前言

当深度学习模型完成训练开始部署、推理阶段，模型的推理速度、性能往往受到关注。目前主流DL framework都有各自的性能分析工具，本文主要介绍PyTorch 的性能分析工具——torch.autograd.profiler

测试环境

ubuntu 18.04
anaconda3 + python 3.7
NVIDIA GPU/CUDA 10.2 (可选）
PyTorch 1.6

Profiler 性能分析工具介绍

Profiler 一般指性能分析工具，用于分析APP、模型的执行时间，执行流程，内存消耗等。除了Pytorch,Tensorflow 这样的深度学习框架，像NVIDIA CUDA， AMD ROCm 等也提供了各自的Profiler性能分析工具，比如 nvprof, rocprofiler。

PyTorch Profiler工具

pytroch Profiler位于torch.autograd.profiler, 目前支持的功能：

CPU/GPU 端Op执行时间统计
CPU/GPU 端Op输入Tensor的维度分析
Op的内存消耗统计

PyTorch 官网关于Profiler的介绍
https://pytorch.org/docs/master/autograd.html

image.png

Profiler分析CPU、GPU端Op执行时间

torch.autograd.profiler.profile(use_cuda=False...)

CPU Only: 设置use_cuda=False
GPU 模式：设置use_cuda=True, 注意：模型以及输入Tensor 需要事先导入显存

CPU Only 模式

import os
import numpy as np
import torch
from torchvision.models import resnet18
import time


if __name__ == '__main__':
    model = resnet18(pretrained=False)
    device = torch.device('cpu')
    model.eval()
    model.to(device)
    dump_input = torch.ones(1,3,224,224).to(device)

    # Warn-up
    for _ in range(5):
        start = time.time()
        outputs = model(dump_input)
        torch.cuda.synchronize()
        end = time.time()
        print('Time:{}ms'.format((end-start)*1000))

    with torch.autograd.profiler.profile(enabled=True, use_cuda=False, record_shapes=False, profile_memory=False) as prof:
        outputs = model(dump_input)
    print(prof.table())

profiler输出：（CPU Only)

image.png

GPU 模式

import os
import numpy as np
import torch
from torchvision.models import resnet18
import time


if __name__ == '__main__':
    model = resnet18(pretrained=False)
    device = torch.device('cuda')
    model.eval()
    model.to(device)
    dump_input = torch.ones(1,3,224,224).to(device)

    # Warn-up
    for _ in range(5):
        start = time.time()
        outputs = model(dump_input)
        torch.cuda.synchronize()
        end = time.time()
        print('Time:{}ms'.format((end-start)*1000))

    with torch.autograd.profiler.profile(enabled=True, use_cuda=True, record_shapes=False, profile_memory=False) as prof:
        outputs = model(dump_input)
    print(prof.table())

profiler输出：（GPU)

image.png

使用Chrome trace可视化Profiler结果

上面的例子中，profiler的结果直接输出到终端，为了更进一步分析模型Op的执行关系，pytroch profiler支持生成 chrome trace json格式的输出，然后采用chrome 浏览器可视化结果：
只需要在上面的代码最后，加上 prof.export_chrome_trace('./resnet_profile.json')

import os
import numpy as np
import torch
from torchvision.models import resnet18
import time

# def process_event(profiler_events):


if __name__ == '__main__':
    model = resnet18(pretrained=False)
    device = torch.device('cuda')
    model.eval()
    model.to(device)
    dump_input = torch.ones(1,3,224,224).to(device)

    # Warn-up
    for _ in range(5):
        start = time.time()
        outputs = model(dump_input)
        torch.cuda.synchronize()
        end = time.time()
        print('Time:{}ms'.format((end-start)*1000))

    with torch.autograd.profiler.profile(enabled=True, use_cuda=True, record_shapes=False, profile_memory=False) as prof:
        outputs = model(dump_input)
    print(prof.table())
    prof.export_chrome_trace('./resnet_profile.json')

