torch.backends.cudnn.benchmark=True 的效果和坑

系统环境
CUDA 11.1
gtx1070-8g
win10 21h1 19043.1266

打开该选项
torch.backends.cudnn.benchmark=True
加快运算速度，但相比加速。。。。。。
最为直观的影响
计算误差会增大约1000倍。
显存占用会变为关闭时的 1.1 - 1.5 倍。

这个选项不能随便打开，影响很大。

以下为实验
cudnn.benchmark = False，显存占用为 3.0 G

card_id: 0 elapsed/total: 1/10 max_diff_1: 0.00003290 max_diff_1: 0.00002837
card_id: 0 elapsed/total: 2/10 max_diff_1: 0.00003082 max_diff_1: 0.00002468
card_id: 0 elapsed/total: 3/10 max_diff_1: 0.00003231 max_diff_1: 0.00002813
card_id: 0 elapsed/total: 4/10 max_diff_1: 0.00003111 max_diff_1: 0.00002661
card_id: 0 elapsed/total: 5/10 max_diff_1: 0.00003064 max_diff_1: 0.00002551
card_id: 0 elapsed/total: 6/10 max_diff_1: 0.00003046 max_diff_1: 0.00002813
card_id: 0 elapsed/total: 7/10 max_diff_1: 0.00003266 max_diff_1: 0.00002570
card_id: 0 elapsed/total: 8/10 max_diff_1: 0.00003067 max_diff_1: 0.00002459
card_id: 0 elapsed/total: 9/10 max_diff_1: 0.00003183 max_diff_1: 0.00002503
card_id: 0 elapsed/total: 10/10 max_diff_1: 0.00003159 max_diff_1: 0.00002563

cudnn.benchmark = True，显存占用为 4.0 G

card_id: 0 elapsed/total: 1/10 max_diff_1: 0.12276840 max_diff_1: 0.09905499
card_id: 0 elapsed/total: 2/10 max_diff_1: 0.10842407 max_diff_1: 0.08467466
card_id: 0 elapsed/total: 3/10 max_diff_1: 0.11710882 max_diff_1: 0.08327144
card_id: 0 elapsed/total: 4/10 max_diff_1: 0.11344403 max_diff_1: 0.08150017
card_id: 0 elapsed/total: 5/10 max_diff_1: 0.10728776 max_diff_1: 0.08666122
card_id: 0 elapsed/total: 6/10 max_diff_1: 0.12120318 max_diff_1: 0.08831707
card_id: 0 elapsed/total: 7/10 max_diff_1: 0.12990451 max_diff_1: 0.08046508
card_id: 0 elapsed/total: 8/10 max_diff_1: 0.11482769 max_diff_1: 0.08473527
card_id: 0 elapsed/total: 9/10 max_diff_1: 0.11562169 max_diff_1: 0.08603296
card_id: 0 elapsed/total: 10/10 max_diff_1: 0.11495388 max_diff_1: 0.08923209

以下为实验代码，你可以通过修改第100行的代码来观察 cudnn.benchmark 的影响

'''
用于检测cuda运算错误
'''
import os
import torch
import torch.nn as nn
from torch.backends import cudnn
import argparse
import time
import platform


def ConvBnAct(in_ch, out_ch, ker_sz, stride, pad, act=nn.Identity(), group=1, dilation=1):
    return nn.Sequential(nn.Conv2d(in_ch, out_ch, ker_sz, stride, pad, groups=group, bias=False, dilation=dilation),
                         nn.GroupNorm(16, out_ch, eps=1e-8),
                         act)


def DeConvBnAct(in_ch, out_ch, ker_sz, stride, pad, act=nn.Identity(), group=1, dilation=1):
    return nn.Sequential(nn.ConvTranspose2d(in_ch, out_ch, ker_sz, stride, pad, groups=group, bias=False, dilation=dilation),
                         nn.GroupNorm(16, out_ch, eps=1e-8),
                         act)


