph_pb2
from mmdnn.conversion.common.IR.IR_graph import load_protobuf_from_file
meta_graph = meta_graph_pb2.MetaGraphDef()
load_protobuf_from_file(meta_graph, model_network_path)
graph = meta_graph.graph_def
print("Tensorflow model file [%s] loaded successfully." %
model_network_path)
return graph
@staticmethod
def _load_weights(model_weight_path):
"""Load a tensorflow checkpoint file from disk
Parameters
----------
model_weight_path: str
Path where the weight path is (checkpoint file)
Returns
-------
model: tensor name --> ndarry
"""
reader = tensorflow.train.NewCheckpointReader(model_weight_path)
var_to_shape_map = reader.get_variable_to_shape_map()
data = dict()
for name in var_to_shape_map:
tensor = reader.get_tensor(name)
name_seg = name.split("/")
if name_seg[-1] == "ExponentialMovingAverage":
name = "/".join(name_seg[:-1])
data[name] = tensor
print(
"Tensorflow checkpoint file [%s] loaded successfully. [%d] variables loaded."
% (model_weight_path, len(data)))
return data
@staticmethod
def _get_scopes(layer_name):
return layer_name.split('/')
def check_const(self, node):
while node:
if node.type == "Const":
return node
elif node.type == "NoOp":
return None
else:
node = self.get_parent(node.name, [0])
def _add_constant_node(self, source_node):
parent_ids = range(len(source_node.in_edges))
for idx in parent_ids:
parent_node = self.tf_graph.get_node(source_node.in_edges[idx])
if parent_node.type == 'Const':
self._rename_Const(parent_node)
def _rename_Const(self, source_node):
IR_node = self._convert_identity_operation(
source_node, in_edge_count=0, new_op='Constant') # Constant
value = source_node.get_attr('value')
if value.float_val:
shape = tuple(self.tensor_shape_to_list(value.tensor_shape))
value = np.full(shape, value.float_val[0])
elif value.int_val:
shape = tuple(self.tensor_shape_to_list(value.tensor_shape))
value = np.full(shape, value.int_val[0])
else:
value = np.array(tensor_util.MakeNdarray(value).tolist())
if value.ndim > 1:
self.set_weight(source_node.name, 'value', value)
else:
kwargs = {'value': value}
assign_IRnode_values(IR_node, kwargs)
def _convert_reduction_operators(self, source_node, new_op=None):
IR_node = self._convert_identity_operation(source_node, 0, 1, new_op)
# keep dims
IR_node.attr['keepdims'].b = source_node.layer.attr['keep_dims'].b
# axes
axes = self.get_parent(source_node.name,
[1]).layer.attr['value'].tensor
axes = tensor_util.MakeNdarray(axes)
IR_node.attr['axes'].list.i.extend(axes)
def _convert_layers_batchnorm(self, source_node):
# name, op
IR_node = self.IR_graph.node.add()
TensorflowParser._copy_and_reop(source_node, IR_node, 'BatchNorm')
# epsilon
epsilon = self.get_parent(source_node.name, [1])
IR_node.attr['epsilon'].f = epsilon.layer.attr[
'value'].tensor.float_val[0]
# moving variance (var)
moving_variance = self.get_parent(source_node.name, [0, 0])
# print(moving_variance.name)
if self.weight_loaded and moving_variance.name in self.ckpt_data.keys(
):
self.set_weight(source_node.name, 'var',
self.ckpt_data[moving_variance.name])
# gamma (scale)
gamma = self.get_son(source_node.name, [0, 0], True)
gamma = self.get_parent(gamma.name, [1, 0], True)
if gamma is None or not gamma.type.startswith('Variable'):
IR_node.attr['scale'].b = False
output_node = self.get_son(source_node.name, [0, 0, 0], True)
else:
IR_node.attr['scale'].b = True
if self.weight_loaded:
self.set_weight(source_node.name, 'scale',
self.ckpt_data[gamma.name])
output_node = self.get_son(source_node.name, [0, 0, 0, 0], True)
if output_node.type == 'Sub':
output_node = self.get_son(source_node.name, [0, 0, 1, 0],
True)
# mean
mean = self.get_parent(output_node.name, [1, 1, 0, 0], True)
if self.weight_loaded and mean.name in self.ckpt_data.keys():
self.set_weight(source_node.name, 'mean',
self.ckpt_data[mean.name])
# bias
bias = self.get_parent(output_node.name, [1, 0, 0], True)
if bias is None or not bias.type.startswith('Variable'):
IR_node.attr['bias'].b = False
else:
IR_node.attr['bias'].b = True
if self.weight_loaded:
self.set_weight(source_node.name, 'bias',
self.ckpt_data[bias.name])
# input node
assert output_node.type == 'Add'
input_node = self.get_parent(output_node.name, [0, 0])
IR_node.input.append(input_node.real_name)
# output node
output_node.real_name = source_node.name
def __init__(self,
meta_file,
checkpoint_file,
dest_nodes,
inputShape=None,
in_nodes=None):
super(TensorflowParser, self).__init__()
# load model files into TensorFlow graph
if meta_file:
model = TensorflowParser._load_meta(meta_file)
if checkpoint_file:
self.ckpt_data = TensorflowParser._load_weights(checkpoint_file)
self.weight_loaded = True
# extract subgraph using in_nodes and dest_nodes
if in_nodes != None and inputShape != None:
from tensorflow.python.tools import strip_unused_lib
from tensorflow.python.framework import dtypes
from tensorflow.python.platform import gfile
model = strip_unused_lib.strip_unused(
input_graph_def=model,
input_node_names=in_nodes,
output_node_names=dest_nodes,
placeholder_type_enum=dtypes.float32.as_datatype_enum)
input_list = [None]
for i in range(len(inputShape)):
input_list.append(tensorflow.Dimension(inputShape[i]))
tensor_input = tensorflow.TensorShape(input_list)
# Build network graph
self.tf_graph = TensorflowGraph(model)
for node in self.tf_graph.model.node:
if node.name in in_nodes:
node.attr['shape'].shape.CopyFrom(tensor_input.as_proto())
node.attr['_output_shapes'].list.shape.pop(
) #unknown_rank pop
node.attr['_output_shapes'].list.shape.extend(
[tensor_input.as_proto()])
# extract subgraph using dest_nodes
elif dest_nodes != None:
from tensorflow.python.framework.graph_util import extract_sub_graph
model = extract_sub_graph(model, dest_nodes)
# Get input node name
if not in_nodes:
in_nodes = []
for node in model.node:
if node.op == 'Placeholder':
in_nodes.append(node.name)
# Graph Transform
transforms = ["fold_constants(ignore_errors=true)"]
transformed_graph_def = TransformGraph(model, in_nodes, dest_nodes,
transforms)
in_type_list = {}
in_shape_list = {}
for n in transformed_graph_def.node:
if n.name in in_nodes:
in_type_list[n.name] = n.attr['dtype'].type
in_node_shape = n.attr['shape'].shape
in_node_shape_str = self._shapeToStr(in_node_shape)
in_shape_list[n.name] = in_node_shape_str
dtype = tensorflow.float32
with tensorflow.Graph().as_default() as g:
input_map = {}
for in_node in in_nodes:
if in_type_list[in_node] == 1 or in_type_list[in_node] == 0:
dtype = tensorflow.float32
elif in_type_list[in_node] == 3:
dtype = tensorflow.int32
elif in_type_list[in_node] == 10:
dtype = tensorflow.bool
x = tensorflow.placeholder(dtype, shape=in_shape_list[in_node])
input_map[in_node] = x
tensorflow.import_graph_def(transformed_graph_def,
name='',
input_map=input_map)
with tensorflow.Session(graph=g) as sess:
tempdir = tempfile.mkdtemp()
meta_graph_def = tensorflow.train.export_meta_graph(
filename=os.path.join(tempdir, 'my-model.meta'))
model = meta_graph_def.graph_def
shutil.rmtree(tempdir)
self.tf_graph = TensorflowGraph(model)
self.tf_graph.build()
process_graph(self.tf_graph, self.ckpt_data)
@classmethod
def _skip_node(cls, source_node):
if source_node.covered:
return True
for prefix in cls.skip_prefix:
if source_node.name.startswith(prefix):
return True
scopes = TensorflowParser._get_scopes(source_node.name)
for s in scopes:
if s in cls.skip_scope:
return True
return False
@staticmethod
def tensor_shape_to_list(shapes):
if isinstance(shapes, attr_value_pb2.AttrValue):
return [dim.size for dim in shapes.shape.dim]
elif isinstance(
shapes, attr_value_pb2.
tensorflow_dot_core_dot_framework_dot_tensor__shape__pb2.
TensorShapeProto):
return [dim.size for dim in shapes.dim]
else:
ret = []
for shape in shapes:
this_one = [dim.size for dim in shape.dim]
ret.append(this_one)
return ret
'''
check current source_node wether has input weights. If it has, set the weights into weight dict and remove the input edge.
