-
Notifications
You must be signed in to change notification settings - Fork 23
/
Copy paththexport.py
157 lines (146 loc) · 6.48 KB
/
thexport.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
import torch
from torch import nn
from ctypes import *
from torchvision import models
from torch.autograd import Variable as V
def writetensor(f, name, t):
f.write(bytearray(c_int8(3))) #Tensor id
f.write(str.encode(name)) #Name
f.write(bytearray(c_int8(0))) #Name terminator
f.write(bytearray(c_int32(t.dim()))) #Number of dimensions
for i in range(t.dim()):
f.write(bytearray(c_int32(t.size(i)))) #Individual dimensions
f.flush()
t.contiguous().storage()._write_file(f)
f.flush()
def writeint(f, name, v):
f.write(bytearray(c_int8(4))) #int32param id
f.write(str.encode(name)) #Param name
f.write(bytearray(c_int8(0))) #Param name terminator
f.write(bytearray(c_int32(v))) #Data
def writefloat(f, name, v):
f.write(bytearray(c_int8(5))) #floatparam id
f.write(str.encode(name)) #Param name
f.write(bytearray(c_int8(0))) #Param name terminator
f.write(bytearray(c_float(v))) #Data
def writeintvect(f, name, v):
f.write(bytearray(c_int8(6))) #32tupleparam id
f.write(str.encode(name)) #Param name
f.write(bytearray(c_int8(0))) #Param name terminator
f.write(bytearray(c_int32(len(v)))) #Tuple elements
for i in range(len(v)):
f.write(bytearray(c_int32(v[i])))
def writefunctionid(f, id):
f.write(bytearray(c_int8(7))) #Function id
f.write(bytearray(c_int32(id))) #Data
def check_layer_class(obj):
if (str(obj.__class__)=="<class 'torch.autograd.function.AddmmBackward'>"):
return (True,'torch.nn._functions.linear.Linear')
elif (str(obj.__class__)=="<class 'torch.autograd.function.ThresholdBackward'>"):
return (True,'torch.nn._functions.thnn.auto.Threshold')
elif (str(obj.__class__)=="<class 'torch.autograd.function.DropoutBackward'>"):
return (True,'torch.nn._functions.dropout.Dropout')
elif (str(obj.__class__)=="<class 'torch.autograd.function.ViewBackward'>"):
return (True,'torch.autograd._functions.tensor.View')
elif (str(obj.__class__)=="<class 'torch.autograd.function.MaxPool2dBackward'>"):
return (True,'torch.nn._functions.thnn.pooling.MaxPool2d')
elif (str(obj.__class__)=="<class 'ConvNdBackward'>"):
if obj.transposed:
return (True,'torch.nn._functions.conv.ConvTransposedNd')
else:
return (True,'torch.nn._functions.conv.ConvNd')
elif (str(obj.__class__)=="<class 'torch.autograd.function.AvgPool2dBackward'>"):
return (True,'torch.nn._functions.thnn.pooling.AvgPool2d')
elif (str(obj.__class__)=="<class 'torch.autograd.function.AddBackward'>"):
obj.inplace=True
return (True,'torch.autograd._functions.basic_ops.Add')
elif (str(obj.__class__)=="<class 'BatchNormBackward'>"):
return (True,'torch.nn._functions.batchnorm.BatchNorm')
elif (str(obj.__class__)=="<class 'torch.autograd.function.ConcatBackward'>"):
return (True,'torch.autograd._functions.tensor.Concat')
elif (str(obj.__class__)=="<class 'torch.autograd.function.CmaxBackward'>"):
return (True,'Cmax')
elif (str(obj.__class__)=="<class 'torch.autograd.function.IndexBackward'>"):
return (True,'Slice')
return (False,'')
def check_parameter_class(obj):
if (str(obj.__class__)=="<class 'AccumulateGrad'>"):
return (True,obj.variable)
return (False,'')
def check_if_linear_weight(obj):
if (str(obj.__class__)=="<class 'torch.autograd.function.TransposeBackward'>"):
return (True,obj.next_functions[0][0].variable)
return (False,'')
class Exporter:
def __init__(self, f):
self.f = f
self.output_id = 0
self.objects = {}
#Write 24 bytes header
f.write(str.encode('PyTorch Graph Dump 1.00'))
f.write(bytearray(c_int8(0))) #String terminator
def end(self):
self.f.write(bytearray(c_int8(0))) #End of function id
def input(self):
self.f.write(bytearray(c_int8(1))) #Input id
def function(self, name, obj):
self.f.write(bytearray(c_int8(2))) #Function id
self.objects[obj] = self.output_id
self.f.write(bytearray(c_int32(self.output_id))) #Unique ID of the output of this function
self.output_id = self.output_id + 1
self.f.write(str.encode(name)) #Function name
self.f.write(bytearray(c_int8(0))) #Function name terminator
if hasattr(obj, 'inplace'):
writeint(self.f, 'inplace', obj.inplace)
if hasattr(obj, 'ceil_mode'):
writeint(self.f, 'ceil_mode', obj.ceil_mode)
if hasattr(obj, 'kernel_size'):
writeintvect(self.f, 'kernel_size', obj.kernel_size)
if hasattr(obj, 'new_sizes'):
writeintvect(self.f, 'sizes', obj.new_sizes)
if hasattr(obj, 'stride'):
writeintvect(self.f, 'stride', obj.stride)
if hasattr(obj, 'padding'):
writeintvect(self.f, 'padding', obj.padding)
if hasattr(obj, 'eps'):
writefloat(self.f, 'eps', obj.eps)
#if hasattr(obj, 'threshold'):
# writefloat(self.f, 'threshold', obj.threshold)
#if hasattr(obj, 'value'):
# writefloat(self.f, 'value', obj.value)
if hasattr(obj, 'running_mean'):
writetensor(self.f, 'running_mean', obj.running_mean)
if hasattr(obj, 'running_var'):
writetensor(self.f, 'running_var', obj.running_var)
if hasattr(obj, 'dim'):
writeint(self.f, 'dim', obj.dim)
if hasattr(obj, 'index'):
writeintvect(self.f, 'index', [obj.index[1].start, obj.index[1].stop])
def tensor(self, t):
writetensor(self.f, '', t.data)
def write(self, obj):
self.function(check_layer_class(obj)[1], obj)
for o in obj.next_functions:
if check_layer_class(o[0])[0]:
if o[0] in self.objects:
writefunctionid(self.f, self.objects[o[0]])
else:
self.write(o[0])
if obj.next_functions[0][0] is None:
self.input()
for o in obj.next_functions:
(check,param)=check_if_linear_weight(o[0])
if check:
self.tensor(param)
for o in obj.next_functions:
(check,param)=check_parameter_class(o[0])
if check:
self.tensor(param)
self.end()
def save(path, output):
with open(path, mode='wb') as f:
e = Exporter(f)
e.write(output.grad_fn)
#model=models.densenet201(pretrained=True).eval()
#out=model(V(torch.FloatTensor(1,3,227,227)))
#save('model.net',out)