def __init__(self):
#图的中间表示：根据src_graph, 统一使用GraphDef，来构造图的中间表示并生成.pb. 然后在具体的Emitter中使用这个中间表示（通过IRGraph包裹），来生成其它指定框架的代码和权重文件.
self.IR_graph = GraphDef() #GraphDef的构造过程(.node.add())，见于具体的parser子类实现
self.weight_loaded = False
# name --> (weight_name --> ndarray)
self.nodes_weights = dict() #权重的中间表示：以dict为元素的dict
class Parser(object):
def __init__(self):
self.IR_graph = GraphDef()
self.weight_loaded = False
# name --> (weight_name --> ndarray)
self.weights = dict()
@property
def src_graph(self):
raise NotImplementedError
def get_son(self, name, path, set_flag=False):
return self.src_graph.get_son(name, path, set_flag)
def get_parent(self, name, path, set_flag=False):
return self.src_graph.get_parent(name, path, set_flag)
def set_weight(self, layer_name, weight_name, data):
if not layer_name in self.weights:
self.weights[layer_name] = dict()
layer = self.weights[layer_name]
layer[weight_name] = data
def save_to_json(self, filename):
import google.protobuf.json_format as json_format
json_str = json_format.MessageToJson(self.IR_graph,
preserving_proto_field_name=True)
with open(filename, "w") as of:
of.write(json_str)
print("IR network structure is saved as [{}].".format(filename))
return json_str
def save_to_proto(self, filename):
proto_str = self.IR_graph.SerializeToString()
with open(filename, 'wb') as of:
of.write(proto_str)
print("IR network structure is saved as [{}].".format(filename))
return proto_str
def save_weights(self, filename):
if self.weight_loaded:
import numpy as np
with open(filename, 'wb') as of:
np.save(of, self.weights)
print("IR weights are saved as [{}].".format(filename))
else:
print("Warning: weights are not loaded.")
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 as_graph_def(self):
graph_pb2 = GraphDef()
graph_pb2.version = self.version
graph_pb2.node.extend([node.node_pb2 for node in self.node_dict.values()])
return graph_pb2
def __init__(self):
self.IR_graph = GraphDef()
self.weight_loaded = False
# name --> (weight_name --> ndarray)
self.weights = dict()
class Parser(object):
def __init__(self):
self.IR_graph = GraphDef()
self.weight_loaded = False
# name --> (weight_name --> ndarray)
self.weights = dict()
def run(self, dest_path):
self.gen_IR()
self.save_to_json(dest_path + ".json")
self.save_to_proto(dest_path + ".pb")
self.save_weights(dest_path + ".npy")
@property
def src_graph(self):
raise NotImplementedError
def get_son(self, name, path, set_flag=False):
return self.src_graph.get_son(name, path, set_flag)
def get_parent(self, name, path, set_flag=False):
return self.src_graph.get_parent(name, path, set_flag)
def set_weight(self, layer_name, weight_name, data):
if not layer_name in self.weights:
self.weights[layer_name] = dict()
layer = self.weights[layer_name]
layer[weight_name] = data
def save_to_json(self, filename):
import google.protobuf.json_format as json_format
json_str = json_format.MessageToJson(self.IR_graph,
preserving_proto_field_name=True)
with open(filename, "w") as of:
of.write(json_str)
print("IR network structure is saved as [{}].".format(filename))
return json_str
def save_to_proto(self, filename):
proto_str = self.IR_graph.SerializeToString()
with open(filename, 'wb') as of:
of.write(proto_str)
print("IR network structure is saved as [{}].".format(filename))
return proto_str
def save_weights(self, filename):
if self.weight_loaded:
with open(filename, 'wb') as of:
np.save(of, self.weights)
print("IR weights are saved as [{}].".format(filename))
else:
print("Warning: weights are not loaded.")