return edges' index which do not include edge connecting weights
'''
def _check_weights(self, source_node, start_edge_id=0, in_edge_count=None):
if in_edge_count == None:
in_edge_count = len(source_node.in_edges) - start_edge_id
valid_pre_ids = []
for pre_idx in range(start_edge_id, start_edge_id + in_edge_count):
pre_node = self.get_parent(source_node.name, [pre_idx])
if pre_node.type == 'Identity' and pre_node.name.split(
'/')[-1] == 'read':
weight_node = self.get_parent(pre_node.name, [0])
assert 'Variable' in weight_node.type
self.set_weight(source_node.name, 'weights',
self.ckpt_data[weight_node.name])
source_node.feed_weights = True
else:
valid_pre_ids.append(pre_idx)
return valid_pre_ids
def _convert_padding(self, source_node, IR_node, kernel_size):
# TODO: Fused conv and pool with padding is different from defused operators
input_node = self.get_parent(source_node.name, [0])
input_shape = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
if source_node.get_attr('padding') == 'VALID':
dims = len(input_shape)
assign_IRnode_values(IR_node, {
'auto_pad': "VALID",
'pads': [0, 0] * dims
})
elif source_node.get_attr('padding') == 'SAME':
padding = compute_tf_same_padding(input_shape, kernel_size,
source_node.get_attr('strides'))
assign_IRnode_values(IR_node, {
'auto_pad': "SAME_UPPER",
'pads': padding
})
else:
assert False
def _convert_pooling(self, source_node, pool_type):
IR_node = self._convert_identity_operation(source_node, new_op='Pool')
kwargs = {}
# strides
kwargs['strides'] = source_node.get_attr('strides')
# window_shape
kwargs['kernel_shape'] = source_node.get_attr('ksize')
# pool type
kwargs['pooling_type'] = pool_type
# padding
self._convert_padding(source_node, IR_node,
kwargs['kernel_shape'][1:-1])
assign_IRnode_values(IR_node, kwargs)
def gen_IR(self):
for layer in self.src_graph.topological_sort:
current_node = self.src_graph.get_node(layer)
if self._skip_node(current_node):
continue
node_type = current_node.type
if hasattr(self, "rename_" + node_type):
func = getattr(self, "rename_" + node_type)
func(current_node)
else:
self.rename_UNKNOWN(current_node)
@staticmethod
def _copy_and_reop(source_node, IR_node, new_op=None):
if new_op == None: new_op = source_node.type
IR_node.name = source_node.name
IR_node.op = new_op
kwargs = {}
if 'data_format' in source_node.layer.attr:
kwargs["data_format"] = source_node.get_attr('data_format')
if 'dtype' in source_node.layer.attr:
assert source_node.layer.attr[
'dtype'].type in TensorflowParser.dtype_map, 'type [{}] is unknown.'.format(
source_node.layer.attr['dtype'].type)
IR_node.attr["dtype"].type = TensorflowParser.dtype_map[
source_node.layer.attr['dtype'].type]
if '_output_shapes' in source_node.layer.attr:
IR_node.attr["_output_shapes"].MergeFromString(
source_node.layer.attr['_output_shapes'].SerializeToString())
if hasattr(source_node, 'feed_weights'):
kwargs["feed_weights"] = True
if hasattr(source_node, 'kwargs'):
kwargs.update(source_node.kwargs)
kwargs['scope'] = source_node.scope
assign_IRnode_values(IR_node, kwargs)
def _convert_inedge(self,
source_node,
IR_node,
start_idx=0,
end_idx=None,
in_ids=None):
if end_idx == None: end_idx = len(source_node.in_edges) - start_idx
if not in_ids:
in_ids = range(start_idx, end_idx + start_idx)
for idx in in_ids:
if ':' in source_node.in_edges[idx]:
input_tensor = self.src_graph.get_node(
source_node.in_edges[idx]
).real_name + ':' + source_node.in_edges[idx].split(':')[1]
else:
input_tensor = self.src_graph.get_node(
source_node.in_edges[idx]).real_name
IR_node.input.append(input_tensor)
def _get_bias(self, source_node, IR_node):
if not source_node.out_edges:
return
add_node = self.tf_graph.get_node(source_node.out_edges[0])
if add_node.type != "Add" and add_node.type != "BiasAdd":
return
variable = self.tf_graph.get_node(add_node.in_edges[1])
if variable.type != "Identity":
return
variable = self.tf_graph.get_node(variable.in_edges[0])
assert variable.layer.attr['shape'].shape.dim[0].size == IR_node.attr[
'kernel_shape'].list.i[-1]
if self.weight_loaded:
assert variable.name in self.ckpt_data
current_layer = self.weights[source_node.name]
current_layer['bias'] = self.ckpt_data[variable.name]
add_node.real_name = IR_node.name
add_node.covered = True
IR_node.attr['use_bias'].b = True
@staticmethod
def _copy_shape(source_node, IR_node):
assert 'shape' in source_node.layer.attr
if source_node.layer.attr['shape'].list.shape:
IR_node.attr['shape'].shape.MergeFromString(
source_node.layer.attr['shape'].list.shape[0].
SerializeToString())
else:
IR_node.attr['shape'].shape.MergeFromString(
source_node.layer.attr['shape'].shape.SerializeToString())
def rename_UNKNOWN(self, source_node):
if source_node.type in self.skip_type:
return
print(
"TensorflowEmitter has not supported operator [%s] with name [%s]."
% (source_node.type, source_node.name))
return
def rename_Placeholder(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='DataInput')
# shape
TensorflowParser._copy_shape(source_node, IR_node)
if len(IR_node.attr['shape'].shape.dim) > 0 and len(
IR_node.attr['_output_shapes'].list.shape) > 0 and len(
IR_node.attr['_output_shapes'].list.shape[0].dim) > 0:
IR_node.attr['shape'].shape.dim[0].size = -1
IR_node.attr['_output_shapes'].list.shape[0].dim[0].size = -1
def rename_Conv2D(self, source_node):
"""
weights: name_weights, name_bias
"""
IR_node = self._convert_identity_operation(source_node, 0, 1, 'Conv')
kwargs = {}
# strides
kwargs['strides'] = source_node.get_attr('strides')
# input[1] : W
# filter
W = self.tf_graph.get_node(source_node.layer.input[1])
if W.type == 'Const':
kwargs['kernel_shape'] = tensor_shape = self.tensor_shape_to_list(
W.layer.attr['value'].tensor.tensor_shape)
else:
W = self.tf_graph.get_node(W.layer.input[0]).layer
kwargs['kernel_shape'] = self.tensor_shape_to_list(W.attr['shape'])
# padding
self._convert_padding(source_node, IR_node,
kwargs['kernel_shape'][:-2])
if self.weight_loaded:
self.set_weight(source_node.name, 'weights',
self.ckpt_data[W.name])
assign_IRnode_values(IR_node, kwargs)
# output[0] : B
self._get_bias(source_node, IR_node)
def _convert_identity_operation(self,
source_node,
start_edge_id=0,
in_edge_count=None,
new_op=None):
IR_node = self.IR_graph.node.add()
in_ids = self._check_weights(source_node, start_edge_id, in_edge_count)
TensorflowParser._copy_and_reop(source_node, IR_node, new_op)
self._convert_inedge(source_node, IR_node, start_edge_id,
in_edge_count, in_ids)
return IR_node
def rename_Relu(self, source_node):
self._convert_identity_operation(source_node)
def rename_Softmax(self, source_node):
self._convert_identity_operation(source_node)
def rename_Relu6(self, source_node):
self._convert_identity_operation(source_node)
def rename_Add(self, source_node):
if not source_node.covered:
scopes = self._get_scopes(source_node.name)
if len(scopes) < 3:
self._convert_identity_operation(source_node, new_op='Add')
elif scopes[-2] == 'dropout':
# converted [dropout]
pass
elif scopes[-2] == 'batchnorm':
# convert [tf.contrib.layers.batch_norm]
self._convert_layers_batchnorm(source_node)
else:
# normal Add
self._add_constant_node(source_node)
self._convert_identity_operation(source_node, new_op='Add')
def rename_AddV2(self, source_node):
self.rename_Add(source_node)
def rename_Sub(self, source_node):
self._add_constant_node(source_node)
self._convert_identity_operation(source_node)
def rename_Reshape(self, source_node):
IR_node = self._convert_identity_operation(source_node,
in_edge_count=1)
kwargs = {
'shape':
self.tensor_shape_to_list(
source_node.get_attr('_output_shapes'))[0]
}
assign_IRnode_values(IR_node, kwargs)
def rename_Abs(self, source_node):
IR_node = self._convert_identity_operation(source_node,
in_edge_count=1,
new_op='Abs')
def rename_Square(self, source_node):
IR_node = self._convert_identity_operation(source_node,
in_edge_count=1,
new_op='Square')
def rename_MatMul(self, source_node):
W = self.tf_graph.get_node(
self.tf_graph.get_node(source_node.in_edges[1]).in_edges[0])
if 'Variable' in W.type:
"""
weights: name_weights, name_bias
"""
IR_node = self._convert_identity_operation(source_node,
in_edge_count=1)
# units
units = source_node.layer.attr['_output_shapes'].list.shape[
-1].dim[-1].size
IR_node.attr['units'].i = units
# Weights
W = self.tf_graph.get_node(
self.tf_graph.get_node(source_node.in_edges[1]).in_edges[0])
if self.weight_loaded:
self.set_weight(source_node.name, 'weights',
self.ckpt_data[W.name])
if source_node.out_edges and (
self.tf_graph.get_node(source_node.out_edges[0]).type
== 'Add' or self.tf_graph.get_node(
source_node.out_edges[0]).type == 'BiasAdd'):
add_node = self.tf_graph.get_node(source_node.out_edges[0])
add_node.covered = True
add_node.real_name = source_node.real_name
# FullyConnected Layer
# name, op
TensorflowParser._copy_and_reop(source_node, IR_node,
'FullyConnected')
# get Bias
B = self.tf_graph.get_node(
self.tf_graph.get_node(
source_node.out_edges[0]).in_edges[1]).in_edges[0]
if self.weight_loaded:
self.set_weight(source_node.name, 'bias',
self.ckpt_data[B])
IR_node.attr['use_bias'].b = True
else:
# Matmul Layer
TensorflowParser._copy_and_reop(source_node, IR_node,
'FullyConnected')
assign_IRnode_values(IR_node, {'use_bias': False})
else:
self._convert_identity_operation(source_node, new_op='MatMul')
def rename_RealDiv(self, source_node):
scopes = self._get_scopes(source_node.name)
# Deal Dropout
if len(scopes) > 1 and scopes[-2][:7] == 'dropout':
IR_node = self._convert_identity_operation(source_node,
in_edge_count=1,
new_op='Dropout')
# keep prob
if 'value' in self.tf_graph.get_node(
source_node.layer.input[1]).layer.attr:
IR_node.attr['keep_prob'].f = self.tf_graph.get_node(
source_node.layer.input[1]
).layer.attr['value'].tensor.float_val[0]
else:
IR_node.attr['keep_prob'].f = 1.0
# Remove nodes
# Mul
mul_node = self.tf_graph.get_node(source_node.out_edges[0])
assert mul_node.type == "Mul"
mul_node.covered = True
mul_node.real_name = source_node.name
# Floor
floor_node = self.tf_graph.get_node(mul_node.in_edges[1])
assert floor_node.type == "Floor"
floor_node.covered = True
else:
# print (source_node)
# print (source_node.layer)
# assert False
self._convert_identity_operation(source_node, new_op='Div')
def rename_Floor(self, source_node):
scopes = self._get_scopes(source_node.name)
assert scopes[-2] == 'dropout'
def rename_MaxPool(self, source_node):
self._convert_pooling(source_node, b'MAX')
def rename_AvgPool(self, source_node):
self._convert_pooling(source_node, b'AVG')
def rename_Identity(self, source_node):
source_node.real_name = self.src_graph.get_node(
source_node.in_edges[0]).real_name
def rename_Squeeze(self, source_node):
IR_node = self._convert_identity_operation(source_node)
IR_node.attr['axes'].MergeFromString(
source_node.layer.attr['squeeze_dims'].SerializeToString())
def rename_QueueDequeueUpToV2(self, source_node):
IR_node = self._convert_identity_operation(source_node,
in_edge_count=0,
new_op='DataInput')
IR_node.attr['shape'].shape.MergeFromString(
source_node.layer.attr['_output_shapes'].list.shape[0].
