PyTorch Lightning 1.1: research : CIFAR100 (ResNet)
作成 : (株)クラスキャットセールスインフォメーション
作成日時 : 02/24/2021 (1.1.x)

* 本ページは以下の CIFAR10 用リソースを参考に CIFAR100 で遂行した実験結果のレポートです：

* ご自由にリンクを張って頂いてかまいませんが、sales-info@classcat.com までご一報いただけると嬉しいです。

★ 無料セミナー実施中 ★ クラスキャット主催人工知能 & ビジネス Web セミナー

人工知能とビジネスをテーマにウェビナー (WEB セミナー) を定期的に開催しています。スケジュールは弊社公式 Web サイトでご確認頂けます。

お住まいの地域に関係なく Web ブラウザからご参加頂けます。事前登録 が必要ですのでご注意ください。
Windows PC のブラウザからご参加が可能です。スマートデバイスもご利用可能です。

クラスキャットは人工知能・テレワークに関する各種サービスを提供しております :

人工知能研究開発支援	人工知能研修サービス	テレワーク & オンライン授業を支援
PoC(概念実証)を失敗させないための支援 (本支援はセミナーに参加しアンケートに回答した方を対象としています。)

◆ お問合せ : 本件に関するお問い合わせ先は下記までお願いいたします。

株式会社クラスキャット セールス・マーケティング本部セールス・インフォメーション

E-Mail：sales-info@classcat.com ; WebSite: https://www.classcat.com/

Facebook: https://www.facebook.com/ClassCatJP/

research: CIFAR100 (ResNet)

ResNet18

仕様

ResNet18
Total params: 11,220,132 (11.2 M)
Trainable params: 11,220,132
Non-trainable params: 0

結果

ResNet18
{‘test_acc’: 0.7301999926567078, ‘test_loss’: 1.1373273134231567}
100 エポック ; Wall time: 1h 39min 34s (‘Tesla M60’ x 2)
ReduceLROnPlateau

ResNet34

仕様

ResNet34
Total params: 21,328,292 (21.3 M)
Trainable params: 21,328,292
Non-trainable params: 0

結果

ResNet34
{‘test_acc’: 0.7281000018119812, ‘test_loss’: 1.155539631843567}
100 エポック ; Wall time: 3h 18s
Tesla T4
ReduceLROnPlateau

CIFAR 100 DM

from typing import Any, Callable, Optional, Sequence, Union

from pl_bolts.datamodules.vision_datamodule import VisionDataModule
#from pl_bolts.datasets import TrialCIFAR10
#from pl_bolts.transforms.dataset_normalizations import cifar10_normalization
from pl_bolts.utils import _TORCHVISION_AVAILABLE
from pl_bolts.utils.warnings import warn_missing_pkg

if _TORCHVISION_AVAILABLE:
    from torchvision import transforms
    #from torchvision import transforms as transform_lib
    from torchvision.datasets import CIFAR100
else:  # pragma: no cover
    warn_missing_pkg('torchvision')
    CIFAR100 = None

def cifar100_normalization():
    if not _TORCHVISION_AVAILABLE:  # pragma: no cover
        raise ModuleNotFoundError(
            'You want to use `torchvision` which is not installed yet, install it with `pip install torchvision`.'
        )

    normalize = transforms.Normalize(
        mean=[x / 255.0 for x in [129.3, 124.1, 112.4]],
        std=[x / 255.0 for x in [68.2, 65.4, 70.4]],
        # cifar10
        #mean=[x / 255.0 for x in [125.3, 123.0, 113.9]],
        #std=[x / 255.0 for x in [63.0, 62.1, 66.7]],
    )
    return normalize

class CIFAR100DataModule(VisionDataModule):
    """
    .. figure:: https://3qeqpr26caki16dnhd19sv6by6v-wpengine.netdna-ssl.com/wp-content/uploads/2019/01/
        Plot-of-a-Subset-of-Images-from-the-CIFAR-10-Dataset.png
        :width: 400
        :alt: CIFAR-10
    Specs:
        - 10 classes (1 per class)
        - Each image is (3 x 32 x 32)
    Standard CIFAR10, train, val, test splits and transforms
    Transforms::
        mnist_transforms = transform_lib.Compose([
            transform_lib.ToTensor(),
            transforms.Normalize(
                mean=[x / 255.0 for x in [125.3, 123.0, 113.9]],
                std=[x / 255.0 for x in [63.0, 62.1, 66.7]]
            )
        ])
    Example::
        from pl_bolts.datamodules import CIFAR10DataModule
        dm = CIFAR10DataModule(PATH)
        model = LitModel()
        Trainer().fit(model, datamodule=dm)
    Or you can set your own transforms
    Example::
        dm.train_transforms = ...
        dm.test_transforms = ...
        dm.val_transforms  = ...
    """
    name = "cifar100"
    dataset_cls = CIFAR100
    dims = (3, 32, 32)

