（beta）PyTorch 中使用 Eager 模式的静态量化¶

创建于：2021 年 3 月 24 日 | 最后更新：2024 年 8 月 27 日 | 最后验证：2024 年 11 月 05 日

作者：Raghuraman Krishnamoorthi 编辑：Seth Weidman、Jerry Zhang

本教程演示了如何进行训练后静态量化，并说明了两种更高级的技术 - 逐通道量化和量化感知训练 - 以进一步提高模型的准确性。请注意，量化目前仅支持 CPU，因此在本教程中我们不会使用 GPU/CUDA。在本教程结束时，您将了解 PyTorch 中的量化如何显著减小模型大小并提高速度。此外，您将了解如何轻松应用此处所示的一些高级量化技术，从而使您的量化模型比其他方式减少更多的精度损失。警告：我们使用了来自其他 PyTorch 代码库的大量样板代码，例如，定义 MobileNetV2 模型架构、定义数据加载器等等。我们当然鼓励您阅读它；但是，如果您想了解量化功能，请随意跳到“4. 训练后静态量化”部分。我们将从必要的导入开始

import os
import sys
import time
import numpy as np

import torch
import torch.nn as nn
from torch.utils.data import DataLoader

import torchvision
from torchvision import datasets
import torchvision.transforms as transforms

# Set up warnings
import warnings
warnings.filterwarnings(
    action='ignore',
    category=DeprecationWarning,
    module=r'.*'
)
warnings.filterwarnings(
    action='default',
    module=r'torch.ao.quantization'
)

# Specify random seed for repeatable results
torch.manual_seed(191009)

1. 模型架构¶

我们首先定义 MobileNetV2 模型架构，并进行了一些值得注意的修改以实现量化

用 nn.quantized.FloatFunctional 替换加法
在网络的开头和结尾插入 QuantStub 和 DeQuantStub。
用 ReLU 替换 ReLU6

注意：此代码取自此处。

from torch.ao.quantization import QuantStub, DeQuantStub

def _make_divisible(v, divisor, min_value=None):
    """
    This function is taken from the original tf repo.
    It ensures that all layers have a channel number that is divisible by 8
    It can be seen here:
    https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py
    :param v:
    :param divisor:
    :param min_value:
    :return:
    """
    if min_value is None:
        min_value = divisor
    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
    # Make sure that round down does not go down by more than 10%.
    if new_v < 0.9 * v:
        new_v += divisor
    return new_v


class ConvBNReLU(nn.Sequential):
    def __init__(self, in_planes, out_planes, kernel_size=3, stride=1, groups=1):
        padding = (kernel_size - 1) // 2
        super(ConvBNReLU, self).__init__(
            nn.Conv2d(in_planes, out_planes, kernel_size, stride, padding, groups=groups, bias=False),
            nn.BatchNorm2d(out_planes, momentum=0.1),
            # Replace with ReLU
            nn.ReLU(inplace=False)
        )


class InvertedResidual(nn.Module):
    def __init__(self, inp, oup, stride, expand_ratio):
        super(InvertedResidual, self).__init__()
        self.stride = stride
        assert stride in [1, 2]

        hidden_dim = int(round(inp * expand_ratio))
        self.use_res_connect = self.stride == 1 and inp == oup

        layers = []
        if expand_ratio != 1:
            # pw
            layers.append(ConvBNReLU(inp, hidden_dim, kernel_size=1))
        layers.extend([
            # dw
            ConvBNReLU(hidden_dim, hidden_dim, stride=stride, groups=hidden_dim),
            # pw-linear
            nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
            nn.BatchNorm2d(oup, momentum=0.1),
        ])
        self.conv = nn.Sequential(*layers)
        # Replace torch.add with floatfunctional
        self.skip_add = nn.quantized.FloatFunctional()

    def forward(self, x):
        if self.use_res_connect:
            return self.skip_add.add(x, self.conv(x))
        else:
            return self.conv(x)


class MobileNetV2(nn.Module):
    def __init__(self, num_classes=1000, width_mult=1.0, inverted_residual_setting=None, round_nearest=8):
        """
        MobileNet V2 main class
        Args:
            num_classes (int): Number of classes
            width_mult (float): Width multiplier - adjusts number of channels in each layer by this amount
            inverted_residual_setting: Network structure
            round_nearest (int): Round the number of channels in each layer to be a multiple of this number
            Set to 1 to turn off rounding
        """
        super(MobileNetV2, self).__init__()
        block = InvertedResidual
        input_channel = 32
        last_channel = 1280

        if inverted_residual_setting is None:
            inverted_residual_setting = [
                # t, c, n, s
                [1, 16, 1, 1],
                [6, 24, 2, 2],
                [6, 32, 3, 2],
                [6, 64, 4, 2],
                [6, 96, 3, 1],
                [6, 160, 3, 2],
                [6, 320, 1, 1],
            ]