生成的JSON 文件

image.png

打开Chrome浏览器，在地址栏输入 chrome://tracing

image.png

导入profiler生成的JSON文件：
操作：
按键盘w, a, s, d键，可以对profiler的结果进行缩放和移动

image.png

Profiler 结果分析

上面内容主要是pytorch profiler的用法，我们更关心的是如何分析profiler的数据， 如何通过profiler发现模型的性能瓶颈，得出结论

模型整体分析

CPU 端的OP List

image.png
GPU 端的OP List

image.png

CPU 和 GPU Op的关系

CNN/RNN/GAN/Transformer 等模型最终都是由许多Op组成的，在采用GPU设备的情况下，首先CPU端负责Op的调度(schedule),将Op的运算发送到GPU, GPU负责Op的具体运算。笔者略微了解CUDA编程知识，在CUDA编程中， host(cpu)端调用GPU kernel function, GPU kernel启动之后，CPU与GPU异步执行。

image.png

Op的wall_duration_time, self_time 区别

wall_duration_time: 此Op的总共执行时间
self_time: 此Op自身的执行时间，不包含调用其他子Op的执行时间

以relu_ Op为例：（relu_ 是in-place ReLU)
调用关系： relu_ ---->threshold_

threshold_: wall_dur=154.624us
relu_: wall_dur=179us, 由于relu_ Op又调用了threshold_ Op,因此relu_的self_time = 179 - 154 = 25us

relu_ op:

image.png

threshold_ op:

image.png

Op Tensor数据维度分析

PyTorch profiler提供了Op 输入维度

image.png

DNN Model Training Profile

代码示例: MNIST训练

from __future__ import print_function
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
from torch.optim.lr_scheduler import StepLR
import torch.cuda.nvtx as nvtx
from torch.profiler import profile, ProfilerActivity


class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 32, 3, 1)
        self.conv2 = nn.Conv2d(32, 64, 3, 1)
        self.dropout1 = nn.Dropout(0.25)
        self.dropout2 = nn.Dropout(0.5)
        self.fc1 = nn.Linear(9216, 128)
        self.fc2 = nn.Linear(128, 10)

    def forward(self, x):
        x = self.conv1(x)
        x = F.relu(x)
        x = self.conv2(x)
        x = F.relu(x)
        x = F.max_pool2d(x, 2)
        x = self.dropout1(x)
        x = torch.flatten(x, 1)
        x = self.fc1(x)
        x = F.relu(x)
        x = self.dropout2(x)
        x = self.fc2(x)
        output = F.log_softmax(x, dim=1)
        return output


prof = torch.profiler.profile(
        schedule=torch.profiler.schedule(wait=1, warmup=1, active=10, repeat=1),
        on_trace_ready=torch.profiler.tensorboard_trace_handler('./profile'),
        record_shapes=True,
        with_stack=False)

def train(args, model, device, train_loader, optimizer, epoch):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        data, target = data.to(device), target.to(device)
        optimizer.zero_grad()

        prof.start()
        
        nvtx.range_push('Forward')
        output = model(data)
        loss = F.nll_loss(output, target)
        nvtx.range_pop()

        nvtx.range_push('Backward')
        loss.backward()
        optimizer.step()
        nvtx.range_pop()

        prof.step()
        prof.stop()

        if batch_idx % args.log_interval == 0:
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                epoch, batch_idx * len(data), len(train_loader.dataset),
                100. * batch_idx / len(train_loader), loss.item()))
            if args.dry_run:
                break


def test(model, device, test_loader):
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for data, target in test_loader:
            data, target = data.to(device), target.to(device)
            output = model(data)
            test_loss += F.nll_loss(output, target, reduction='sum').item()  # sum up batch loss
            pred = output.argmax(dim=1, keepdim=True)  # get the index of the max log-probability
            correct += pred.eq(target.view_as(pred)).sum().item()