SerializeToString())
IR_node.attr['shape'].shape.dim[0].size = -1
IR_node.attr['dtype'].type = self.dtype_map[
source_node.layer.attr['component_types'].list.type[0]]
def rename_QueueDequeueManyV2(self, source_node):
IR_node = self._convert_identity_operation(source_node,
in_edge_count=0,
new_op='DataInput')
IR_node.attr['shape'].shape.MergeFromString(
source_node.layer.attr['_output_shapes'].list.shape[0].
SerializeToString())
IR_node.attr['shape'].shape.dim[0].size = -1
IR_node.attr['dtype'].type = self.dtype_map[
source_node.layer.attr['component_types'].list.type[0]]
# def rename_RandomShuffleQueueV2(self, source_node):
# # print(source_node.layer)
# IR_node = self._convert_identity_operation(source_node, in_edge_count = 0, new_op = 'DataInput')
# # IR_node.attr['shape'].shape.MergeFromString(source_node.layer.attr['_output_shapes'].list.shape[0].SerializeToString())
# # IR_node.attr['shape'].shape.dim[0].size = -1
# IR_node.attr['dtype'].type = self.dtype_map[source_node.layer.attr['component_types'].list.type[0]]
def rename_Pad(self, source_node):
IR_node = self._convert_identity_operation(source_node,
in_edge_count=1,
new_op='Pad')
kwargs = {}
kwargs['mode'] = 'constant'
kwargs['constant_values'] = 0.0
# paddings
padding = self.get_parent(source_node.name,
[1]).layer.attr['value'].tensor
shapes = tensor_util.MakeNdarray(padding)
kwargs['pads'] = convert_tf_pad_to_onnx(shapes)
assign_IRnode_values(IR_node, kwargs)
def rename_Mean(self, source_node):
self._convert_reduction_operators(source_node, new_op='ReduceMean')
def rename_ConcatV2(self, source_node):
n = len(source_node.in_edges) - 1
self._add_constant_node(source_node)
IR_node = self._convert_identity_operation(source_node,
in_edge_count=n,
new_op='Concat')
axis = self.tf_graph.get_parent(source_node.name, [n])
IR_node.attr['axis'].i = axis.layer.attr['value'].tensor.int_val[0]
def rename_DepthwiseConv2dNative(self, source_node):
IR_node = self._convert_identity_operation(source_node,
in_edge_count=1,
new_op='DepthwiseConv')
kwargs = {}
kwargs['strides'] = source_node.get_attr('strides')
input_node = self.src_graph.get_parent(source_node.name, [1])
kwargs['kernel_shape'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
self._convert_padding(source_node, IR_node,
kwargs['kernel_shape'][:-2])
if self.weight_loaded:
weight = self.src_graph.get_parent(source_node.name, [1, 0])
self.set_weight(source_node.name, 'weights',
self.ckpt_data[weight.name])
assign_IRnode_values(IR_node, kwargs)
self._get_bias(source_node, IR_node)
def rename_FusedBatchNorm(self, source_node):
IR_node = self._convert_identity_operation(source_node,
in_edge_count=1,
new_op='BatchNorm')
IR_node.attr['epsilon'].f = source_node.get_attr('epsilon', 0)
# gamma (scale)
scale = self.get_parent(source_node.name, [1], True)
if scale.type == 'Const':
value = scale.get_attr('value')
shape = value.tensor_shape
assert len(shape.dim) == 1
shape = shape.dim[0].size
assert len(value.float_val) == 1
value = value.float_val[0]
if np.isclose(value, 1.0):
IR_node.attr['scale'].b = False
else:
IR_node.attr['scale'].b = True
if self.weight_loaded:
self.set_weight(source_node.name, 'scale',
np.array([value] * shape))
else:
scale = self.get_parent(scale.name, [0], True)
if self.weight_loaded:
self.set_weight(source_node.name, 'scale',
self.ckpt_data[scale.name])
IR_node.attr['scale'].b = True
# bias
bias = self.get_parent(source_node.name, [2, 0], True)
IR_node.attr['bias'].b = True
# Mean
mean = self.get_parent(source_node.name, [3, 0], True)
# Var
var = self.get_parent(source_node.name, [4, 0], True)
if self.weight_loaded:
self.set_weight(source_node.name, 'bias',
self.ckpt_data[bias.name])
self.set_weight(source_node.name, 'mean',
self.ckpt_data[mean.name])
self.set_weight(source_node.name, 'var', self.ckpt_data[var.name])
def rename_FusedBatchNormV3(self, source_node):
self.rename_FusedBatchNorm(source_node)
def rename_Shape(self, source_node):
IR_node = self._convert_identity_operation(source_node,
in_edge_count=1,
new_op='Shape')
def rename_Pack(self, source_node):
N = len(source_node.layer.input)
for i in range(N):
this_node = self.get_parent(source_node.name, [i])
if this_node.type == 'Const':
IR_node = self.IR_graph.node.add()
TensorflowParser._copy_and_reop(this_node, IR_node, 'Constant')
kwargs = {
'value': this_node.layer.attr['value'].tensor.int_val[0],
}
assign_IRnode_values(IR_node, kwargs)
IR_node = self._convert_identity_operation(source_node, new_op='Pack')
kwargs = {
'axis': source_node.layer.attr['axis'].i,
'N': source_node.layer.attr['N'].i
}
assign_IRnode_values(IR_node, kwargs)
def rename_Gather(self, source_node):
W = self.src_graph.get_parent(source_node.name, [0])
W = self.src_graph.get_parent(W.name, [0])
if 'Variable' in W.type:
IR_node = self._convert_identity_operation(source_node,
new_op='Embedding')
self.set_weight(source_node.name, "weights",
self.ckpt_data[W.name])
kwargs = {
'input_dim': self.ckpt_data[W.name].shape[0],
'output_dim': self.ckpt_data[W.name].shape[1],
'mask_zero': False
}
kwargs['axis'] = 0 # add default
assign_IRnode_values(IR_node, kwargs)
else:
IR_node = self._convert_identity_operation(source_node,
new_op='Gather')
return IR_node
def rename_GatherV2(self, source_node):
IR_node = self.rename_Gather(source_node)
kwargs = {}
kwargs['axis'] = source_node.layer.attr['axis'].i
assign_IRnode_values(IR_node, kwargs)
def rename_Transpose(self, source_node):
IR_node = self._convert_identity_operation(source_node)
def rename_Sigmoid(self, source_node):
self._convert_identity_operation(source_node)
def rename_Mul(self, source_node):
scale1 = self.get_parent(source_node.name, [1], True)
scale2 = self.get_parent(source_node.name, [0], True)
if scale1.type == 'Const' or scale2.type == 'Const':
self._add_constant_node(source_node)
self._convert_identity_operation(source_node)
elif scale2.type == 'Identity':
scale2 = self.get_parent(scale2.name, [0], True)
assert scale2.type == "VariableV2"
self.set_weight(source_node.name, 'alpha',
self.ckpt_data[scale2.name])
self._convert_identity_operation(source_node)
else:
self._convert_identity_operation(source_node)
'''
tf.unpack has been deprecated with replaced tf.unstack
'''
def rename_Unpack(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='Unstack')
kwargs = {
'axis': source_node.get_attr('axis'),
'num': source_node.get_attr('num')
}
assign_IRnode_values(IR_node, kwargs)
def rename_Split(self, source_node):
if source_node.get_attr('num_split') == 1:
for n in source_node.out_nodes:
for idx, e in enumerate(n.in_edges):
if source_node.name in e:
n.in_edges[idx] = e.split(':')[0]
source_node.real_name = self.get_parent(source_node.name,
[1]).real_name
else:
IR_node = self._convert_identity_operation(source_node, 1, 1)
kwargs = {
'axis':
self.get_parent(source_node.name,
[0]).layer.attr['value'].tensor.int_val[0],
'split':
source_node.get_attr('num_split')
}
assign_IRnode_values(IR_node, kwargs)
def rename_StridedSlice(self, source_node):
# TODO: Current it is only for slice
if self.get_parent(source_node.name, [1]).type != 'Const':
self._add_constant_node(source_node)
IR_node = self._convert_identity_operation(source_node,
new_op='Slice')
return
IR_node = self._convert_identity_operation(source_node,
in_edge_count=1,
new_op='Slice')
kwargs = {
'begin_mask': source_node.get_attr('begin_mask'),
'end_mask': source_node.get_attr('end_mask'),
'shrink_axis_mask': source_node.get_attr('shrink_axis_mask'),
'new_axis_mask': source_node.get_attr('new_axis_mask')
}
starts = self.get_parent(source_node.name,
[1]).layer.attr['value'].tensor
starts = tensor_util.MakeNdarray(starts).tolist()
kwargs['starts'] = starts
ends = self.get_parent(source_node.name,
[2]).layer.attr['value'].tensor
ends = tensor_util.MakeNdarray(ends).tolist()
kwargs['ends'] = ends
if self.get_parent(source_node.name, [3]) != None:
strides = self.get_parent(source_node.name,
[3]).layer.attr['value'].tensor
strides = tensor_util.MakeNdarray(strides).tolist()
kwargs['strides'] = strides
assign_IRnode_values(IR_node, kwargs)
def rename_Slice(self, source_node):
input_node_begin = self.get_parent(source_node.name, [1])
input_node_size = self.get_parent(source_node.name, [2])
begin = tensor_util.MakeNdarray(
input_node_begin.layer.attr['value'].tensor)
size = tensor_util.MakeNdarray(
input_node_size.layer.attr['value'].tensor)
IR_node = self._convert_identity_operation(source_node,
in_edge_count=1,
new_op='Slice')
# TODO: axis
end = size + begin
kwargs = {'starts': begin, 'ends': end}
assign_IRnode_values(IR_node, kwargs)
def rename_LRN(self, source_node):
IR_node = self._convert_identity_operation(source_node)
size = source_node.get_attr('depth_radius') * 2 + 1
alpha = source_node.get_attr('alpha') * size
beta = source_node.get_attr('beta')
bias = source_node.get_attr('bias')
kwargs = {
"alpha": alpha,
"beta": beta,
"bias": bias,
'size': size,
}
assign_IRnode_values(IR_node, kwargs)
def rename_Tanh(self, source_node):
self._convert_identity_operation(source_node)
def rename_ExpandDims(self, source_node):
IR_node = self._convert_identity_operation(source_node,
0,
1,
new_op='Unsqueeze')
ax_node = self.get_parent(source_node.name, [1])
kwargs = {'axes': [ax_node.layer.attr['value'].tensor.int_val[0]]}
assign_IRnode_values(IR_node, kwargs)
def rename_Fill(self, source_node):
IR_node = self._convert_identity_operation(source_node,
0,
1,
new_op='Fill')
value_node = self.get_parent(source_node.name, [1])
if value_node.layer.attr['value'].tensor.float_val:
IR_node.attr['value'].f = value_node.layer.attr[