        # only check the first element, assuming user knows t,c,n,s are required
        if len(inverted_residual_setting) == 0 or len(inverted_residual_setting[0]) != 4:
            raise ValueError("inverted_residual_setting should be non-empty "
                             "or a 4-element list, got {}".format(inverted_residual_setting))

        # building first layer
        input_channel = _make_divisible(input_channel * width_mult, round_nearest)
        self.last_channel = _make_divisible(last_channel * max(1.0, width_mult), round_nearest)
        features = [ConvBNReLU(3, input_channel, stride=2)]
        # building inverted residual blocks
        for t, c, n, s in inverted_residual_setting:
            output_channel = _make_divisible(c * width_mult, round_nearest)
            for i in range(n):
                stride = s if i == 0 else 1
                features.append(block(input_channel, output_channel, stride, expand_ratio=t))
                input_channel = output_channel
        # building last several layers
        features.append(ConvBNReLU(input_channel, self.last_channel, kernel_size=1))
        # make it nn.Sequential
        self.features = nn.Sequential(*features)
        self.quant = QuantStub()
        self.dequant = DeQuantStub()
        # building classifier
        self.classifier = nn.Sequential(
            nn.Dropout(0.2),
            nn.Linear(self.last_channel, num_classes),
        )

        # weight initialization
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out')
                if m.bias is not None:
                    nn.init.zeros_(m.bias)
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.ones_(m.weight)
                nn.init.zeros_(m.bias)
            elif isinstance(m, nn.Linear):
                nn.init.normal_(m.weight, 0, 0.01)
                nn.init.zeros_(m.bias)

    def forward(self, x):
        x = self.quant(x)
        x = self.features(x)
        x = x.mean([2, 3])
        x = self.classifier(x)
        x = self.dequant(x)
        return x

    # Fuse Conv+BN and Conv+BN+Relu modules prior to quantization
    # This operation does not change the numerics
    def fuse_model(self, is_qat=False):
        fuse_modules = torch.ao.quantization.fuse_modules_qat if is_qat else torch.ao.quantization.fuse_modules
        for m in self.modules():
            if type(m) == ConvBNReLU:
                fuse_modules(m, ['0', '1', '2'], inplace=True)
            if type(m) == InvertedResidual:
                for idx in range(len(m.conv)):
                    if type(m.conv[idx]) == nn.Conv2d:
                        fuse_modules(m.conv, [str(idx), str(idx + 1)], inplace=True)

2. 辅助函数¶

接下来，我们定义几个辅助函数来帮助进行模型评估。这些函数主要来自此处。

class AverageMeter(object):
    """Computes and stores the average and current value"""
    def __init__(self, name, fmt=':f'):
        self.name = name
        self.fmt = fmt
        self.reset()

    def reset(self):
        self.val = 0
        self.avg = 0
        self.sum = 0
        self.count = 0

    def update(self, val, n=1):
        self.val = val
        self.sum += val * n
        self.count += n
        self.avg = self.sum / self.count

    def __str__(self):
        fmtstr = '{name} {val' + self.fmt + '} ({avg' + self.fmt + '})'
        return fmtstr.format(**self.__dict__)


def accuracy(output, target, topk=(1,)):
    """Computes the accuracy over the k top predictions for the specified values of k"""
    with torch.no_grad():
        maxk = max(topk)
        batch_size = target.size(0)

        _, pred = output.topk(maxk, 1, True, True)
        pred = pred.t()
        correct = pred.eq(target.view(1, -1).expand_as(pred))

        res = []
        for k in topk:
            correct_k = correct[:k].reshape(-1).float().sum(0, keepdim=True)
            res.append(correct_k.mul_(100.0 / batch_size))
        return res


def evaluate(model, criterion, data_loader, neval_batches):
    model.eval()
    top1 = AverageMeter('Acc@1', ':6.2f')
    top5 = AverageMeter('Acc@5', ':6.2f')
    cnt = 0
    with torch.no_grad():
        for image, target in data_loader:
            output = model(image)
            loss = criterion(output, target)
            cnt += 1
            acc1, acc5 = accuracy(output, target, topk=(1, 5))
            print('.', end = '')
            top1.update(acc1[0], image.size(0))
            top5.update(acc5[0], image.size(0))
            if cnt >= neval_batches:
                 return top1, top5

    return top1, top5

def load_model(model_file):
    model = MobileNetV2()
    state_dict = torch.load(model_file, weights_only=True)
    model.load_state_dict(state_dict)
    model.to('cpu')
    return model

def print_size_of_model(model):
    torch.save(model.state_dict(), "temp.p")
    print('Size (MB):', os.path.getsize("temp.p")/1e6)
    os.remove('temp.p')

