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
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
|
# Owner(s): ["oncall: export"]
def random_dag(n: int):
"""
Util to generate a random DAG with n nodes.
The nodes are numbered 0, 1, ..., n-1. The DAG is generated by randomly
choosing a subset of edges from the complete graph on n nodes, such that
for each (i, j) we have i < j.
"""
import random
edges = {}
for i in range(n):
edges[i] = []
for j in range(i + 1, n):
if random.choice([True, False]):
edges[i].append(j)
return edges
class Block:
"""
Util to generate a block of Python-formatted code.
"""
def __init__(self):
self._code = []
def __repr__(self):
return "".join(self._code)
def new_line(self, line: str):
"""
Add a new line of code. The line is automatically suffixed
with a newline character.
"""
self._code.append(line + "\n")
def new_block(self, block: "Block"):
"""
Add a new block of code. All lines in the new block are
automatically prefixed by a tab character.
"""
self._code.extend(" " + line for line in block._code)
class TestGenerator:
"""
Abstract base class for generating test code.
Users should subclass this class and implement the test_name() and
test_body() methods. The test_name() method should return a string
that uniquely identifies the test. The test_body() method should
yield blocks of code.
"""
def __init__(self):
self._count = 0
def _generate_test_name(self):
self._count += 1
return f"{self.test_name()}_{self._count}"
def generate_test(self):
test_name = self._generate_test_name()
code = Block()
code.new_line(f"def {test_name}():")
for block in self.test_body():
code.new_block(block)
code.new_line(f"{test_name}()")
return str(code)
def test_name(self):
raise NotImplementedError
def test_body(self):
raise NotImplementedError
class NNModuleGenerator:
"""
Abstract base class for generating a nn.Module.
Users should subclass this class and implement the gen_init_body() and
gen_forward_body() methods. The gen_init_body() method should return a
block of code that initializes the nn.Module. The gen_forward_body() method
should return a block of code that defines the forward() of the nn.Module.
"""
def gen_init_body(self, i: int):
raise NotImplementedError
def gen_forward_body(self, i: int):
raise NotImplementedError
def gen_nn_module(self, i: int):
def gen_nn_module_body():
code = Block()
code.new_line("def __init__(self):")
code.new_block(self.gen_init_body(i))
code.new_line("def forward(self, x):")
code.new_block(self.gen_forward_body(i))
return code
code = Block()
code.new_line(f"class N{i}(torch.nn.Module):")
code.new_block(gen_nn_module_body())
return code
class Unflatten(TestGenerator):
"""
Generates test that unflattens a model with several nn.Modules that call
each other. The modules are generated by calling the nn_module_generator()
method.
The model is exported and then unflattened. The unflattened model is then
compared against the eager model.
"""
def __init__(self, n: int):
super().__init__()
self.n = n
def nn_module_generator(self):
class GenNNModule(NNModuleGenerator):
def __init__(self, n: int):
super().__init__()
self.n = n
self.calls = random_dag(self.n)
def gen_init_body(self, i: int):
code = Block()
code.new_line("super().__init__()")
if i < self.n - 1:
code.new_line(f"self.n{i+1} = N{i+1}()")
return code
def gen_forward_body(self, i: int):
def path(i, j):
if i + 1 == j:
return f"n{j}"
else:
return f"n{i + 1}.{path(i + 1, j)}"
code = Block()
for j in self.calls[i]:
code.new_line(f"x = self.{path(i, j)}(x + 1)")
code.new_line("return x + 1")
return code
return GenNNModule(self.n)
def test_name(self):
return f"{self.__class__.__name__}_{self.n}"
def test_body(self):
def path(i, j):
if i + 1 == j:
return f"n{j}"
else:
return f"n{i + 1}.{path(i + 1, j)}"
nn_module_generator = self.nn_module_generator()
for i in range(self.n):
yield nn_module_generator.gen_nn_module(self.n - 1 - i)
fqns = "".join(f"'{path(0, j)},'" for j in range(1, self.n))
def gen_main():
code = Block()
code.new_line("inp = (torch.ones(1),)")
code.new_line("eager = N0()(*inp)")
code.new_line(
f"ep = torch.export.export(N0(), inp, strict=False, preserve_module_call_signature=({fqns}))"
)
code.new_line("epm = ep.module()")
code.new_line("ufm = torch.export.unflatten(ep)")
code.new_line("assert torch.allclose(epm(*inp), eager)")
code.new_line("assert torch.allclose(ufm(*inp), eager)")
return code
yield gen_main()
class ConstantUnflatten(Unflatten):
"""
Generates test that unflattens a model with several nn.Modules that call
each other and access constants. The modules are generated by calling the
nn_module_generator() method.
"""
def nn_module_generator(self):
class GenNNModule(NNModuleGenerator):
def __init__(self, n):
super().__init__()
self.n = n
self.accesses = random_dag(self.n)
self.calls = random_dag(self.n)
def gen_init_body(self, i: int):
code = Block()
code.new_line("super().__init__()")
code.new_line("self.const = torch.ones(1)")
if i < self.n - 1:
code.new_line(f"self.n{i+1} = N{i+1}()")
return code
def gen_forward_body(self, i: int):
def path(i, j):
if i + 1 == j:
return f"n{j}"
else:
return f"n{i + 1}.{path(i + 1, j)}"
code = Block()
for j in self.accesses[i]:
code.new_line(f"x = x + self.{path(i, j)}.const")
for j in self.calls[i]:
code.new_line(f"x = self.{path(i, j)}(x + 1)")
code.new_line("return x + 1")
return code
return GenNNModule(self.n)
class BufferUnflatten(Unflatten):
"""
Generates test that unflattens a model with several nn.Modules that call
each other and access and mutate buffers. The modules are generated by
calling the nn_module_generator() method.
"""
def nn_module_generator(self):
class GenNNModule(NNModuleGenerator):
def __init__(self, n):
super().__init__()
self.n = n
self.accesses = random_dag(self.n)
self.mutations = random_dag(self.n)
self.calls = random_dag(self.n)
def gen_init_body(self, i: int):
code = Block()
code.new_line("super().__init__()")
code.new_line("self.buf = torch.nn.Buffer(torch.ones(1))")
if i < self.n - 1:
code.new_line(f"self.n{i+1} = N{i+1}()")
return code
def gen_forward_body(self, i: int):
def path(i, j):
if i + 1 == j:
return f"n{j}"
else:
return f"n{i + 1}.{path(i + 1, j)}"
code = Block()
for j in self.accesses[i]:
code.new_line(f"x = x + self.{path(i, j)}.buf")
for j in self.calls[i]:
code.new_line(f"x = self.{path(i, j)}(x + 1)")
for j in self.mutations[i]:
code.new_line(f"self.{path(i, j)}.buf.add_(1)")
code.new_line("return x + 1")
return code
return GenNNModule(self.n)
|
