Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_adam.py 6.31 KB
Newer Older
Deyu Fu's avatar
initial commit to make fused optimizers compatible with AMP  
Deyu Fu committed
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 
import unittest
import os
import random

import torch
import apex

class TestFusedAdam(unittest.TestCase):
    def setUp(self, max_abs_diff=1e-3, max_rel_diff=1, iters=7):
        self.max_abs_diff = max_abs_diff
        self.max_rel_diff = max_rel_diff
        self.iters = iters
        torch.cuda.manual_seed(9876)

    def tearDown(self):
        pass

    def gen_param_optim(self, tensors, adam_option):
        ref_param = []
        tst_param = []
        for tensor in tensors:
            ref_param.append(torch.nn.Parameter(tensor.clone()))
            tst_param.append(torch.nn.Parameter(tensor.clone()))

        ref_optim = torch.optim.Adam(ref_param, **adam_option)
Deyu Fu's avatar
Reverse to Fused* naming, clean up accordingly:  
Deyu Fu committed
26 
        tst_optim = apex.optimizers.FusedAdam(tst_param, **adam_option)
Deyu Fu's avatar
initial commit to make fused optimizers compatible with AMP  
Deyu Fu committed
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 

        return (ref_param, tst_param, ref_optim, tst_optim)

    def gen_grad(self, ref_param, tst_param):
        for p_ref, p_tst in zip(ref_param, tst_param):
            p_ref.grad = torch.rand_like(p_ref)
            p_tst.grad = p_ref.grad

    def gen_mixed_grad(self, ref_param, tst_param, scale=1.0):
        half_grads = []
        for p_ref, p_tst in zip(ref_param, tst_param):
            half_grads.append(torch.rand_like(p_ref).half())
            p_ref.grad = half_grads[-1].float() / scale
        return half_grads

    def get_max_diff(self, ref_param, tst_param):
        max_abs_diff = max_rel_diff = 0
        for p_ref, p_tst in zip(ref_param, tst_param):
            max_abs_diff_p = (p_ref - p_tst).abs().max().item()
            max_rel_diff_p = ((p_ref - p_tst) / p_ref).abs().max().item()

            if max_abs_diff_p > max_abs_diff:  max_abs_diff = max_abs_diff_p
            if max_rel_diff_p > max_rel_diff:  max_rel_diff = max_rel_diff_p

        return max_abs_diff, max_rel_diff

    def gen_single_type_test(self, param_type=torch.float):
        nelem = 278011
        adam_option = {'lr':5e-4, 'betas':(0.9, 0.999), 'eps':1e-08,
            'weight_decay':0, 'amsgrad':False}

        tensor = torch.rand(nelem, dtype=param_type, device='cuda')
        ref_param, tst_param, ref_optim, tst_optim = \
            self.gen_param_optim([tensor], adam_option)

        for i in range(self.iters):
            self.gen_grad(ref_param, tst_param)
            ref_optim.step()
            tst_optim.step()
            max_abs_diff, max_rel_diff = self.get_max_diff(ref_param, tst_param)

            self.assertLessEqual(max_abs_diff, self.max_abs_diff)
            self.assertLessEqual(max_rel_diff, self.max_rel_diff)

    def test_double(self):
        self.gen_single_type_test(param_type=torch.double)

    def test_float(self):
        self.gen_single_type_test(param_type=torch.float)

    def test_half(self):
        self.gen_single_type_test(param_type=torch.float16)
Deyu Fu's avatar
disable breaking test until switch to test against upstream v1.2.0  
Deyu Fu committed
80 
    @unittest.skip('Disable until 8/1/2019 adam/adamw upstream picked')
Deyu Fu's avatar
initial commit to make fused optimizers compatible with AMP  
Deyu Fu committed
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 
    def test_multi_params(self):
        sizes = [[4096, 1024], [4096], [4096, 2048], [32320, 1024], [1]]
        adam_option = {'lr':5e-4, 'betas':(0.9, 0.999), 'eps':1e-08,
            'weight_decay':0, 'amsgrad':False}

        tensors = []
        for size in sizes:
            tensors.append(torch.rand(size, dtype=torch.float, device='cuda'))
        ref_param, tst_param, ref_optim, tst_optim = \
            self.gen_param_optim(tensors, adam_option)

        for i in range(self.iters):
            self.gen_grad(ref_param, tst_param)
            ref_optim.step()
            tst_optim.step()
            max_abs_diff, max_rel_diff = self.get_max_diff(ref_param, tst_param)
            self.assertLessEqual(max_abs_diff, self.max_abs_diff)
            self.assertLessEqual(max_rel_diff, self.max_rel_diff)

    @unittest.skip('No longer support fuse scaling')
    def test_scale(self):
        nelem = 278011
        adam_option = {'lr':5e-4, 'betas':(0.9, 0.999), 'eps':1e-08,
            'weight_decay':0, 'amsgrad':False}

        tensor = torch.rand(nelem, dtype=torch.float, device='cuda')
        ref_param, tst_param, ref_optim, tst_optim = \
            self.gen_param_optim([tensor], adam_option)

        for i in range(self.iters):
            scale = random.random() * 1000
            half_grads = self.gen_mixed_grad(ref_param, tst_param, scale)
            ref_optim.step()
            tst_optim.step(grads=half_grads, scale=scale)
            max_abs_diff, max_rel_diff = self.get_max_diff(ref_param, tst_param)

            self.assertLessEqual(max_abs_diff, self.max_abs_diff)
            self.assertLessEqual(max_rel_diff, self.max_rel_diff)

    @unittest.skip('No longer support output fp16 param')
    def test_fp16_output(self):
        nelem = 278011
        adam_option = {'lr':5e-4, 'betas':(0.9, 0.999), 'eps':1e-08,
            'weight_decay':0, 'amsgrad':False}

        tensor = torch.rand(nelem, dtype=torch.float, device='cuda')
        ref_param, tst_param, ref_optim, tst_optim = \
            self.gen_param_optim([tensor], adam_option)

        fp16_param = torch.nn.Parameter(tensor.clone().half())

        for i in range(self.iters):
            half_grads = self.gen_mixed_grad(ref_param, tst_param)
            ref_optim.step()
            tst_optim.step(grads=half_grads, output_params=[fp16_param])

            max_abs_diff, max_rel_diff = self.get_max_diff(ref_param, tst_param)
            self.assertLessEqual(max_abs_diff, self.max_abs_diff)
            self.assertLessEqual(max_rel_diff, self.max_rel_diff)

            max_abs_diff, max_rel_diff = self.get_max_diff(tst_param, \
                [fp16_param.float()])
            self.assertLessEqual(max_abs_diff, self.max_abs_diff)
            self.assertLessEqual(max_rel_diff, self.max_rel_diff)

    def test_adam_option(self):
        nelem = 1
        adam_option = {'lr':0.01, 'betas':(0.6, 0.9), 'eps':3e-06,
            'weight_decay':0, 'amsgrad':False}

        tensor = torch.rand(nelem, dtype=torch.float, device='cuda')
        ref_param, tst_param, ref_optim, tst_optim = \
            self.gen_param_optim([tensor], adam_option)

        for i in range(self.iters):
            self.gen_grad(ref_param, tst_param)
            ref_optim.step()
            tst_optim.step()
            max_abs_diff, max_rel_diff = self.get_max_diff(ref_param, tst_param)

            self.assertLessEqual(max_abs_diff, self.max_abs_diff)
            self.assertLessEqual(max_rel_diff, self.max_rel_diff)


if __name__ == '__main__':
    script_path = os.path.dirname(os.path.realpath(__file__))
    unittest.main()