'value'].tensor.float_val[0]
elif value_node.layer.attr['value'].tensor.int_val:
IR_node.attr['value'].i = value_node.layer.attr[
'value'].tensor.int_val[0]
else:
raise NotImplementedError()
def rename_Conv2DBackpropInput(self, source_node):
"""
weights: name_weights, name_bias
"""
IR_node = self._convert_identity_operation(source_node,
new_op='ConvTranspose')
kwargs = {}
# strides
kwargs['strides'] = source_node.get_attr('strides')
# input[1] : W
# filter
W = self.tf_graph.get_node(source_node.layer.input[1])
W = self.tf_graph.get_node(W.layer.input[0]).layer
kwargs['kernel_shape'] = self.tensor_shape_to_list(W.attr['shape'])
# padding
self._convert_padding(source_node, IR_node,
kwargs['kernel_shape'][:-2])
if self.weight_loaded:
self.set_weight(source_node.name, 'weights',
self.ckpt_data[W.name])
assign_IRnode_values(IR_node, kwargs)
# output[0] : B
self._get_bias(source_node, IR_node)
def rename_Minimum(self, source_node):
self._add_constant_node(source_node)
self._convert_identity_operation(source_node)
def rename_Maximum(self, source_node):
self._add_constant_node(source_node)
self._convert_identity_operation(source_node)
def rename_Cast(self, source_node):
IR_node = self._convert_identity_operation(source_node)
dst = source_node.get_attr('DstT')
if dst == 1:
dst = 'float'
elif dst == 3:
dst = 'int'
else:
raise NotImplementedError
kwargs = {'dstType': dst}
assign_IRnode_values(IR_node, kwargs)
class TensorflowParser(Parser):
skip_prefix = [
"^",
"train_op",
"save",
"gradients",
"init",
"global_step",
"distort_image",
"Adagrad",
]
skip_scope = [
"random_uniform", "Initializer", "optimizer", "weight_loss",
"parallel_read", "case"
]
skip_type = set([
"L2Loss", "VariableV2", "Const", "Assign", "RandomUniform",
"FIFOQueueV2"
])
dtype_map = {
0: graph_pb2.DT_UNDEFINED,
1: graph_pb2.DT_FLOAT32,
2: graph_pb2.DT_FLOAT64,
3: graph_pb2.DT_INT32,
4: graph_pb2.DT_UINT8,
5: graph_pb2.DT_INT16,
6: graph_pb2.DT_INT8,
7: graph_pb2.DT_STRING,
9: graph_pb2.DT_INT64
}
@property
def src_graph(self):
return self.tf_graph
@staticmethod
def _load_meta(model_network_path):
"""Load a tensorflow meta file from disk
Parameters
----------
model_network_path: str
Path where the model network path is (protobuf meta file)
Returns
-------
model: A tensorflow protobuf file
"""
from tensorflow.core.protobuf import meta_graph_pb2
from mmdnn.conversion.common.IR.IR_graph import load_protobuf_from_file
meta_graph = meta_graph_pb2.MetaGraphDef()
load_protobuf_from_file(meta_graph, model_network_path)
graph = meta_graph.graph_def
print("Tensorflow model file [%s] loaded successfully." %
model_network_path)
return graph
@staticmethod
def _load_weights(model_weight_path):
"""Load a tensorflow checkpoint file from disk
Parameters
----------
model_weight_path: str
Path where the weight path is (checkpoint file)
Returns
-------
model: tensor name --> ndarry
"""
reader = tensorflow.train.NewCheckpointReader(model_weight_path)
var_to_shape_map = reader.get_variable_to_shape_map()
data = dict()
for name in var_to_shape_map:
tensor = reader.get_tensor(name)
data[name] = tensor
print(
"Tensorflow checkpoint file [%s] loaded successfully. [%d] variables loaded."
% (model_weight_path, len(data)))
return data
@staticmethod
def _get_scopes(layer_name):
return layer_name.split('/')
def _convert_reduction_operators(self, source_node, new_op=None):
IR_node = self._convert_identity_operation(source_node, 1, new_op)
# keep dims
IR_node.attr['keepdims'].b = source_node.layer.attr['keep_dims'].b
# axes
axes = self.get_parent(source_node.name,
[1]).layer.attr['value'].tensor
axes = tensor_util.MakeNdarray(axes)
IR_node.attr['axes'].list.i.extend(axes)
def _convert_layers_batchnorm(self, source_node):
# name, op
IR_node = self.IR_graph.node.add()
TensorflowParser._copy_and_reop(source_node, IR_node, 'BatchNorm')
# epsilon
epsilon = self.get_parent(source_node.name, [1])
IR_node.attr['epsilon'].f = epsilon.layer.attr[
'value'].tensor.float_val[0]
# moving variance (var)
moving_variance = self.get_parent(source_node.name, [0, 0])
if self.weight_loaded:
self.set_weight(source_node.name, 'var',
self.ckpt_data[moving_variance.name])
# gamma (scale)
gamma = self.get_son(source_node.name, [0, 0], True)
gamma = self.get_parent(gamma.name, [1, 0], True)
if gamma is None or not gamma.type.startswith('Variable'):
IR_node.attr['scale'].b = False
output_node = self.get_son(source_node.name, [0, 0, 0], True)
else:
IR_node.attr['scale'].b = True
if self.weight_loaded:
self.set_weight(source_node.name, 'scale',
self.ckpt_data[gamma.name])
output_node = self.get_son(source_node.name, [0, 0, 0, 0], True)
# mean
mean = self.get_parent(output_node.name, [1, 1, 0, 0], True)
if self.weight_loaded:
self.set_weight(source_node.name, 'mean',
self.ckpt_data[mean.name])
# bias
bias = self.get_parent(output_node.name, [1, 0, 0], True)
if bias is None or not bias.type.startswith('Variable'):
IR_node.attr['bias'].b = False
else:
IR_node.attr['bias'].b = True
if self.weight_loaded:
self.set_weight(source_node.name, 'bias',
self.ckpt_data[bias.name])
# input node
assert output_node.type == 'Add'
input_node = self.get_parent(output_node.name, [0, 0])
IR_node.input.append(input_node.real_name)
# output node
output_node.real_name = source_node.name
def __init__(self, input_args, dest_nodes=None):
super(TensorflowParser, self).__init__()
# load model files into Keras graph
from six import string_types as _string_types
if isinstance(input_args, _string_types):
model = TensorflowParser._load_meta(input_args)
elif isinstance(input_args, tuple):
model = TensorflowParser._load_meta(input_args[0])
self.ckpt_data = TensorflowParser._load_weights(input_args[1])
self.weight_loaded = True
if dest_nodes != None:
from tensorflow.python.framework.graph_util import extract_sub_graph
model = extract_sub_graph(model, dest_nodes.split(','))
# Build network graph
self.tf_graph = TensorflowGraph(model)
self.tf_graph.build()
@classmethod
def _skip_node(cls, source_node):
if source_node.covered:
return True
for prefix in cls.skip_prefix:
if source_node.name.startswith(prefix):
return True
scopes = TensorflowParser._get_scopes(source_node.name)
for s in scopes:
if s in cls.skip_scope:
return True
return False
@staticmethod
def tensor_shape_to_list(shapes):
if isinstance(shapes, attr_value_pb2.AttrValue):
return [dim.size for dim in shapes.shape.dim]
else:
ret = []
for shape in shapes:
this_one = [dim.size for dim in shape.dim]
ret.append(this_one)
return ret
def _convert_padding(self, source_node, IR_node, kernel_size):
# TODO: Fused conv and pool with padding is different from defused operators
input_node = self.get_parent(source_node.name, [0])
input_shape = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
if source_node.get_attr('padding') == 'VALID':
dims = len(input_shape)
assign_IRnode_values(IR_node, {
'auto_pad': "VALID",
'pads': [0, 0] * dims
})
elif source_node.get_attr('padding') == 'SAME':
padding = compute_tf_same_padding(input_shape, kernel_size,
source_node.get_attr('strides'))
assign_IRnode_values(IR_node, {
'auto_pad': "SAME_LOWER",
'pads': padding
})
else:
assert False
def _convert_pooling(self, source_node, pool_type):
IR_node = self._convert_identity_operation(source_node, new_op='Pool')
kwargs = {}
# strides
kwargs['strides'] = source_node.get_attr('strides')
# window_shape
kwargs['kernel_shape'] = source_node.get_attr('ksize')
# pool type
kwargs['pooling_type'] = pool_type
# padding
self._convert_padding(source_node, IR_node,
kwargs['kernel_shape'][1:-1])
assign_IRnode_values(IR_node, kwargs)
def gen_IR(self):
for layer in self.src_graph.topological_sort:
current_node = self.src_graph.get_node(layer)
if self._skip_node(current_node):
continue
node_type = current_node.type
if hasattr(self, "rename_" + node_type):
func = getattr(self, "rename_" + node_type)
func(current_node)
else:
self.rename_UNKNOWN(current_node)
@staticmethod
def _copy_and_reop(source_node, IR_node, new_op=None):
if new_op == None: new_op = source_node.type
IR_node.name = source_node.name
IR_node.op = new_op
kwargs = {}
if 'data_format' in source_node.layer.attr:
kwargs['data_format'] = source_node.get_attr('data_format')
if 'dtype' in source_node.layer.attr:
assert source_node.layer.attr[
'dtype'].type in TensorflowParser.dtype_map, 'type [{}] is unknown.'.format(
source_node.layer.attr['dtype'].type)
IR_node.attr["dtype"].type = TensorflowParser.dtype_map[
source_node.layer.attr['dtype'].type]
if '_output_shapes' in source_node.layer.attr:
IR_node.attr["_output_shapes"].MergeFromString(
source_node.layer.attr['_output_shapes'].SerializeToString())
assign_IRnode_values(IR_node, kwargs)
def _convert_inedge(self, source_node, IR_node, start_idx=0, end_idx=None):
if end_idx == None: end_idx = len(source_node.in_edges)
for idx in range(start_idx, end_idx):
IR_node.input.append(
self.src_graph.get_node(source_node.in_edges[idx]).real_name)
def _get_bias(self, source_node, IR_node):
if not source_node.out_edges:
return
add_node = self.tf_graph.get_node(source_node.out_edges[0])
if add_node.type != "Add" and add_node.type != "BiasAdd":
return
variable = self.tf_graph.get_node(add_node.in_edges[1])
variable = self.tf_graph.get_node(variable.in_edges[0])
assert variable.layer.attr['shape'].shape.dim[0].size == IR_node.attr[
'kernel_shape'].list.i[-1]
if self.weight_loaded:
assert variable.name in self.ckpt_data
current_layer = self.weights[source_node.name]
current_layer['bias'] = self.ckpt_data[variable.name]
add_node.real_name = IR_node.name
add_node.covered = True
IR_node.attr['use_bias'].b = True