class RevSequential(nn.ModuleList):
    '''
    功能大部分与ModuleList重叠
    '''
    def __init__(self, modules=None):
        super().__init__(modules)

    def append(self, module):
        assert hasattr(module, 'invert') and callable(module.invert)
        super().append(module)

    def extend(self, modules):
        for m in modules:
            self.append(m)

    def forward(self, x1, x2):
        y1, y2 = x1, x2
        for m in self:
            y1, y2 = m(y1, y2)
        return y1, y2

    def invert(self, y1, y2):
        x1, x2 = y1, y2
        for m in list(self)[::-1]:
            x1, x2 = m.invert(x1, x2)
        return x1, x2


class RevGroupBlock(RevSequential):
    '''
    当前只支持输入通道等于输出通道，并且不允许下采样
    '''
    def __init__(self, in_ch, out_ch, stride, act, block_type, blocks, **kwargs):
        assert in_ch == out_ch
        assert stride == 1
        mods = []
        for _ in range(blocks):
            mods.append(block_type(in_ch=in_ch, out_ch=out_ch, stride=1, act=act, **kwargs))
        # self.extend(mods)
        super().__init__(mods)


class RevBlockC(nn.Module):
    def __init__(self, in_ch, out_ch, stride, act, **kwargs):
        super().__init__()
        inter_ch = in_ch // 2
        self.conv1 = ConvBnAct(in_ch, inter_ch, ker_sz=5, stride=1, pad=2, act=act)
        self.conv2 = ConvBnAct(inter_ch, inter_ch, ker_sz=5, stride=1, pad=2, act=act, group=inter_ch)
        self.conv3 = ConvBnAct(in_ch, in_ch, ker_sz=1, stride=1, pad=0, act=nn.Identity())

    def func(self, x):
        y1 = self.conv1(x)
        y2 = self.conv2(y1)
        y = torch.cat([y1, y2], dim=1)
        y = self.conv3(y)
        return y

    def forward(self, x1, x2):
        y = x1 + self.func(x2)
        return x2, y

    def invert(self, y1, y2):
        x2, y = y1, y2
        x1 = y - self.func(x2)
        return x1, x2


def new_model():
    act = nn.ELU()
    rvb = RevGroupBlock(128, 128, 1, act, RevBlockC, 12).to(device)
    rvb.eval()
    return rvb


if __name__ == '__main__':
    cudnn.benchmark = True
    cudnn.deterministic = True
    torch.set_grad_enabled(False)

    parse = argparse.ArgumentParser(description='Used to check pytorch cudnn.benchmark flag.')
    parse.add_argument('-i', type=int, help='Card id. Which cuda card do you want to test. default: 0', default=0)
    parse.add_argument('-e', type=int, help='Epoch. defaule: 10', default=10)
    parse.add_argument('-bp', type=bool, help='Use backward. defaule: True', default=True)
    parse.add_argument('-bs', type=int, help='Batch size. defaule: 8', default=8)
    parse = parse.parse_args()

    card_id = parse.i
    epoch = parse.e
    use_backward = parse.bp
    batch_size = parse.bs

    # 使用cpu测试理论上是永远不会报错的
    device = 'cpu' if card_id == -1 else f'cuda:{card_id}'
    device = torch.device(device)
    assert epoch > 0
    assert batch_size > 0

    rvb = new_model()

    torch.set_grad_enabled(use_backward)

    for e in range(epoch):
        e = e+1

        a1 = torch.randn(batch_size, 128, 64, 64, device=device)
        b1, b2 = rvb(a1, a1)
        o_a1, o_a2 = rvb.invert(b1, b2)

        if use_backward:
            (o_a1.max() + o_a2.max()).backward()

        with torch.no_grad():
            max_diff_1 = torch.abs(o_a1 - o_a2).max().item()
            max_diff_2 = torch.abs(a1 - o_a1).max().item()

        line = f'card_id: {card_id} elapsed/total: {e}/{epoch} max_diff_1: {max_diff_1:.8f} max_diff_1: {max_diff_2:.8f}'
        print(line)