3. 定义数据集和数据加载器¶

作为我们最后的主要设置步骤，我们为训练集和测试集定义数据加载器。

ImageNet 数据¶

要使用整个 ImageNet 数据集运行本教程中的代码，请首先按照 ImageNet Data 中的说明下载 imagenet。将下载的文件解压缩到“data_path”文件夹中。

下载数据后，我们将在下面展示定义数据加载器的函数，我们将使用这些加载器来读取数据。这些函数主要来自此处。

def prepare_data_loaders(data_path):
    normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                     std=[0.229, 0.224, 0.225])
    dataset = torchvision.datasets.ImageNet(
        data_path, split="train", transform=transforms.Compose([
            transforms.RandomResizedCrop(224),
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor(),
            normalize,
        ]))
    dataset_test = torchvision.datasets.ImageNet(
        data_path, split="val", transform=transforms.Compose([
            transforms.Resize(256),
            transforms.CenterCrop(224),
            transforms.ToTensor(),
            normalize,
        ]))

    train_sampler = torch.utils.data.RandomSampler(dataset)
    test_sampler = torch.utils.data.SequentialSampler(dataset_test)

    data_loader = torch.utils.data.DataLoader(
        dataset, batch_size=train_batch_size,
        sampler=train_sampler)

    data_loader_test = torch.utils.data.DataLoader(
        dataset_test, batch_size=eval_batch_size,
        sampler=test_sampler)

    return data_loader, data_loader_test

接下来，我们将加载预训练的 MobileNetV2 模型。我们在此处提供下载模型的 URL here。

data_path = '~/.data/imagenet'
saved_model_dir = 'data/'
float_model_file = 'mobilenet_pretrained_float.pth'
scripted_float_model_file = 'mobilenet_quantization_scripted.pth'
scripted_quantized_model_file = 'mobilenet_quantization_scripted_quantized.pth'

train_batch_size = 30
eval_batch_size = 50

data_loader, data_loader_test = prepare_data_loaders(data_path)
criterion = nn.CrossEntropyLoss()
float_model = load_model(saved_model_dir + float_model_file).to('cpu')

# Next, we'll "fuse modules"; this can both make the model faster by saving on memory access
# while also improving numerical accuracy. While this can be used with any model, this is
# especially common with quantized models.

print('\n Inverted Residual Block: Before fusion \n\n', float_model.features[1].conv)
float_model.eval()

# Fuses modules
float_model.fuse_model()

# Note fusion of Conv+BN+Relu and Conv+Relu
print('\n Inverted Residual Block: After fusion\n\n',float_model.features[1].conv)

最后，为了获得“基线”准确率，让我们看看融合模块的未量化模型的准确率

num_eval_batches = 1000

print("Size of baseline model")
print_size_of_model(float_model)

top1, top5 = evaluate(float_model, criterion, data_loader_test, neval_batches=num_eval_batches)
print('Evaluation accuracy on %d images, %2.2f'%(num_eval_batches * eval_batch_size, top1.avg))
torch.jit.save(torch.jit.script(float_model), saved_model_dir + scripted_float_model_file)

在整个模型上，我们在包含 50,000 张图像的评估数据集上获得了 71.9% 的准确率。

这将是我们的基线，用于进行比较。接下来，让我们尝试不同的量化方法

4. 训练后静态量化¶

训练后静态量化不仅涉及将权重从浮点数转换为整数，就像动态量化一样，还涉及执行额外的步骤，即首先通过网络馈送成批数据，并计算不同激活的结果分布（具体而言，这是通过在不同点插入 observer 模块来记录此数据来完成的）。然后，这些分布用于确定在推理时应如何具体量化不同的激活（一种简单的技术是将整个激活范围简单地划分为 256 个级别，但我们也支持更复杂的方法）。重要的是，这个额外的步骤使我们能够在操作之间传递量化值，而不是在每次操作之间将这些值转换为浮点数 - 然后再转换回整数 - 从而显著提高速度。

num_calibration_batches = 32

myModel = load_model(saved_model_dir + float_model_file).to('cpu')
myModel.eval()