@staticmethod
def _copy_shape(source_node, IR_node):
assert 'shape' in source_node.layer.attr
IR_node.attr['shape'].shape.MergeFromString(
source_node.layer.attr['shape'].shape.SerializeToString())
def rename_UNKNOWN(self, source_node):
if source_node.type in self.skip_type:
return
print("Tensorflow has not supported operator [%s] with name [%s]." %
(source_node.type, source_node.name))
return
def rename_Placeholder(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='DataInput')
# shape
TensorflowParser._copy_shape(source_node, IR_node)
def rename_Conv2D(self, source_node):
"""
weights: name_weights, name_bias
"""
IR_node = self._convert_identity_operation(source_node, 1, 'Conv')
kwargs = {}
# strides
kwargs['strides'] = source_node.get_attr('strides')
# input[1] : W
# filter
W = self.tf_graph.get_node(source_node.layer.input[1])
W = self.tf_graph.get_node(W.layer.input[0]).layer
kwargs['kernel_shape'] = self.tensor_shape_to_list(W.attr['shape'])
# padding
self._convert_padding(source_node, IR_node,
kwargs['kernel_shape'][:-2])
if self.weight_loaded:
self.set_weight(source_node.name, 'weights',
self.ckpt_data[W.name])
assign_IRnode_values(IR_node, kwargs)
# output[0] : B
self._get_bias(source_node, IR_node)
def _convert_identity_operation(self,
source_node,
in_edge_count=None,
new_op=None):
IR_node = self.IR_graph.node.add()
TensorflowParser._copy_and_reop(source_node, IR_node, new_op)
self._convert_inedge(source_node, IR_node, 0, in_edge_count)
return IR_node
def rename_Relu(self, source_node):
self._convert_identity_operation(source_node)
def rename_Relu6(self, source_node):
self._convert_identity_operation(source_node)
def rename_Add(self, source_node):
if not source_node.covered:
scopes = self._get_scopes(source_node.name)
if len(scopes) < 3:
self._convert_identity_operation(source_node)
elif scopes[-2] == 'dropout':
# converted [dropout]
pass
elif scopes[-2] == 'batchnorm':
# convert [tf.contrib.layers.batch_norm]
self._convert_layers_batchnorm(source_node)
else:
# normal Add
self._convert_identity_operation(source_node)
def rename_Sub(self, source_node):
self._convert_identity_operation(source_node)
def rename_Reshape(self, source_node):
IR_node = self._convert_identity_operation(source_node, 1)
kwargs = {
'shape':
self.tensor_shape_to_list(
source_node.get_attr('_output_shapes'))[0]
}
assign_IRnode_values(IR_node, kwargs)
def rename_MatMul(self, source_node):
"""
weights: name_weights, name_bias
"""
IR_node = self._convert_identity_operation(source_node, 1)
# units
units = source_node.layer.attr['_output_shapes'].list.shape[-1].dim[
-1].size
IR_node.attr['units'].i = units
# Weights
W = self.tf_graph.get_node(
self.tf_graph.get_node(source_node.in_edges[1]).in_edges[0])
if self.weight_loaded:
self.set_weight(source_node.name, 'weights',
self.ckpt_data[W.name])
if source_node.out_edges:
add_node = self.tf_graph.get_node(source_node.out_edges[0])
if add_node.type == 'Add':
add_node.covered = True
add_node.real_name = source_node.real_name
# FullyConnected Layer
# name, op
TensorflowParser._copy_and_reop(source_node, IR_node,
'FullyConnected')
# get Bias
B = self.tf_graph.get_node(
self.tf_graph.get_node(
source_node.out_edges[0]).in_edges[1]).in_edges[0]
if self.weight_loaded:
self.set_weight(source_node.name, 'bias',
self.ckpt_data[B])
IR_node.attr['use_bias'].b = True
else:
raise NotImplementedError(
"Not implemented yet. Please submit a issue in github and provide your models for reproduce."
)
else:
# Matmul Layer
TensorflowParser._copy_and_reop(source_node, IR_node,
'FullyConnected')
assign_IRnode_values(IR_node, {'use_bias': False})
def rename_RealDiv(self, source_node):
scopes = self._get_scopes(source_node.name)
# Deal Dropout
if scopes[-2] == 'dropout':
IR_node = self._convert_identity_operation(source_node, 1,
'Dropout')
# keep prob
if 'value' in self.tf_graph.get_node(
source_node.layer.input[1]).layer.attr:
IR_node.attr['keep_prob'].f = self.tf_graph.get_node(
source_node.layer.input[1]
).layer.attr['value'].tensor.float_val[0]
else:
IR_node.attr['keep_prob'].f = 1.0
# Remove nodes
# Mul
mul_node = self.tf_graph.get_node(source_node.out_edges[0])
assert mul_node.type == "Mul"
mul_node.covered = True
mul_node.real_name = source_node.name
# Floor
floor_node = self.tf_graph.get_node(mul_node.in_edges[1])
assert floor_node.type == "Floor"
floor_node.covered = True
else:
assert False
def rename_Floor(self, source_node):
scopes = self._get_scopes(source_node.name)
assert scopes[-2] == 'dropout'
def rename_MaxPool(self, source_node):
self._convert_pooling(source_node, b'MAX')
def rename_AvgPool(self, source_node):
self._convert_pooling(source_node, b'AVG')
def rename_Identity(self, source_node):
source_node.real_name = self.src_graph.get_node(
source_node.in_edges[0]).real_name
def rename_Squeeze(self, source_node):
IR_node = self._convert_identity_operation(source_node)
IR_node.attr['axes'].MergeFromString(
source_node.layer.attr['squeeze_dims'].SerializeToString())
def rename_QueueDequeueManyV2(self, source_node):
IR_node = self._convert_identity_operation(source_node, 0, 'DataInput')
IR_node.attr['shape'].shape.MergeFromString(
source_node.layer.attr['_output_shapes'].list.shape[0].
SerializeToString())
IR_node.attr['shape'].shape.dim[0].size = -1
IR_node.attr['dtype'].type = self.dtype_map[
source_node.layer.attr['component_types'].list.type[0]]
def rename_Pad(self, source_node):
IR_node = self._convert_identity_operation(source_node, 1, 'Pad')
kwargs = {}
kwargs['mode'] = 'constant'
kwargs['constant_values'] = 0.0
# paddings
padding = self.get_parent(source_node.name,
[1]).layer.attr['value'].tensor
shapes = tensor_util.MakeNdarray(padding)
kwargs['pads'] = convert_tf_pad_to_onnx(shapes)
assign_IRnode_values(IR_node, kwargs)
def rename_Mean(self, source_node):
self._convert_reduction_operators(source_node, new_op='ReduceMean')
def rename_ConcatV2(self, source_node):
n = len(source_node.in_edges) - 1
IR_node = self._convert_identity_operation(source_node, n, 'Concat')
axis = self.tf_graph.get_parent(source_node.name, [n])
IR_node.attr['axis'].i = axis.layer.attr['value'].tensor.int_val[0]
def rename_DepthwiseConv2dNative(self, source_node):
IR_node = self._convert_identity_operation(source_node, 1,
'DepthwiseConv')
kwargs = {}
kwargs['strides'] = source_node.get_attr('strides')
input_node = self.src_graph.get_parent(source_node.name, [1])
kwargs['kernel_shape'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
self._convert_padding(source_node, IR_node,
kwargs['kernel_shape'][:-2])
if self.weight_loaded:
weight = self.src_graph.get_parent(source_node.name, [1, 0])
self.set_weight(source_node.name, 'weights',
self.ckpt_data[weight.name])
assign_IRnode_values(IR_node, kwargs)
def rename_FusedBatchNorm(self, source_node):
IR_node = self._convert_identity_operation(source_node, 1, 'BatchNorm')
IR_node.attr['epsilon'].f = source_node.get_attr('epsilon', 0)
# gamma (scale)
scale = self.get_parent(source_node.name, [1], True)
if scale.type == 'Const':
value = scale.get_attr('value')
shape = value.tensor_shape
assert len(shape.dim) == 1
shape = shape.dim[0].size
assert len(value.float_val) == 1
value = value.float_val[0]
if np.isclose(value, 1.0):
IR_node.attr['scale'].b = False
else:
IR_node.attr['scale'].b = True
if self.weight_loaded:
self.set_weight(source_node.name, 'scale',
np.array([value] * shape))
else:
scale = self.get_parent(scale.name, [0], True)
if self.weight_loaded:
self.set_weight(source_node.name, 'scale',
self.ckpt_data[scale.name])
IR_node.attr['scale'].b = True
# bias
bias = self.get_parent(source_node.name, [2, 0], True)
IR_node.attr['bias'].b = True
# Mean
mean = self.get_parent(source_node.name, [3, 0], True)
# Var
var = self.get_parent(source_node.name, [4, 0], True)
if self.weight_loaded:
self.set_weight(source_node.name, 'bias',
self.ckpt_data[bias.name])
self.set_weight(source_node.name, 'mean',
self.ckpt_data[mean.name])
self.set_weight(source_node.name, 'var', self.ckpt_data[var.name])
def rename_Transpose(self, source_node):
IR_node = self._convert_identity_operation(source_node, 1)
perm = self.get_parent(source_node.name,
[1]).layer.attr['value'].tensor
perm = tensor_util.MakeNdarray(perm).tolist()
assign_IRnode_values(IR_node, {'perm': perm})
def rename_Sigmoid(self, source_node):
IR_node = self._convert_identity_operation(source_node)
def rename_Mul(self, source_node):
IR_node = self._convert_identity_operation(source_node)
def __init__(self, frozen_file, inputshape, dest_nodes):
super(TensorflowParser2, self).__init__()
self.weight_loaded = True
# load model files into TensorFlow graph
with open(frozen_file, 'rb') as f:
serialized = f.read()
tensorflow.reset_default_graph()
original_gdef = tensorflow.GraphDef()
original_gdef.ParseFromString(serialized)
model = original_gdef
# if dest_nodes != None:
# from tensorflow.python.framework.graph_util import extract_sub_graph
# model = extract_sub_graph(model, dest_nodes.split(','))
output_shape_map = dict()
input_shape_map = dict()
with tensorflow.Graph().as_default() as g:
tensorflow.import_graph_def(original_gdef, name='')
ops = g.get_operations()
N = len(ops)
p = 0
for i in range(N):
for x in ops[i].inputs:
input_shape_map[x.name] = x.get_shape()
for x in ops[i].outputs:
output_shape_map[x.name] = x.get_shape()
tensor_input = tensorflow.TensorShape([
None,
tensorflow.Dimension(inputshape[0]),
tensorflow.Dimension(inputshape[1]),
tensorflow.Dimension(inputshape[2])
])
output_shape_map['input:0'] = tensor_input
self.tf_graph = TensorflowGraph(model)
