-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathacc_benchmark.py
201 lines (169 loc) · 7.73 KB
/
acc_benchmark.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
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
import torch
from grid_sample1d import GridSample1d
import torch.nn.functional as F
from tqdm import tqdm
import numpy as np
import random
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
args_groups = [
{'original': {'padding_mode': 'zeros', 'align_corners': True},
'mine': {'padding_mode': False, 'align_corners': True}},
{'original': {'padding_mode': 'zeros', 'align_corners': False},
'mine': {'padding_mode': False, 'align_corners': False}},
{'original': {'padding_mode': 'border', 'align_corners': True},
'mine': {'padding_mode': True, 'align_corners': True}},
{'original': {'padding_mode': 'border', 'align_corners': False},
'mine': {'padding_mode': True, 'align_corners': False}}
]
def original(input, grid, padding_mode, align_corners):
shape = grid.shape
grid = grid.sin() # batch_size * L_out
input = input.unsqueeze(-1) # batch_size * C * L_in * 1
# grid = grid.unsqueeze(-1) # batch_size * L_out * 1
grid = grid.unsqueeze(1) # batch_size * 1 * L_out
grid = torch.stack([-torch.ones_like(grid), grid], dim=-1)
z = F.grid_sample(input, grid, padding_mode=padding_mode, align_corners=align_corners)
C = input.shape[1]
out_shape = [shape[0], C, shape[1]]
z = z.view(*out_shape) # batch_size * C * L_out
return z
def mine(input, grid, module):
shape = grid.shape
grid = grid.sin()
z = module(input, grid)
C = input.shape[1]
out_shape = [shape[0], C, shape[1]]
z = z.view(*out_shape)
return z
def inspect(output, output_origin, verbose_matrix=False, verbose=False):
err = torch.abs(output - output_origin)
max_err = torch.max(err).item()
pos = torch.argmax(err)
rela_err = err / torch.abs(output_origin)
max_err_rela = torch.max(rela_err)
pos_rela = torch.argmax(rela_err)
N_err = torch.sum(err > eps).item()
N_rela_err = torch.sum(rela_err > eps_r).item()
# if max_err > eps:
# if verbose_matrix:
# print('output')
# print(output)
# print('origin')
# print(output_origin)
# print(output - output_origin)
# print('different!')
# print(f'max_err={max_err}')
# print(f'where origin={output_origin.view(-1)[pos]}')
# print(f'mine={output.view(-1)[pos]}')
# print(f'N err > eps={N_err}')
# print(f'err% = {N_err / torch.numel(output) * 100:.2f}')
# print('-' * 50)
if max_err_rela > eps_r:
if verbose:
if verbose_matrix:
print('output')
print(output)
print('origin')
print(output_origin)
print(output - output_origin)
print('different!')
print(f'max_err_rela={max_err_rela}')
print(f'where origin={output_origin.view(-1)[pos_rela]}')
print(f'mine={output.view(-1)[pos_rela]}')
print(f'N err > eps={N_err}')
print(f'err% = {N_rela_err / torch.numel(output) * 100:.2f}')
# if N_err == 0:
# print('same!')
return N_rela_err
if __name__ == '__main__':
setup_seed(0)
batch_size = 20
C = 256
L_in = 16
L_out = 32
eps = 1e-6
eps_r = 1e-5
N_samples = 100
print('forward')
for args in args_groups:
print('testing')
print(args)
module = GridSample1d(**args['mine'])
running_err_forward = 0.
running_err_backward_input = 0.
running_err_backward_grid = 0.
try:
with torch.no_grad():
for i in tqdm(range(N_samples)):
input = torch.randn((batch_size, C, L_in)).cuda()
grid = torch.randn(batch_size, L_out).cuda()
output = mine(input, grid, module).cpu()
output_origin = original(input, grid, **args['original']).cpu()
try:
if (not args['mine']['padding_mode']) and (not args['mine']['align_corners']):
torch.allclose(output, output_origin * 2, atol=eps, rtol=eps_r)
else:
assert torch.allclose(output, output_origin, atol=eps, rtol=eps_r)
except:
N_err = inspect(output, output_origin)
running_err_forward += N_err / torch.numel(output)
if N_err / torch.numel(output) >= 0.05:
raise
else:
pass
print(f'Forward ACC test done on {N_samples} samples with eps={eps}')
print('backward')
for i in tqdm(range(N_samples)):
setup_seed(i)
grid_original = torch.randn((batch_size, L_out), requires_grad=True).cuda()
input_original = torch.randn((batch_size, C, L_in), requires_grad=True).cuda()
grid_original.retain_grad()
input_original.retain_grad()
setup_seed(i)
grid_mine = torch.randn((batch_size, L_out), requires_grad=True).cuda()
input_mine = torch.randn((batch_size, C, L_in), requires_grad=True).cuda()
grid_mine.retain_grad()
input_mine.retain_grad()
output_origin = original(input_original, grid_original, **args['original'])
output = mine(input_mine, grid_mine, module)
if (not args['mine']['padding_mode']) and (not args['mine']['align_corners']):
assert torch.allclose(output, output_origin*2, atol=eps, rtol=eps_r)
else:
assert torch.allclose(output, output_origin, atol=eps, rtol=eps_r)
output_origin = torch.sum(output_origin.view(-1))
output = torch.sum(output.view(-1))
output.backward()
output_origin.backward()
grad_grid_original = grid_original.grad
grad_input_original = input_original.grad
grad_grid_mine = grid_mine.grad
grad_input_mine = input_mine.grad
try:
if (not args['mine']['padding_mode']) and (not args['mine']['align_corners']):
assert torch.allclose(2*grad_grid_original, grad_grid_mine, atol=eps, rtol=eps_r)
assert torch.allclose(2*grad_input_original, grad_input_mine, atol=eps, rtol=eps_r)
else:
assert torch.allclose(grad_grid_original, grad_grid_mine, atol=eps, rtol=eps_r)
assert torch.allclose(grad_input_original, grad_input_mine, atol=eps, rtol=eps_r)
except AssertionError:
N_err_grid = inspect(grad_grid_mine, grad_grid_original,verbose=True)
N_err_input = inspect(grad_input_mine, grad_input_original,verbose=True)
running_err_backward_grid += N_err_grid / torch.numel(grad_grid_mine)
running_err_backward_input += N_err_input / torch.numel(grad_input_mine)
if N_err_grid / torch.numel(grad_grid_mine) >= 0.05 or N_err_input / torch.numel(
grad_input_mine) >= 0.05:
raise
else:
pass
print(f'Backward ACC test done on {N_samples} samples with eps={eps}')
print(f'running err forward:{running_err_forward * 100:.2f}%')
print(f'running err backward input:{running_err_backward_input:.2f}%')
print(f'running err backward grid:{running_err_backward_grid:.2f}%')
except AssertionError:
raise
print('Done')