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
|
import sys
import argparse
import unittest
import contextlib
from itertools import product
from copy import deepcopy
import torch
import torch.cuda
from torch.autograd import Variable, Function
torch.set_default_tensor_type('torch.DoubleTensor')
def run_tests():
parser = argparse.ArgumentParser(add_help=False)
parser.add_argument('--seed', type=int, default=123)
args, remaining = parser.parse_known_args()
torch.manual_seed(args.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(args.seed)
remaining = [sys.argv[0]] + remaining
unittest.main(argv=remaining)
TEST_NUMPY = True
try:
import numpy
except ImportError:
TEST_NUMPY = False
def get_cpu_type(t):
assert t.__module__ == 'torch.cuda'
return getattr(torch, t.__class__.__name__)
def get_gpu_type(t):
assert t.__module__ == 'torch'
return getattr(torch.cuda, t.__name__)
def to_gpu(obj, type_map={}):
if torch.is_tensor(obj):
t = type_map.get(type(obj), get_gpu_type(type(obj)))
return obj.clone().type(t)
elif torch.is_storage(obj):
return obj.new().resize_(obj.size()).copy_(obj)
elif isinstance(obj, Variable):
assert obj.creator is None
t = type_map.get(type(obj.data), get_gpu_type(type(obj.data)))
return Variable(obj.data.clone().type(t), requires_grad=obj.requires_grad)
elif isinstance(obj, list):
return [to_gpu(o, type_map) for o in obj]
elif isinstance(obj, tuple):
return tuple(to_gpu(o, type_map) for o in obj)
else:
return deepcopy(obj)
@contextlib.contextmanager
def freeze_rng_state():
rng_state = torch.get_rng_state()
if torch.cuda.is_available():
cuda_rng_state = torch.cuda.get_rng_state()
yield
if torch.cuda.is_available():
torch.cuda.set_rng_state(cuda_rng_state)
torch.set_rng_state(rng_state)
def iter_indices(tensor):
if tensor.dim() == 0:
return range(0)
if tensor.dim() == 1:
return range(tensor.size(0))
return product(*(range(s) for s in tensor.size()))
def is_iterable(obj):
try:
iter(obj)
return True
except:
return False
class TestCase(unittest.TestCase):
precision = 1e-5
def assertEqual(self, x, y, prec=None, message=''):
if prec is None:
prec = self.precision
if isinstance(x, Variable) and isinstance(y, Variable):
x = x.data
y = y.data
if torch.is_tensor(x) and torch.is_tensor(y):
max_err = 0
super(TestCase, self).assertEqual(x.size(), y.size())
for index in iter_indices(x):
max_err = max(max_err, abs(x[index] - y[index]))
self.assertLessEqual(max_err, prec, message)
elif type(x) == str and type(y) == str:
super(TestCase, self).assertEqual(x, y)
elif is_iterable(x) and is_iterable(y):
for x_, y_ in zip(x, y):
self.assertEqual(x_, y_, prec, message)
else:
try:
self.assertLessEqual(abs(x - y), prec, message)
return
except:
pass
super(TestCase, self).assertEqual(x, y, message)
def assertNotEqual(self, x, y, prec=None, message=''):
if prec is None:
prec = self.precision
if isinstance(x, Variable) and isinstance(y, Variable):
x = x.data
y = y.data
if torch.is_tensor(x) and torch.is_tensor(y):
max_err = 0
if x.size() != y.size():
super(TestCase, self).assertNotEqual(x.size(), y.size())
for index in iter_indices(x):
max_err = max(max_err, abs(x[index] - y[index]))
self.assertGreaterEqual(max_err, prec, message)
elif type(x) == str and type(y) == str:
super(TestCase, self).assertNotEqual(x, y)
elif is_iterable(x) and is_iterable(y):
super(TestCase, self).assertNotEqual(x, y)
else:
try:
self.assertGreaterEqual(abs(x - y), prec, message)
return
except:
pass
super(TestCase, self).assertNotEqual(x, y, message)
def assertObjectIn(self, obj, iterable):
for elem in iterable:
if id(obj) == id(elem):
return
raise AssertionError("object not found in iterable")
def make_jacobian(input, num_out):
if isinstance(input, Variable) and not input.requires_grad:
return None
if torch.is_tensor(input) or isinstance(input, Variable):
return torch.zeros(input.nelement(), num_out)
else:
return type(input)(filter(lambda x: x is not None,
(make_jacobian(elem, num_out) for elem in input)))
def iter_tensors(x, only_requiring_grad=False):
if torch.is_tensor(x):
yield x
elif isinstance(x, Variable):
if x.requires_grad or not only_requiring_grad:
yield x.data
else:
for elem in x:
for result in iter_tensors(elem, only_requiring_grad):
yield result
def contiguous(input):
if torch.is_tensor(input):
return input.contiguous()
elif isinstance(input, Variable):
return input.contiguous()
else:
return type(input)(contiguous(e) for e in input)
def get_numerical_jacobian(fn, input, target):
perturbation = 1e-6
# To be able to use .view(-1) input must be contiguous
input = contiguous(input)
output_size = fn(input).numel()
jacobian = make_jacobian(target, output_size)
# It's much easier to iterate over flattened lists of tensors.
# These are reference to the same objects in jacobian, so any changes
# will be reflected in it as well.
x_tensors = [t for t in iter_tensors(target, True)]
j_tensors = [t for t in iter_tensors(jacobian)]
outa = torch.DoubleTensor(output_size)
outb = torch.DoubleTensor(output_size)
# TODO: compare structure
for x_tensor, d_tensor in zip(x_tensors, j_tensors):
flat_tensor = x_tensor.view(-1)
for i in range(flat_tensor.nelement()):
orig = flat_tensor[i]
flat_tensor[i] = orig - perturbation
outa.copy_(fn(input))
flat_tensor[i] = orig + perturbation
outb.copy_(fn(input))
flat_tensor[i] = orig
outb.add_(-1, outa).div_(2 * perturbation)
d_tensor[i] = outb
return jacobian
|