for node in self.tf_graph.model.node:
if node.name == 'input' and node.op == 'Placeholder':
node.attr['shape'].list.shape.extend(
[output_shape_map[node.name + ':0'].as_proto()])
if (node.name + ':0') in output_shape_map:
node.attr['_output_shapes'].list.shape.extend(
[output_shape_map[node.name + ':0'].as_proto()])
if node.op == 'MirrorPad':
node.attr['paddings'].list.shape.extend(
[input_shape_map[node.name + '/paddings:0'].as_proto()])
if node.op == 'QuantizeV2':
node.attr['shape'].list.shape.extend(
[input_shape_map[node.name + ':0'].as_proto()])
if node.op == 'RequantizationRange':
map_key = node.name.split(
'eightbit')[0] + "eightbit_quantized_conv:0"
node.attr['shape'].list.shape.extend(
[input_shape_map[map_key].as_proto()])
if node.op == 'Requantize':
map_key = node.name.replace("requantize",
"quantized_conv") + ":0"
node.attr['shape'].list.shape.extend(
[input_shape_map[map_key].as_proto()])
self.tf_graph.build()
class TensorflowParser2(Parser):
# skip_prefix = [
# "^",
# "train_op",
# "save",
# "gradients",
# "init",
# "global_step",
# "distort_image",
# "Adagrad",
# ]
skip_prefix = []
skip_scope = [
"random_uniform", "Initializer", "optimizer", "weight_loss",
"parallel_read", "case"
]
skip_type = set([
"L2Loss",
"VariableV2",
"Const",
"Assign",
"RandomUniform",
"FIFOQueueV2",
"Assert",
"Unpack",
"NextIteration",
"TensorArrayV3",
"Range",
"TensorArrayScatterV3",
"TensorArrayReadV3",
"TensorArrayWriteV3",
# "Switch"
"Dequantize",
"RequantizationRange",
"Requantize",
"ExpandDims",
"Cast"
])
q_type = set([
"Dequantize", "QuantizeV2", "QuantizedConv2D", "QuantizedReshape",
"RequantizationRange"
])
dtype_map = {
0: graph_pb2.DT_UNDEFINED,
1: graph_pb2.DT_FLOAT32,
2: graph_pb2.DT_FLOAT64,
3: graph_pb2.DT_INT32,
4: graph_pb2.DT_UINT8,
5: graph_pb2.DT_INT16,
6: graph_pb2.DT_INT8,
7: graph_pb2.DT_STRING,
9: graph_pb2.DT_INT64
}
@property
def src_graph(self):
return self.tf_graph
def __init__(self, frozen_file, inputshape, dest_nodes):
super(TensorflowParser2, self).__init__()
self.weight_loaded = True
# load model files into TensorFlow graph
with open(frozen_file, 'rb') as f:
serialized = f.read()
tensorflow.reset_default_graph()
original_gdef = tensorflow.GraphDef()
original_gdef.ParseFromString(serialized)
model = original_gdef
# if dest_nodes != None:
# from tensorflow.python.framework.graph_util import extract_sub_graph
# model = extract_sub_graph(model, dest_nodes.split(','))
output_shape_map = dict()
input_shape_map = dict()
with tensorflow.Graph().as_default() as g:
tensorflow.import_graph_def(original_gdef, name='')
ops = g.get_operations()
N = len(ops)
p = 0
for i in range(N):
for x in ops[i].inputs:
input_shape_map[x.name] = x.get_shape()
for x in ops[i].outputs:
output_shape_map[x.name] = x.get_shape()
tensor_input = tensorflow.TensorShape([
None,
tensorflow.Dimension(inputshape[0]),
tensorflow.Dimension(inputshape[1]),
tensorflow.Dimension(inputshape[2])
])
output_shape_map['input:0'] = tensor_input
self.tf_graph = TensorflowGraph(model)
for node in self.tf_graph.model.node:
if node.name == 'input' and node.op == 'Placeholder':
node.attr['shape'].list.shape.extend(
[output_shape_map[node.name + ':0'].as_proto()])
if (node.name + ':0') in output_shape_map:
node.attr['_output_shapes'].list.shape.extend(
[output_shape_map[node.name + ':0'].as_proto()])
if node.op == 'MirrorPad':
node.attr['paddings'].list.shape.extend(
[input_shape_map[node.name + '/paddings:0'].as_proto()])
if node.op == 'QuantizeV2':
node.attr['shape'].list.shape.extend(
[input_shape_map[node.name + ':0'].as_proto()])
if node.op == 'RequantizationRange':
map_key = node.name.split(
'eightbit')[0] + "eightbit_quantized_conv:0"
node.attr['shape'].list.shape.extend(
[input_shape_map[map_key].as_proto()])
if node.op == 'Requantize':
map_key = node.name.replace("requantize",
"quantized_conv") + ":0"
node.attr['shape'].list.shape.extend(
[input_shape_map[map_key].as_proto()])
self.tf_graph.build()
@staticmethod
def _get_scopes(layer_name):
return layer_name.split('/')
def _convert_reduction_operators(self, source_node, new_op=None):
IR_node = self._convert_identity_operation(source_node, 1, new_op)
# keep dims
IR_node.attr['keepdims'].b = source_node.layer.attr['keep_dims'].b
# axes
axes = self.get_parent(source_node.name,
[1]).layer.attr['value'].tensor
axes = tensor_util.MakeNdarray(axes)
IR_node.attr['axes'].list.i.extend(axes)
def _convert_layers_batchnorm(self, source_node):
# name, op
IR_node = self.IR_graph.node.add()
TensorflowParser2._copy_and_reop(source_node, IR_node, 'BatchNorm')
# epsilon
epsilon = self.get_parent(source_node.name, [1])
IR_node.attr['epsilon'].f = epsilon.layer.attr[
'value'].tensor.float_val[0]
# moving variance (var) /read
moving_variance = self.get_parent(source_node.name, [0])
if moving_variance.type == 'Identity':
moving_variance_read = self.src_graph.get_parent(
moving_variance.name, [0])
tensor_content = moving_variance_read.get_attr('value')
moving_variance_content = tensor_util.MakeNdarray(tensor_content)
self.set_weight(source_node.name, 'var', moving_variance_content)
else:
print(moving_variance.layer)
assert False
# gamma (scale)
Rsqrt = self.get_son(source_node.name, [0], True)
if len(Rsqrt.out_edges) == 2:
IR_node.attr['scale'].b = False
output_node = self.get_son(source_node.name, [0, 0, 0], True)
else:
IR_node.attr['scale'].b = True
# assert False
son = self.get_son(source_node.name, [0, 0, 0], True)
gamma = self.get_parent(son.name, [1, 1, 0, 0, 0, 1], True)
gamma_tensor = gamma.get_attr('value')
scale = tensor_util.MakeNdarray(gamma_tensor)
self.set_weight(source_node.name, 'scale', scale)
output_node = self.get_son(source_node.name, [0, 0, 0, 0], True)
# print(output_node.layer)
# beta (bias)
Sub = self.get_son(source_node.name, [0, 1, 0], True)
beta = self.get_parent(Sub.name, [0, 0]).get_attr('value')
bias = tensor_util.MakeNdarray(beta) #(96,)
IR_node.attr['bias'].b = True
self.set_weight(source_node.name, 'bias', bias)
# moving mean (mean)
son2 = self.get_son(source_node.name, [0, 1], True)
moving_mean = self.get_parent(son2.name, [0, 0]).get_attr('value')
mean = tensor_util.MakeNdarray(moving_mean)
self.set_weight(source_node.name, 'mean', mean)
# input node
assert output_node.type == 'Add'
input_node = self.get_parent(output_node.name, [0, 0])
IR_node.input.append(input_node.real_name)
# output node
output_node.real_name = source_node.name
def _convert_layers_instancenorm(self, source_node):
IR_node = self.IR_graph.node.add()
TensorflowParser2._copy_and_reop(source_node, IR_node, 'InstanceNorm')
# epsilon
epsilon = self.get_parent(source_node.name, [1])
epsilon_value = epsilon.get_attr('value').float_val[0]
IR_node.attr['epsilon'].f = epsilon_value
# beta
# gamma (scale)
IR_node.attr['scale'].b = True
son = self.get_son(source_node.name, [0, 0, 0], True)
gamma = self.get_parent(son.name, [1, 1, 0, 0, 0, 1], True)
gamma_tensor = gamma.get_attr('value')
scale = tensor_util.MakeNdarray(gamma_tensor)
self.set_weight(source_node.name, 'scale', scale)
# output_node = self.get_son(source_node.name, [0, 0, 0, 0], True)
assert False
@classmethod
def _skip_node(cls, source_node):
if source_node.covered:
return True
for prefix in cls.skip_prefix:
if source_node.name.startswith(prefix):
return True
scopes = TensorflowParser2._get_scopes(source_node.name)
for s in scopes:
if s in cls.skip_scope:
return True
return False
def gen_IR(self):
for layer in self.src_graph.topological_sort:
current_node = self.src_graph.get_node(layer)
if self._skip_node(current_node):
continue
node_type = current_node.type
if hasattr(self, "rename_" + node_type):
func = getattr(self, "rename_" + node_type)
func(current_node)
else:
self.rename_UNKNOWN(current_node)
@staticmethod
def tensor_shape_to_list(shapes):
if isinstance(shapes, attr_value_pb2.AttrValue):
return [dim.size for dim in shapes.shape.dim]
else:
ret = []
for shape in shapes:
this_one = [dim.size for dim in shape.dim]
ret.append(this_one)
return ret
@staticmethod
def _copy_and_reop(source_node, IR_node, new_op=None):
if new_op == None: new_op = source_node.type
IR_node.name = source_node.name
IR_node.op = new_op
kwargs = {}
if 'data_format' in source_node.layer.attr:
kwargs['data_format'] = source_node.get_attr('data_format')
if 'T' in source_node.layer.attr:
if source_node.type not in TensorflowParser2.q_type:
if source_node.type == 'Enter':
IR_node.attr["dtype"].type = TensorflowParser2.dtype_map[6]
else:
assert source_node.layer.attr[
'T'].type in TensorflowParser2.dtype_map, 'type [{}] is unknown.'.format(
source_node.layer.attr['dtype'].type)
IR_node.attr["dtype"].type = TensorflowParser2.dtype_map[
source_node.layer.attr['T'].type]
else:
IR_node.attr["dtype"].type = TensorflowParser2.dtype_map[6]
if '_output_shapes' in source_node.layer.attr:
IR_node.attr["_output_shapes"].MergeFromString(
source_node.layer.attr['_output_shapes'].SerializeToString())
if 'paddings' in source_node.layer.attr:
IR_node.attr["paddings"].MergeFromString(
source_node.layer.attr['paddings'].SerializeToString())
assign_IRnode_values(IR_node, kwargs)
def _convert_inedge(self, source_node, IR_node, start_idx=0, end_idx=None):
if end_idx == None: end_idx = len(source_node.in_edges)
for idx in range(start_idx, end_idx):
IR_node.input.append(
self.src_graph.get_node(source_node.in_edges[idx]).real_name)
@staticmethod
def _copy_shape(source_node, IR_node):
assert 'shape' in source_node.layer.attr
if source_node.layer.attr['shape'].list.shape:
IR_node.attr['shape'].shape.MergeFromString(
source_node.layer.attr['shape'].list.shape[0].