    def __init__(
        self,
        data_dir: Optional[str] = None,
        val_split: Union[int, float] = 0.2,
        num_workers: int = 16,
        normalize: bool = False,
        batch_size: int = 32,
        seed: int = 42,
        shuffle: bool = False,
        pin_memory: bool = False,
        drop_last: bool = False,
        *args: Any,
        **kwargs: Any,
    ) -> None:
        """
        Args:
            data_dir: Where to save/load the data
            val_split: Percent (float) or number (int) of samples to use for the validation split
            num_workers: How many workers to use for loading data
            normalize: If true applies image normalize
            batch_size: How many samples per batch to load
            seed: Random seed to be used for train/val/test splits
            shuffle: If true shuffles the train data every epoch
            pin_memory: If true, the data loader will copy Tensors into CUDA pinned memory before
                        returning them
            drop_last: If true drops the last incomplete batch
        """
        super().__init__(  # type: ignore[misc]
            data_dir=data_dir,
            val_split=val_split,
            num_workers=num_workers,
            normalize=normalize,
            batch_size=batch_size,
            seed=seed,
            shuffle=shuffle,
            pin_memory=pin_memory,
            drop_last=drop_last,
            *args,
            **kwargs,
        )

    @property
    def num_samples(self) -> int:
        train_len, _ = self._get_splits(len_dataset=50_000)
        return train_len

    @property
    def num_classes(self) -> int:
        """
        Return:
            10
        """
        return 100

    def default_transforms(self) -> Callable:
        if self.normalize:
            cf100_transforms = transforms.Compose([transform_lib.ToTensor(), cifar100_normalization()])
        else:
            cf100_transforms = transforms.Compose([transform_lib.ToTensor()])

        return cf100_transforms

モデル

import torch
import torch.nn as nn
import torch.nn.functional as F

class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, in_planes, planes, stride=1):
        super(BasicBlock, self).__init__()
        self.conv1 = nn.Conv2d(
            in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(planes)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3,
                               stride=1, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(planes)

        self.shortcut = nn.Sequential()
        if stride != 1 or in_planes != self.expansion*planes:
            self.shortcut = nn.Sequential(
                nn.Conv2d(in_planes, self.expansion*planes,
                          kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(self.expansion*planes)
            )

    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))
        out += self.shortcut(x)
        out = F.relu(out)
        return out


class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, in_planes, planes, stride=1):
        super(Bottleneck, self).__init__()
        self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, bias=False)
        self.bn1 = nn.BatchNorm2d(planes)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3,
                               stride=stride, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(planes)
        self.conv3 = nn.Conv2d(planes, self.expansion *
                               planes, kernel_size=1, bias=False)
        self.bn3 = nn.BatchNorm2d(self.expansion*planes)

        self.shortcut = nn.Sequential()
        if stride != 1 or in_planes != self.expansion*planes:
            self.shortcut = nn.Sequential(
                nn.Conv2d(in_planes, self.expansion*planes,
                          kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(self.expansion*planes)
            )

    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = F.relu(self.bn2(self.conv2(out)))
        out = self.bn3(self.conv3(out))
        out += self.shortcut(x)
        out = F.relu(out)
        return out


class ResNet(nn.Module):
    def __init__(self, block, num_blocks, num_classes=100):
        super(ResNet, self).__init__()
        self.in_planes = 64

        self.conv1 = nn.Conv2d(3, 64, kernel_size=3,
                               stride=1, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(64)
        self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1)
        self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2)
        self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2)
        self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2)
        self.linear = nn.Linear(512*block.expansion, num_classes)

    def _make_layer(self, block, planes, num_blocks, stride):
        strides = [stride] + [1]*(num_blocks-1)
        layers = []
        for stride in strides:
            layers.append(block(self.in_planes, planes, stride))
            self.in_planes = planes * block.expansion
        return nn.Sequential(*layers)

    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.layer4(out)
        out = F.avg_pool2d(out, 4)
        out = out.view(out.size(0), -1)
        out = self.linear(out)
        return out


def ResNet18():
    return ResNet(BasicBlock, [2, 2, 2, 2])


def ResNet34():
    return ResNet(BasicBlock, [3, 4, 6, 3])


def ResNet50():
    return ResNet(Bottleneck, [3, 4, 6, 3])


def ResNet101():
    return ResNet(Bottleneck, [3, 4, 23, 3])


def ResNet152():
    return ResNet(Bottleneck, [3, 8, 36, 3])

net = ResNet18()
print(net)
y = net(torch.randn(1, 3, 32, 32))
print(y.size())