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.strip('_'))
@staticmethod
def channel_first_conv_kernel_to_IR(tensor):
dim = tensor.ndim
tensor = np.transpose(tensor, list(range(2, dim)) + [1, 0])
return tensor
@staticmethod
def channel_first_shape_to_IR(shape):
return [shape[0]] + list(shape[2:]) + [shape[1]]
@staticmethod
def channel_first_axis_to_IR(index):
if index == 0:
return 0
elif index == 1:
return -1
else:
return index - 1
class Parser(object):
node_index = 0
def __init__(self):
#图的中间表示：根据src_graph, 统一使用GraphDef，来构造图的中间表示并生成.pb. 然后在具体的Emitter中使用这个中间表示（通过IRGraph包裹），来生成其它指定框架的代码和权重文件.
self.IR_graph = GraphDef() #GraphDef的构造过程(.node.add())，见于具体的parser子类实现
self.weight_loaded = False
# name --> (weight_name --> ndarray)
self.nodes_weights = dict() #权重的中间表示：以dict为元素的dict
def run(self, dest_path):
self.gen_IR() #生成中间表示
print('save files turned off.')
self.save_to_json(dest_path + ".json")
# self.save_to_proto(dest_path + ".pb")
self.save_weights(dest_path + ".npy")
self.save_weights_txt(dest_path + ".txt")
@property
def src_graph(
self
): #图的原始表示：特定于框架. 比如tensorflow_parser的src_graph就是：self.tf_graph = TensorflowGraph(model) ,而model又是从.ckpt.meta中读出来的GraphDef
raise NotImplementedError
def get_node_son_by_name(self, name, path, set_flag=False):
return self.src_graph.get_node_son_by_name(name, path, set_flag)
def get_node_parent_by_name(self, name, path, set_flag=False):
return self.src_graph.get_node_parent_by_name(name, path, set_flag)
def set_node_weight(self, node_name, weight_name, data):
if node_name not in self.nodes_weights:
self.nodes_weights[node_name] = dict() #以dict为元素的dict
self.nodes_weights[node_name]['index'] = self.node_index
self.node_index += 1
node_weight = self.nodes_weights[node_name]
node_weight[weight_name] = data
#IR_graph本来就是GraphDef对象，因此可以直接序列化
def save_to_json(self, filename):
import google.protobuf.json_format as json_format
json_str = json_format.MessageToJson(
self.IR_graph, preserving_proto_field_name=True) #中间表示序列化为.json
with open(filename, "w") as of:
of.write(json_str)
print("IR network structure is saved as [{}].".format(filename))
return json_str
def save_to_proto(self, filename):
proto_str = self.IR_graph.SerializeToString() #中间表示序列化为.pb
with open(filename, 'wb') as of:
of.write(proto_str)
print("IR network structure is saved as [{}].".format(filename))
return proto_str
def save_weights(self, filename):
if self.weight_loaded:
with open(filename, 'wb') as of:
np.save(
of, self.nodes_weights
) #weights就是dict(), key:节点名, val:权重dict. 保存到.npy供Emitter进一步转换为特定框架的权重.
print("IR weights are saved as [{}].".format(filename))
else:
print("Warning: weights are not loaded.")
def save_weights_txt(self, filename):
if self.weight_loaded:
with open(filename, 'w') as of:
for name in self.nodes_weights:
of.write(name + ':\n')
for name2 in self.nodes_weights[name]:
if name2 == 'index':
of.write('\t id:{}\n'.format(
self.nodes_weights[name]['index']))
else:
of.write('\t {} shape:{} \n'.format(
name2, self.nodes_weights[name][name2].shape))
print("IR weights are saved as [{}].".format(filename))
else:
print("Warning: weights are not loaded.")
#原名：convert_inedge
#注意这个方法在tensorflow_parser中没有用到.而在tensorflow_parser中定义了新函数_convert_in_nodes实现类似功能.
def convert_in_nodes(self, src_node, IR_node, start_idx=0, end_idx=None):
if end_idx == None:
end_idx = len(src_node.in_nodes_names)
for idx in range(start_idx, end_idx):
#把node的输入关系，转换到中间表示NodeDef的input顶级属性中去.
IR_node.input.append(
self.src_graph.get_node_by_name(
src_node.in_nodes_names[idx]).real_name.lstrip('_'))
@staticmethod
def channel_first_conv_kernel_to_IR(tensor):
dim = tensor.ndim
tensor = np.transpose(tensor, list(range(2, dim)) + [1, 0])
return tensor
@staticmethod
def channel_first_shape_to_IR(shape):
return [shape[0]] + list(shape[2:]) + [shape[1]]
@staticmethod
def channel_first_axis_to_IR(index):
if index == 0:
return 0
elif index == 1:
return -1
else:
return index - 1
def as_graph_def(self):
graph_pb2 = GraphDef()
graph_pb2.version = self.version
graph_pb2.node.extend([node.node_pb2 for node in self.node_dict.values()])
return graph_pb2