SerializeToString())
else:
IR_node.attr['shape'].shape.MergeFromString(
source_node.layer.attr['shape'].shape.SerializeToString())
# print(IR_node)
def rename_UNKNOWN(self, source_node):
if source_node.type in self.skip_type:
return
print("Tensorflow has not supported operator [%s] with name [%s]." %
(source_node.type, source_node.name))
assert False
def rename_NoOp(self, source_node):
return
def _convert_padding(self, source_node, IR_node, kernel_size):
# TODO: Fused conv and pool with padding is different from defused operators
input_node = self.get_parent(source_node.name, [0])
input_shape = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
if source_node.get_attr('padding') == 'VALID':
dims = len(input_shape)
assign_IRnode_values(IR_node, {
'auto_pad': "VALID",
'pads': [0, 0] * dims
})
elif source_node.get_attr('padding') == 'SAME':
padding = compute_tf_same_padding(input_shape, kernel_size,
source_node.get_attr('strides'))
assign_IRnode_values(IR_node, {
'auto_pad': "SAME_LOWER",
'pads': padding
})
else:
assert False
def _get_bias(self, source_node, IR_node):
if not source_node.out_edges:
return
add_node = self.tf_graph.get_node(source_node.out_edges[0])
if add_node.type != "Add" and add_node.type != "BiasAdd":
return
variable = self.tf_graph.get_node(add_node.in_edges[1]) #add_bias node
if variable.type == 'Identity':
variable = self.tf_graph.get_node(variable.in_edges[0])
bias_value = variable.get_attr('value')
bias = tensor_util.MakeNdarray(bias_value)
# assert variable.get_attr('_output_shapes')[0].dim[0].size == IR_node.attr['kernel_shape'].list.i[-1]
add_node.real_name = IR_node.name
add_node.covered = True
IR_node.attr['use_bias'].b = True
current_layer = self.weights[source_node.name]
current_layer['bias'] = bias
def _convert_pooling(self, source_node, pool_type):
IR_node = self._convert_identity_operation(source_node, new_op='Pool')
kwargs = {}
# strides
kwargs['strides'] = source_node.get_attr('strides')
# window_shape
kwargs['kernel_shape'] = source_node.get_attr('ksize')
# pool type
kwargs['pooling_type'] = pool_type
# padding
self._convert_padding(source_node, IR_node,
kwargs['kernel_shape'][1:-1])
assign_IRnode_values(IR_node, kwargs)
def _convert_identity_operation(self,
source_node,
start_idx=0,
end_idx=None,
new_op=None):
IR_node = self.IR_graph.node.add()
TensorflowParser2._copy_and_reop(source_node, IR_node, new_op)
self._convert_inedge(source_node, IR_node, start_idx, end_idx)
return IR_node
def rename_Placeholder(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='DataInput')
TensorflowParser2._copy_shape(source_node, IR_node)
def rename_Reshape(self, source_node):
IR_node = self._convert_identity_operation(source_node, end_idx=1)
kwargs = {
'shape':
self.tensor_shape_to_list(
source_node.get_attr('_output_shapes'))[0]
}
assign_IRnode_values(IR_node, kwargs)
def rename_MirrorPad(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='MirrorPad')
input_node = self.src_graph.get_parent(source_node.name, [1])
tensor_content = tensor_util.MakeNdarray(
input_node.get_attr('value')).reshape(-1)
# print(tensor_content)
kwargs = {}
kwargs['mode'] = source_node.get_attr('mode')
kwargs['pads'] = tensor_content.tolist()
assign_IRnode_values(IR_node, kwargs)
# print(IR_node)
def rename_Min(self, source_node):
# print(source_node.layer)
IR_node = self._convert_identity_operation(source_node,
start_idx=0,
end_idx=1,
new_op='Min')
kwargs = {}
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs['shape_0'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
input_node = self.src_graph.get_parent(source_node.name, [1])
kwargs['shape_1'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
# assert False
def rename_Max(self, source_node):
# print(source_node.layer)
IR_node = self._convert_identity_operation(source_node,
start_idx=0,
end_idx=1,
new_op='Max')
kwargs = {}
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs['shape_0'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
input_node = self.src_graph.get_parent(source_node.name, [1])
kwargs['shape_1'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
# assert False
def rename_Mul(self, source_node):
input_node_0 = self.src_graph.get_parent(source_node.name, [0])
# mean/read
if input_node_0.type == 'Identity':
input_node_0_read = self.src_graph.get_parent(
input_node_0.name, [0])
tensor_content = input_node_0_read.get_attr('value')
tensor_content = tensor_util.MakeNdarray(tensor_content)
self.set_weight(source_node.name, 'weights', tensor_content)
IR_node = self._convert_identity_operation(source_node,
start_idx=1)
else:
IR_node = self._convert_identity_operation(source_node)
def rename_Add(self, source_node):
scopes = self._get_scopes(source_node.name)
# print(scopes)
if scopes[-2] == 'batchnorm':
if scopes[-3] == 'BatchNorm':
self._convert_layers_batchnorm(source_node)
elif scopes[-3] == 'InstanceNorm':
print(source_node.layer)
# self._convert_layers_instancenorm(source_node)
else:
# print(scopes)
IR_node = self._convert_identity_operation(source_node,
new_op="Add")
# assert False
def rename_Sub(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op="Sub")
# kwargs = {}
# print(source_node.in_edges)
# input_node = self.src_graph.get_parent(source_node.name, [0])
# kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
# print(IR_node)
# assign_IRnode_values(IR_node, kwargs)
def rename_Sum(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op='Sum')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs = {}
kwargs['cal_shape'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
input_node_indices = self.src_graph.get_parent(source_node.name, [1])
indice_value = input_node_indices.get_attr('value')
if indice_value.tensor_content:
shapes = tensor_util.MakeNdarray(indice_value)
c = shapes.tolist()
kwargs['sum_indices'] = c
else:
kwargs['sum_indices'] = input_node_indices.get_attr(
'value').int_val[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Rsqrt(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op="Rsqrt")
kwargs = {}
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs['shape'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Square(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='Square')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs = {}
kwargs['shape'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Sigmoid(self, source_node):
IR_node = self._convert_identity_operation(source_node)
kwargs = {}
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs['shape'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Reciprocal(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='Reciprocal')
def rename_Minimum(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='Minimum')
def rename_Maximum(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='Maximum')
def rename_RealDiv(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='RealDiv')
def rename_Enter(self, source_node):
# print(source_node.layer)
IR_node = self._convert_identity_operation(source_node, new_op='Enter')
def rename_Switch(self, source_node):
source_node.real_name = self.src_graph.get_node(
source_node.in_edges[0]).real_name
def rename_Exp(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op='Exp')
def rename_ResizeBilinear(self, source_node):
# print(source_node.layer)
IR_node = self._convert_identity_operation(source_node,
new_op='ResizeBilinear')
# def rename_Cast(self, source_node):
# IR_node = self._convert_identity_operation(source_node, new_op = 'Cast')
# input_node = self.src_graph.get_parent(source_node.name, [0])
# kwargs = {}
# kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
# assign_IRnode_values(IR_node, kwargs)
def rename_Prod(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op='Prod')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs = {}
kwargs['shape'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
input_node_const = self.src_graph.get_parent(source_node.name, [1])
kwargs['const'] = input_node_const.get_attr('value').int_val[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Shape(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op='Shape')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs = {}
kwargs['shape'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Squeeze(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='Squeeze')
def rename_Pack(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op='Pack')
def rename_Gather(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='Gather')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs = {}
# kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
input_node_indices = self.src_graph.get_parent(source_node.name, [1])
indice_value = input_node_indices.get_attr('value')
shapes = tensor_util.MakeNdarray(indice_value)
c = shapes.tolist()
kwargs['gather_indices'] = c
assign_IRnode_values(IR_node, kwargs)
def rename_StridedSlice(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='StridedSlice')
kwargs = {}
input_node = self.src_graph.get_parent(source_node.name, [0])
# kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
input_node_const0 = self.src_graph.get_parent(source_node.name, [1])
input_node_const1 = self.src_graph.get_parent(source_node.name, [2])
input_node_const2 = self.src_graph.get_parent(source_node.name, [3])
if input_node_const0.get_attr('value').int_val:
kwargs['const0'] = input_node_const0.get_attr('value').int_val[0]
if input_node_const1.get_attr('value').int_val:
kwargs['const1'] = input_node_const1.get_attr('value').int_val[0]
if input_node_const2.get_attr('value').int_val:
kwargs['const2'] = input_node_const2.get_attr('value').int_val[0]
assign_IRnode_values(IR_node, kwargs)
# def rename_ExpandDims(self, source_node):
# IR_node = self._convert_identity_operation(source_node, new_op = 'ExpandDims')
# input_node = self.src_graph.get_parent(source_node.name, [0])
# kwargs = {}
# kwargs['shape'] = self.tensor_shape_to_list(input_node.get_attr('_output_shapes'))[0]
# input_node_indices = self.src_graph.get_parent(source_node.name, [1])
# kwargs['exp_dim'] = input_node_indices.get_attr('value').int_val[0]
# assign_IRnode_values(IR_node, kwargs)
def rename_ResizeNearestNeighbor(self, source_node):
IR_node = self._convert_identity_operation(source_node)
kwargs = {}
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs['shape'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
input_node_size = self.src_graph.get_parent(source_node.name, [1])
kwargs['size'] = self.tensor_shape_to_list(
input_node_size.get_attr('_output_shapes'))[0]
assign_IRnode_values(IR_node, kwargs)
def rename_Conv2D(self, source_node):
IR_node = self._convert_identity_operation(source_node,
end_idx=1,
new_op='Conv')
kwargs = {}
kwargs['strides'] = source_node.get_attr('strides')
kwargs['padding'] = source_node.get_attr('padding')
input_node = self.src_graph.get_parent(source_node.name, [0])
kwargs['shape'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
# weights
input_node_weight = self.src_graph.get_parent(source_node.name, [1])
if input_node_weight.type == 'Const':
tensor_content = input_node_weight.get_attr('value')
else:
input_node_weight_read = self.src_graph.get_parent(
input_node_weight.name, [0])
tensor_content = input_node_weight_read.get_attr('value')
W = tensor_util.MakeNdarray(tensor_content)
kwargs['kernel_shape'] = self.tensor_shape_to_list(
input_node_weight.get_attr('_output_shapes'))[0]
self.set_weight(source_node.name, 'weights', W)
self._convert_padding(source_node, IR_node,
kwargs['kernel_shape'][:-2])
assign_IRnode_values(IR_node, kwargs)
self._get_bias(source_node, IR_node)
def rename_Relu(self, source_node):
IR_node = self._convert_identity_operation(source_node)
kwargs = {
'shape':
self.tensor_shape_to_list(
source_node.get_attr('_output_shapes'))[0]
}
assign_IRnode_values(IR_node, kwargs)
def rename_MaxPool(self, source_node):
# print(source_node.layer.attr)
self._convert_pooling(source_node, b'MAX')
def rename_AvgPool(self, source_node):
self._convert_pooling(source_node, b'AVG')
def rename_LRN(self, source_node):
IR_node = self._convert_identity_operation(source_node)
# alpha
IR_node.attr["alpha"].f = float(source_node.get_attr(
"alpha", "0.0001"))
# beta
IR_node.attr["beta"].f = float(source_node.get_attr("beta", "0.75"))
IR_node.attr["size"].i = source_node.get_attr("depth_radius")
IR_node.attr["bias"].f = float(source_node.get_attr("bias"))
def rename_Concat(self, source_node):
n = len(source_node.in_edges)
IR_node = self._convert_identity_operation(source_node,
start_idx=1,
end_idx=n,
new_op='Concat')
axis = self.tf_graph.get_parent(source_node.name, [0])
IR_node.attr["axis"].i = axis.get_attr('value').int_val[0]
def rename_ConcatV2(self, source_node):
n = len(source_node.in_edges)
IR_node = self._convert_identity_operation(source_node,
start_idx=0,
end_idx=n - 1,
new_op='Concat')
axis = self.tf_graph.get_parent(source_node.name, [n - 1])
IR_node.attr["axis"].i = axis.get_attr('value').int_val[0]
def rename_MatMul(self, source_node):
IR_node = self._convert_identity_operation(source_node, end_idx=1)
input_weight_node = self.src_graph.get_parent(source_node.name, [1])
weight_value = input_weight_node.get_attr('value')
weight = tensor_util.MakeNdarray(weight_value)
self.set_weight(source_node.name, 'weights', weight)
units = source_node.layer.attr['_output_shapes'].list.shape[-1].dim[
-1].size
IR_node.attr['units'].i = units
if source_node.out_edges and self.tf_graph.get_node(
source_node.out_edges[0]).type == 'BiasAdd':
add_node = self.tf_graph.get_node(source_node.out_edges[0])
add_node.covered = True
add_node.real_name = source_node.real_name
TensorflowParser2._copy_and_reop(source_node, IR_node,
'FullyConnected')
variable = self.tf_graph.get_node(
add_node.in_edges[1]) #add_bias node
bias_value = variable.get_attr('value')
bias = tensor_util.MakeNdarray(bias_value)
self.set_weight(source_node.name, 'bias', bias)
IR_node.attr['use_bias'].b = True
def rename_Softmax(self, source_node):
IR_node = self._convert_identity_operation(source_node)
kwargs = {
'shape':
self.tensor_shape_to_list(
source_node.get_attr('_output_shapes'))[0]
}
IR_node.attr["dim"].i = 1
assign_IRnode_values(IR_node, kwargs)
def rename_Identity(self, source_node):
source_node.real_name = self.src_graph.get_node(
source_node.in_edges[0]).real_name
def rename_QuantizeV2(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op='QuantizeV2')
TensorflowParser2._copy_shape(source_node, IR_node)
def rename_QuantizedRelu(self, source_node):
IR_node = self._convert_identity_operation(source_node,
new_op="QuantizedRelu")
kwargs = {
'shape':
self.tensor_shape_to_list(
source_node.get_attr('_output_shapes'))[0]
}
assign_IRnode_values(IR_node, kwargs)
def rename_QuantizedReshape(self, source_node):
IR_node = self._convert_identity_operation(source_node, end_idx=1)
kwargs = {
'shape':
self.tensor_shape_to_list(
source_node.get_attr('_output_shapes'))[0]
}
assign_IRnode_values(IR_node, kwargs)
def rename_QuantizedConv2D(self, source_node):
IR_node = self._convert_identity_operation(source_node, new_op='QConv')
kwargs = {}
kwargs['strides'] = source_node.get_attr('strides')
kwargs['padding'] = source_node.get_attr('padding')
# weights
input_node = self.src_graph.get_parent(source_node.name, [1])
tensor_content = input_node.get_attr('value')
W = tensor_util.MakeNdarray(tensor_content)
W = W.astype(np.uint8)
kwargs['kernel_shape'] = self.tensor_shape_to_list(
input_node.get_attr('_output_shapes'))[0]
input_node_minw = self.src_graph.get_parent(source_node.name, [4])
min_W = input_node_minw.get_attr('value').float_val[0]
input_node_maxw = self.src_graph.get_parent(source_node.name, [5])
max_W = input_node_maxw.get_attr('value').float_val[0]
if source_node.get_attr('Tfilter') == tensorflow.quint8:
W = ((max_W - min_W) / 255.0) * W + min_W
else:
assert False, (
'Only uint8 weights handled currently by the converter')
self.set_weight(source_node.name, 'kernel_weights', W)
assign_IRnode_values(IR_node, kwargs)
def __init__(self, frozen_file, inputshape, in_nodes, dest_nodes):
super(TensorflowParser2, self).__init__()
self.weight_loaded = True
# load model files into TensorFlow graph
with open(frozen_file, 'rb') as f:
serialized = f.read()
tensorflow.reset_default_graph()
original_gdef = tensorflow.GraphDef()
original_gdef.ParseFromString(serialized)
# model = original_gdef
from tensorflow.python.tools import strip_unused_lib
from tensorflow.python.framework import dtypes
from tensorflow.python.platform import gfile
input_node_names = in_nodes.split(',')
output_node_names = dest_nodes.split(',')
gdef = strip_unused_lib.strip_unused(
input_graph_def=original_gdef,
input_node_names=input_node_names,
output_node_names=output_node_names,
placeholder_type_enum=dtypes.float32.as_datatype_enum)
# Save it to an output file
frozen_model_file = './frozen.pb'
with gfile.GFile(frozen_model_file, "wb") as f:
f.write(gdef.SerializeToString())
with open(frozen_model_file, 'rb') as f:
serialized = f.read()
tensorflow.reset_default_graph()
model = tensorflow.GraphDef()
model.ParseFromString(serialized)
output_shape_map = dict()
input_shape_map = dict()
with tensorflow.Graph().as_default() as g:
tensorflow.import_graph_def(original_gdef, name='')
ops = g.get_operations()
N = len(ops)
p = 0
for i in range(N):
for x in ops[i].inputs:
input_shape_map[x.name] = x.get_shape()
for x in ops[i].outputs:
output_shape_map[x.name] = x.get_shape()
tensor_input = tensorflow.TensorShape([
None,
tensorflow.Dimension(inputshape[0]),
tensorflow.Dimension(inputshape[1]),
tensorflow.Dimension(inputshape[2])
])
output_shape_map[input_node_names[0]] = tensor_input
self.tf_graph = TensorflowGraph(model)
for node in self.tf_graph.model.node:
if node.name == 'input' and node.op == 'Placeholder':
node.attr['shape'].list.shape.extend(
[output_shape_map[node.name + ':0'].as_proto()])
if (node.name + ':0') in output_shape_map:
node.attr['_output_shapes'].list.shape.extend(
[output_shape_map[node.name + ':0'].as_proto()])
if node.op == 'MirrorPad':
node.attr['paddings'].list.shape.extend(
[input_shape_map[node.name + '/paddings:0'].as_proto()])
if node.op == 'QuantizeV2':
node.attr['shape'].list.shape.extend(
[input_shape_map[node.name + ':0'].as_proto()])
if node.op == 'RequantizationRange':
map_key = node.name.split(
'eightbit')[0] + "eightbit_quantized_conv:0"
node.attr['shape'].list.shape.extend(
[input_shape_map[map_key].as_proto()])
if node.op == 'Requantize':
map_key = node.name.replace("requantize",
"quantized_conv") + ":0"
node.attr['shape'].list.shape.extend(
[input_shape_map[map_key].as_proto()])
if node.name == in_nodes:
node.attr['shape'].list.shape.extend(
[output_shape_map[node.name].as_proto()])
self.tf_graph.build()