ResNet(
  (conv1): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
  (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  (layer1): Sequential(
    (0): BasicBlock(
      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (shortcut): Sequential()
    )
    (1): BasicBlock(
      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (shortcut): Sequential()
    )
  )
  (layer2): Sequential(
    (0): BasicBlock(
      (conv1): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (shortcut): Sequential(
        (0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): BasicBlock(
      (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (shortcut): Sequential()
    )
  )
  (layer3): Sequential(
    (0): BasicBlock(
      (conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (shortcut): Sequential(
        (0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): BasicBlock(
      (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (shortcut): Sequential()
    )
  )
  (layer4): Sequential(
    (0): BasicBlock(
      (conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (shortcut): Sequential(
        (0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): BasicBlock(
      (conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (shortcut): Sequential()
    )
  )
  (linear): Linear(in_features=512, out_features=100, bias=True)
)
torch.Size([1, 100])

from torchsummary import summary

summary(ResNet18().to('cuda'), (3, 32, 32))

----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1           [-1, 64, 32, 32]           1,728
       BatchNorm2d-2           [-1, 64, 32, 32]             128
            Conv2d-3           [-1, 64, 32, 32]          36,864
       BatchNorm2d-4           [-1, 64, 32, 32]             128
            Conv2d-5           [-1, 64, 32, 32]          36,864
       BatchNorm2d-6           [-1, 64, 32, 32]             128
        BasicBlock-7           [-1, 64, 32, 32]               0
            Conv2d-8           [-1, 64, 32, 32]          36,864
       BatchNorm2d-9           [-1, 64, 32, 32]             128
           Conv2d-10           [-1, 64, 32, 32]          36,864
      BatchNorm2d-11           [-1, 64, 32, 32]             128
       BasicBlock-12           [-1, 64, 32, 32]               0
           Conv2d-13          [-1, 128, 16, 16]          73,728
      BatchNorm2d-14          [-1, 128, 16, 16]             256
           Conv2d-15          [-1, 128, 16, 16]         147,456
      BatchNorm2d-16          [-1, 128, 16, 16]             256
           Conv2d-17          [-1, 128, 16, 16]           8,192
      BatchNorm2d-18          [-1, 128, 16, 16]             256
       BasicBlock-19          [-1, 128, 16, 16]               0
           Conv2d-20          [-1, 128, 16, 16]         147,456
      BatchNorm2d-21          [-1, 128, 16, 16]             256
           Conv2d-22          [-1, 128, 16, 16]         147,456
      BatchNorm2d-23          [-1, 128, 16, 16]             256
       BasicBlock-24          [-1, 128, 16, 16]               0
           Conv2d-25            [-1, 256, 8, 8]         294,912
      BatchNorm2d-26            [-1, 256, 8, 8]             512
           Conv2d-27            [-1, 256, 8, 8]         589,824
      BatchNorm2d-28            [-1, 256, 8, 8]             512
           Conv2d-29            [-1, 256, 8, 8]          32,768
      BatchNorm2d-30            [-1, 256, 8, 8]             512
       BasicBlock-31            [-1, 256, 8, 8]               0
           Conv2d-32            [-1, 256, 8, 8]         589,824
      BatchNorm2d-33            [-1, 256, 8, 8]             512
           Conv2d-34            [-1, 256, 8, 8]         589,824
      BatchNorm2d-35            [-1, 256, 8, 8]             512
       BasicBlock-36            [-1, 256, 8, 8]               0
           Conv2d-37            [-1, 512, 4, 4]       1,179,648
      BatchNorm2d-38            [-1, 512, 4, 4]           1,024
           Conv2d-39            [-1, 512, 4, 4]       2,359,296
      BatchNorm2d-40            [-1, 512, 4, 4]           1,024
           Conv2d-41            [-1, 512, 4, 4]         131,072
      BatchNorm2d-42            [-1, 512, 4, 4]           1,024
       BasicBlock-43            [-1, 512, 4, 4]               0
           Conv2d-44            [-1, 512, 4, 4]       2,359,296
      BatchNorm2d-45            [-1, 512, 4, 4]           1,024
           Conv2d-46            [-1, 512, 4, 4]       2,359,296
      BatchNorm2d-47            [-1, 512, 4, 4]           1,024
       BasicBlock-48            [-1, 512, 4, 4]               0
           Linear-49                  [-1, 100]          51,300
================================================================
Total params: 11,220,132
Trainable params: 11,220,132
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.01
Forward/backward pass size (MB): 11.25
Params size (MB): 42.80
Estimated Total Size (MB): 54.06
----------------------------------------------------------------