    test_loss /= len(test_loader.dataset)

    print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
        test_loss, correct, len(test_loader.dataset),
        100. * correct / len(test_loader.dataset)))


def inference():
    print('Inference')
    model.eval()
    data = torch.rand(64, 1, 28, 28).cuda()
    output = model(data)
    print(output.size())



def main():
    # Training settings
    parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
    parser.add_argument('--batch-size', type=int, default=64, metavar='N',
                        help='input batch size for training (default: 64)')
    parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
                        help='input batch size for testing (default: 1000)')
    parser.add_argument('--epochs', type=int, default=14, metavar='N',
                        help='number of epochs to train (default: 14)')
    parser.add_argument('--lr', type=float, default=1.0, metavar='LR',
                        help='learning rate (default: 1.0)')
    parser.add_argument('--gamma', type=float, default=0.7, metavar='M',
                        help='Learning rate step gamma (default: 0.7)')
    parser.add_argument('--no-cuda', action='store_true', default=False,
                        help='disables CUDA training')
    parser.add_argument('--no-mps', action='store_true', default=False,
                        help='disables macOS GPU training')
    parser.add_argument('--dry-run', action='store_true', default=False,
                        help='quickly check a single pass')
    parser.add_argument('--seed', type=int, default=1, metavar='S',
                        help='random seed (default: 1)')
    parser.add_argument('--log-interval', type=int, default=10, metavar='N',
                        help='how many batches to wait before logging training status')
    parser.add_argument('--save-model', action='store_true', default=False,
                        help='For Saving the current Model')
    args = parser.parse_args()
    use_cuda = not args.no_cuda and torch.cuda.is_available()
    use_mps = not args.no_mps and torch.backends.mps.is_available()

    torch.manual_seed(args.seed)

    if use_cuda:
        device = torch.device("cuda")
    elif use_mps:
        device = torch.device("mps")
    else:
        device = torch.device("cpu")

    train_kwargs = {'batch_size': args.batch_size}
    test_kwargs = {'batch_size': args.test_batch_size}
    if use_cuda:
        cuda_kwargs = {'num_workers': 1,
                       'pin_memory': True,
                       'shuffle': True}
        train_kwargs.update(cuda_kwargs)
        test_kwargs.update(cuda_kwargs)

    transform=transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize((0.1307,), (0.3081,))
        ])
    dataset1 = datasets.MNIST('../data', train=True, download=True,
                       transform=transform)
    dataset2 = datasets.MNIST('../data', train=False,
                       transform=transform)
    train_loader = torch.utils.data.DataLoader(dataset1,**train_kwargs)
    test_loader = torch.utils.data.DataLoader(dataset2, **test_kwargs)

    model = Net().to(device)
    optimizer = optim.Adadelta(model.parameters(), lr=args.lr)

    scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)
    for epoch in range(1, args.epochs + 1):
        train(args, model, device, train_loader, optimizer, epoch)
        test(model, device, test_loader)
        scheduler.step()

    if args.save_model:
        torch.save(model.state_dict(), "mnist_cnn.pt")


if __name__ == '__main__':
    main()

发现生成了很多json文件，保存了Profile的信息:

image.png

采用Tensorboard打开:
tensorboard --logdir=./profile

image.png

查看Training interation=2的运行情况:
可以观察到3部分内容: Forward, backword, Optimizer.step (weight update)

image.png

PyTorch 模型性能分析——PyTorch Profiler

PyTorch 模型性能分析——PyTorch Profiler

前言

测试环境

Profiler 性能分析工具介绍

PyTorch Profiler工具

Profiler分析CPU、GPU端Op执行时间

CPU Only 模式

GPU 模式

使用Chrome trace可视化Profiler结果

Profiler 结果分析

模型整体分析

CPU 和 GPU Op的关系

Op的wall_duration_time, self_time 区别

Op Tensor数据维度分析

DNN Model Training Profile