# Fuse Conv, bn and relu
myModel.fuse_model()

# Specify quantization configuration
# Start with simple min/max range estimation and per-tensor quantization of weights
myModel.qconfig = torch.ao.quantization.default_qconfig
print(myModel.qconfig)
torch.ao.quantization.prepare(myModel, inplace=True)

# Calibrate first
print('Post Training Quantization Prepare: Inserting Observers')
print('\n Inverted Residual Block:After observer insertion \n\n', myModel.features[1].conv)

# Calibrate with the training set
evaluate(myModel, criterion, data_loader, neval_batches=num_calibration_batches)
print('Post Training Quantization: Calibration done')

# Convert to quantized model
torch.ao.quantization.convert(myModel, inplace=True)
# You may see a user warning about needing to calibrate the model. This warning can be safely ignored.
# This warning occurs because not all modules are run in each model runs, so some
# modules may not be calibrated.
print('Post Training Quantization: Convert done')
print('\n Inverted Residual Block: After fusion and quantization, note fused modules: \n\n',myModel.features[1].conv)

print("Size of model after quantization")
print_size_of_model(myModel)

top1, top5 = evaluate(myModel, criterion, data_loader_test, neval_batches=num_eval_batches)
print('Evaluation accuracy on %d images, %2.2f'%(num_eval_batches * eval_batch_size, top1.avg))

对于此量化模型，我们在评估数据集上看到了 56.7% 的准确率。这是因为我们使用了简单的最小/最大观察器来确定量化参数。尽管如此，我们确实将模型的大小减少到略低于 3.6 MB，减少了近 4 倍。

此外，我们可以仅通过使用不同的量化配置来显著提高准确率。我们使用推荐的配置重复相同的练习，以针对 x86 架构进行量化。此配置执行以下操作

在逐通道的基础上量化权重
使用直方图观察器，该观察器收集激活的直方图，然后以最佳方式选择量化参数。

per_channel_quantized_model = load_model(saved_model_dir + float_model_file)
per_channel_quantized_model.eval()
per_channel_quantized_model.fuse_model()
# The old 'fbgemm' is still available but 'x86' is the recommended default.
per_channel_quantized_model.qconfig = torch.ao.quantization.get_default_qconfig('x86')
print(per_channel_quantized_model.qconfig)

torch.ao.quantization.prepare(per_channel_quantized_model, inplace=True)
evaluate(per_channel_quantized_model,criterion, data_loader, num_calibration_batches)
torch.ao.quantization.convert(per_channel_quantized_model, inplace=True)
top1, top5 = evaluate(per_channel_quantized_model, criterion, data_loader_test, neval_batches=num_eval_batches)
print('Evaluation accuracy on %d images, %2.2f'%(num_eval_batches * eval_batch_size, top1.avg))
torch.jit.save(torch.jit.script(per_channel_quantized_model), saved_model_dir + scripted_quantized_model_file)

仅更改此量化配置方法就使准确率提高到 67.3% 以上！但是，这仍然比上面达到的 71.9% 基线差 4%。因此，让我们尝试量化感知训练。

5. 量化感知训练¶

量化感知训练 (QAT) 是通常产生最高准确率的量化方法。使用 QAT，所有权重和激活在训练的前向和后向传递中都被“伪量化”：也就是说，浮点值被舍入以模拟 int8 值，但所有计算仍然使用浮点数完成。因此，训练期间的所有权重调整都是在“感知”到模型最终将被量化的事实下进行的；因此，在量化之后，此方法通常会比动态量化或训练后静态量化产生更高的准确率。

实际执行 QAT 的总体工作流程与之前非常相似

我们可以使用与之前相同的模型：量化感知训练不需要额外的准备。
我们需要使用 qconfig 指定要在权重和激活之后插入哪种伪量化，而不是指定观察器

我们首先定义一个训练函数

def train_one_epoch(model, criterion, optimizer, data_loader, device, ntrain_batches):
    model.train()
    top1 = AverageMeter('Acc@1', ':6.2f')
    top5 = AverageMeter('Acc@5', ':6.2f')
    avgloss = AverageMeter('Loss', '1.5f')