Lightning モジュール

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim.lr_scheduler import OneCycleLR, CyclicLR, ExponentialLR, CosineAnnealingLR, ReduceLROnPlateau
from torch.optim.swa_utils import AveragedModel, update_bn
import torchvision
 
import pytorch_lightning as pl
from pytorch_lightning.callbacks import LearningRateMonitor, GPUStatsMonitor, EarlyStopping
from pytorch_lightning.metrics.functional import accuracy
#from pl_bolts.datamodules import CIFAR10DataModule
#from pl_bolts.transforms.dataset_normalizations import cifar10_normalization

pl.seed_everything(7);

batch_size = 100
 
train_transforms = torchvision.transforms.Compose([
    torchvision.transforms.RandomCrop(32, padding=4),
    torchvision.transforms.RandomHorizontalFlip(),
    torchvision.transforms.ToTensor(),
    cifar100_normalization(),
])
 
test_transforms = torchvision.transforms.Compose([
    torchvision.transforms.ToTensor(),
    cifar100_normalization(),
])
 
cifar100_dm = CIFAR100DataModule(
    batch_size=batch_size,
    num_workers=8,
    train_transforms=train_transforms,
    test_transforms=test_transforms,
    val_transforms=test_transforms,
)

class LitCifar100(pl.LightningModule):
    def __init__(self, model_name='resnet18', lr=0.05, factor=0.8):
        super().__init__()
 
        self.save_hyperparameters()
        if model_name == 'resnet18':
            self.model = ResNet18()

    def forward(self, x):
        out = self.model(x)
        return F.log_softmax(out, dim=1)
 
    def training_step(self, batch, batch_idx):
        x, y = batch
        logits = F.log_softmax(self.model(x), dim=1)
        loss = F.nll_loss(logits, y)
        self.log('train_loss', loss)
        return loss
 
    def evaluate(self, batch, stage=None):
        x, y = batch
        logits = self(x)
        loss = F.nll_loss(logits, y)
        preds = torch.argmax(logits, dim=1)
        acc = accuracy(preds, y)
 
        if stage:
            self.log(f'{stage}_loss', loss, prog_bar=True)
            self.log(f'{stage}_acc', acc, prog_bar=True)
 
    def validation_step(self, batch, batch_idx):
        self.evaluate(batch, 'val')
 
    def test_step(self, batch, batch_idx):
        self.evaluate(batch, 'test')
 
    def configure_optimizers(self):
        optimizer = torch.optim.SGD(self.parameters(), lr=self.hparams.lr, momentum=0.9, weight_decay=5e-4)

        return {
          'optimizer': optimizer,
          'lr_scheduler': ReduceLROnPlateau(optimizer, 'max', patience=5, factor=self.hparams.factor, verbose=True, threshold=0.0001, threshold_mode='abs', cooldown=1, min_lr=1e-5),
          'monitor': 'val_acc'
        }

訓練 / 評価

%%time

model = LitCifar100('resnet18', lr=0.05, factor=0.5)
model.datamodule = cifar100_dm
 
trainer = pl.Trainer(
    gpus=2,
    accelerator='dp',
    max_epochs=100,
    progress_bar_refresh_rate=100,
    logger=pl.loggers.TensorBoardLogger('tblogs/', name='resnet18'),
    callbacks=[LearningRateMonitor(logging_interval='step')],
)
 
trainer.fit(model, cifar100_dm)
trainer.test(model, datamodule=cifar100_dm);

GPU available: True, used: True
TPU available: None, using: 0 TPU cores
(...)
  | Name  | Type   | Params
---------------------------------
0 | model | ResNet | 11.2 M
---------------------------------
11.2 M    Trainable params
0         Non-trainable params
11.2 M    Total params
44.881    Total estimated model params size (MB)
(...)
Epoch    35: reducing learning rate of group 0 to 2.5000e-02.
Epoch    42: reducing learning rate of group 0 to 1.2500e-02.
Epoch    62: reducing learning rate of group 0 to 6.2500e-03.
Epoch    88: reducing learning rate of group 0 to 3.1250e-03.
Epoch    95: reducing learning rate of group 0 to 1.5625e-03.
(...)
--------------------------------------------------------------------------------
DATALOADER:0 TEST RESULTS
{'test_acc': 0.7301999926567078, 'test_loss': 1.1373273134231567}
--------------------------------------------------------------------------------
CPU times: user 1h 32min 48s, sys: 20min 14s, total: 1h 53min 2s
Wall time: 1h 39min 34s

以上

2021年2月
月	火	水	木	金	土	日
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