File size: 4,383 Bytes
079c32c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import time
import torch
from hpc_rll.origin.upgo import upgo_loss
from hpc_rll.rl_utils.upgo import UPGO
from testbase import mean_relative_error, times

assert torch.cuda.is_available()
use_cuda = True

T = 256
B = 256
N = 256


def upgo_val():
    ori_target_output = torch.randn(T, B, N)
    ori_rhos = torch.randn(T, B)
    ori_action = torch.randint(
        0, N, size=(
            T,
            B,
        )
    )
    ori_rewards = torch.randn(T, B)
    ori_bootstrap_values = torch.randn(T + 1, B)

    hpc_target_output = ori_target_output.clone().detach()
    hpc_rhos = ori_rhos.clone().detach()
    hpc_action = ori_action.clone().detach()
    hpc_rewards = ori_rewards.clone().detach()
    hpc_bootstrap_values = ori_bootstrap_values.clone().detach()
    hpc_upgo = UPGO(T, B, N)

    if use_cuda:
        ori_target_output = ori_target_output.cuda()
        ori_rhos = ori_rhos.cuda()
        ori_action = ori_action.cuda()
        ori_rewards = ori_rewards.cuda()
        ori_bootstrap_values = ori_bootstrap_values.cuda()

        hpc_target_output = hpc_target_output.cuda()
        hpc_rhos = hpc_rhos.cuda()
        hpc_action = hpc_action.cuda()
        hpc_rewards = hpc_rewards.cuda()
        hpc_bootstrap_values = hpc_bootstrap_values.cuda()
        hpc_upgo = hpc_upgo.cuda()

    ori_target_output.requires_grad_(True)
    ori_loss = upgo_loss(ori_target_output, ori_rhos, ori_action, ori_rewards, ori_bootstrap_values)
    ori_loss = ori_loss.mean()
    ori_loss.backward()
    if use_cuda:
        torch.cuda.synchronize()

    hpc_target_output.requires_grad_(True)
    hpc_loss = hpc_upgo(hpc_target_output, hpc_rhos, hpc_action, hpc_rewards, hpc_bootstrap_values)
    hpc_loss = hpc_loss.mean()
    hpc_loss.backward()
    if use_cuda:
        torch.cuda.synchronize()

    mre = mean_relative_error(
        torch.flatten(ori_loss).cpu().detach().numpy(),
        torch.flatten(hpc_loss).cpu().detach().numpy()
    )
    print("upgo fp mean_relative_error: " + str(mre))
    mre = mean_relative_error(
        torch.flatten(ori_target_output.grad).cpu().detach().numpy(),
        torch.flatten(hpc_target_output.grad).cpu().detach().numpy()
    )
    print("upgo bp mean_relative_error: " + str(mre))


def upgo_perf():
    ori_target_output = torch.randn(T, B, N)
    ori_rhos = torch.randn(T, B)
    ori_action = torch.randint(
        0, N, size=(
            T,
            B,
        )
    )
    ori_rewards = torch.randn(T, B)
    ori_bootstrap_values = torch.randn(T + 1, B)

    hpc_target_output = ori_target_output.clone().detach()
    hpc_rhos = ori_rhos.clone().detach()
    hpc_action = ori_action.clone().detach()
    hpc_rewards = ori_rewards.clone().detach()
    hpc_bootstrap_values = ori_bootstrap_values.clone().detach()
    hpc_upgo = UPGO(T, B, N)

    if use_cuda:
        ori_target_output = ori_target_output.cuda()
        ori_rhos = ori_rhos.cuda()
        ori_action = ori_action.cuda()
        ori_rewards = ori_rewards.cuda()
        ori_bootstrap_values = ori_bootstrap_values.cuda()

        hpc_target_output = hpc_target_output.cuda()
        hpc_rhos = hpc_rhos.cuda()
        hpc_action = hpc_action.cuda()
        hpc_rewards = hpc_rewards.cuda()
        hpc_bootstrap_values = hpc_bootstrap_values.cuda()
        hpc_upgo = hpc_upgo.cuda()

    ori_target_output.requires_grad_(True)
    for i in range(times):
        t = time.time()
        ori_loss = upgo_loss(ori_target_output, ori_rhos, ori_action, ori_rewards, ori_bootstrap_values)
        ori_loss = ori_loss.mean()
        ori_loss.backward()
        if use_cuda:
            torch.cuda.synchronize()
        print('epoch: {}, original upgo cost time: {}'.format(i, time.time() - t))

    hpc_target_output.requires_grad_(True)
    for i in range(times):
        t = time.time()
        hpc_loss = hpc_upgo(hpc_target_output, hpc_rhos, hpc_action, hpc_rewards, hpc_bootstrap_values)
        hpc_loss = hpc_loss.mean()
        hpc_loss.backward()
        if use_cuda:
            torch.cuda.synchronize()
        print('epoch: {}, hpc upgo cost time: {}'.format(i, time.time() - t))


if __name__ == '__main__':
    print("target problem: T = {}, B = {}, N = {}".format(T, B, N))
    print("================run upgo validation test================")
    upgo_val()
    print("================run upgo performance test================")
    upgo_perf()