    cnt = 0
    for image, target in data_loader:
        start_time = time.time()
        print('.', end = '')
        cnt += 1
        image, target = image.to(device), target.to(device)
        output = model(image)
        loss = criterion(output, target)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        acc1, acc5 = accuracy(output, target, topk=(1, 5))
        top1.update(acc1[0], image.size(0))
        top5.update(acc5[0], image.size(0))
        avgloss.update(loss, image.size(0))
        if cnt >= ntrain_batches:
            print('Loss', avgloss.avg)

            print('Training: * Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}'
                  .format(top1=top1, top5=top5))
            return

    print('Full imagenet train set:  * Acc@1 {top1.global_avg:.3f} Acc@5 {top5.global_avg:.3f}'
          .format(top1=top1, top5=top5))
    return

我们像以前一样融合模块

qat_model = load_model(saved_model_dir + float_model_file)
qat_model.fuse_model(is_qat=True)

optimizer = torch.optim.SGD(qat_model.parameters(), lr = 0.0001)
# The old 'fbgemm' is still available but 'x86' is the recommended default.
qat_model.qconfig = torch.ao.quantization.get_default_qat_qconfig('x86')

最后，prepare_qat 执行“伪量化”，为量化感知训练准备模型

torch.ao.quantization.prepare_qat(qat_model, inplace=True)
print('Inverted Residual Block: After preparation for QAT, note fake-quantization modules \n',qat_model.features[1].conv)

训练具有高准确率的量化模型需要在推理时准确地建模数值。因此，对于量化感知训练，我们通过以下方式修改训练循环

在训练结束时切换批归一化以使用运行均值和方差，以更好地匹配推理数值。
我们还冻结量化器参数（比例和零点）并微调权重。

num_train_batches = 20

# QAT takes time and one needs to train over a few epochs.
# Train and check accuracy after each epoch
for nepoch in range(8):
    train_one_epoch(qat_model, criterion, optimizer, data_loader, torch.device('cpu'), num_train_batches)
    if nepoch > 3:
        # Freeze quantizer parameters
        qat_model.apply(torch.ao.quantization.disable_observer)
    if nepoch > 2:
        # Freeze batch norm mean and variance estimates
        qat_model.apply(torch.nn.intrinsic.qat.freeze_bn_stats)

    # Check the accuracy after each epoch
    quantized_model = torch.ao.quantization.convert(qat_model.eval(), inplace=False)
    quantized_model.eval()
    top1, top5 = evaluate(quantized_model,criterion, data_loader_test, neval_batches=num_eval_batches)
    print('Epoch %d :Evaluation accuracy on %d images, %2.2f'%(nepoch, num_eval_batches * eval_batch_size, top1.avg))

量化感知训练在整个 imagenet 数据集上产生了超过 71.5% 的准确率，这接近于 71.9% 的浮点准确率。

有关量化感知训练的更多信息

QAT 是训练后量化技术的超集，可以进行更多调试。例如，我们可以分析模型的准确率是否受到权重或激活量化的限制。
我们还可以在浮点数中模拟量化模型的准确率，因为我们正在使用伪量化来建模实际量化算术的数值。
我们也可以轻松地模拟训练后量化。

量化带来的加速¶

最后，让我们确认一下我们在上面暗示的内容：我们的量化模型是否真的执行推理速度更快？让我们测试一下

def run_benchmark(model_file, img_loader):
    elapsed = 0
    model = torch.jit.load(model_file)
    model.eval()
    num_batches = 5
    # Run the scripted model on a few batches of images
    for i, (images, target) in enumerate(img_loader):
        if i < num_batches:
            start = time.time()
            output = model(images)
            end = time.time()
            elapsed = elapsed + (end-start)
        else:
            break
    num_images = images.size()[0] * num_batches

    print('Elapsed time: %3.0f ms' % (elapsed/num_images*1000))
    return elapsed

run_benchmark(saved_model_dir + scripted_float_model_file, data_loader_test)

run_benchmark(saved_model_dir + scripted_quantized_model_file, data_loader_test)

在 MacBook Pro 上本地运行此代码，常规模型耗时 61 毫秒，而量化模型仅耗时 20 毫秒，这说明了与浮点模型相比，量化模型通常可以看到 2-4 倍的加速。

结论¶

在本教程中，我们展示了两种量化方法 - 训练后静态量化和量化感知训练 - 描述了它们在“底层”的工作原理以及如何在 PyTorch 中使用它们。

感谢您的阅读！与往常一样，我们欢迎任何反馈，因此如果您有任何问题，请在此处创建问题 here。