PyTorch Lightning 1.1: research : CIFAR100 (RegNet)
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 : 02/25/2021 (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：
- notebooks : PyTorch Lightning CIFAR10 ~94% Baseline Tutorial
- Train CIFAR10 with PyTorch
- Pytorch-cifar100
* 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 (RegNet)
RegNetX_200MF
仕様
- Total params: 2,355,156 (2.4M)
- Trainable params: 2,355,156
- Non-trainable params: 0
結果
- RegNetX_200MF
- {‘test_acc’: 0.7339000105857849, ‘test_loss’: 1.152081847190857}
- 100 エポック ; Wall time: 2h 8min 30s
- Tesla T4
- ReduceLROnPlateau
RegNetX_400MF
仕様
- Total params: 4,813,988 (4.8M)
- Trainable params: 4,813,988
- Non-trainable params: 0
結果
- RegNetX_400MF
- {‘test_acc’: 0.732200026512146, ‘test_loss’: 1.1259433031082153}
- 100 エポック ; Wall time: Wall time: 3h 19min 18s
- Tesla T4
- ReduceLROnPlateau
RegNetY_400MF
仕様
- Total params: 5,749,012 (5.7M)
- Trainable params: 5,749,012
- Non-trainable params: 0
結果
- RegNetY_400MF
- {‘test_acc’: 0.7128999829292297, ‘test_loss’: 1.2059353590011597}
- 100 エポック ; Wall time: Wall time: 3h 47min 37s
- 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 SE(nn.Module):
'''Squeeze-and-Excitation block.'''
def __init__(self, in_planes, se_planes):
super(SE, self).__init__()
self.se1 = nn.Conv2d(in_planes, se_planes, kernel_size=1, bias=True)
self.se2 = nn.Conv2d(se_planes, in_planes, kernel_size=1, bias=True)
def forward(self, x):
out = F.adaptive_avg_pool2d(x, (1, 1))
out = F.relu(self.se1(out))
out = self.se2(out).sigmoid()
out = x * out
return out
class Block(nn.Module):
def __init__(self, w_in, w_out, stride, group_width, bottleneck_ratio, se_ratio):
super(Block, self).__init__()
# 1x1
w_b = int(round(w_out * bottleneck_ratio))
self.conv1 = nn.Conv2d(w_in, w_b, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(w_b)
# 3x3
num_groups = w_b // group_width
self.conv2 = nn.Conv2d(w_b, w_b, kernel_size=3,
stride=stride, padding=1, groups=num_groups, bias=False)
self.bn2 = nn.BatchNorm2d(w_b)
# se
self.with_se = se_ratio > 0
if self.with_se:
w_se = int(round(w_in * se_ratio))
self.se = SE(w_b, w_se)
# 1x1
self.conv3 = nn.Conv2d(w_b, w_out, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(w_out)
self.shortcut = nn.Sequential()
if stride != 1 or w_in != w_out:
self.shortcut = nn.Sequential(
nn.Conv2d(w_in, w_out,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(w_out)
)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = F.relu(self.bn2(self.conv2(out)))
if self.with_se:
out = self.se(out)
out = self.bn3(self.conv3(out))
out += self.shortcut(x)
out = F.relu(out)
return out
class RegNet(nn.Module):
def __init__(self, cfg, num_classes=100):
super(RegNet, self).__init__()
self.cfg = cfg
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(0)
self.layer2 = self._make_layer(1)
self.layer3 = self._make_layer(2)
self.layer4 = self._make_layer(3)
self.linear = nn.Linear(self.cfg['widths'][-1], num_classes)
def _make_layer(self, idx):
depth = self.cfg['depths'][idx]
width = self.cfg['widths'][idx]
stride = self.cfg['strides'][idx]
group_width = self.cfg['group_width']
bottleneck_ratio = self.cfg['bottleneck_ratio']
se_ratio = self.cfg['se_ratio']
layers = []
for i in range(depth):
s = stride if i == 0 else 1
layers.append(Block(self.in_planes, width,
s, group_width, bottleneck_ratio, se_ratio))
self.in_planes = width
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.adaptive_avg_pool2d(out, (1, 1))
out = out.view(out.size(0), -1)
out = self.linear(out)
return out
def RegNetX_200MF():
cfg = {
'depths': [1, 1, 4, 7],
'widths': [24, 56, 152, 368],
'strides': [1, 1, 2, 2],
'group_width': 8,
'bottleneck_ratio': 1,
'se_ratio': 0,
}
return RegNet(cfg)
def RegNetX_400MF():
cfg = {
'depths': [1, 2, 7, 12],
'widths': [32, 64, 160, 384],
'strides': [1, 1, 2, 2],
'group_width': 16,
'bottleneck_ratio': 1,
'se_ratio': 0,
}
return RegNet(cfg)
def RegNetY_400MF():
cfg = {
'depths': [1, 2, 7, 12],
'widths': [32, 64, 160, 384],
'strides': [1, 1, 2, 2],
'group_width': 16,
'bottleneck_ratio': 1,
'se_ratio': 0.25,
}
return RegNet(cfg)
net = RegNetX_200MF() print(net) y = net(torch.randn(1, 3, 32, 32)) print(y.size())
RegNet(
(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): Block(
(conv1): Conv2d(64, 24, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(24, 24, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=3, bias=False)
(bn2): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(24, 24, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn3): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(shortcut): Sequential(
(0): Conv2d(64, 24, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
)
(layer2): Sequential(
(0): Block(
(conv1): Conv2d(24, 56, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(56, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(56, 56, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=7, bias=False)
(bn2): BatchNorm2d(56, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(56, 56, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn3): BatchNorm2d(56, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(shortcut): Sequential(
(0): Conv2d(24, 56, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(56, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
)
(layer3): Sequential(
(0): Block(
(conv1): Conv2d(56, 152, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(152, 152, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=19, bias=False)
(bn2): BatchNorm2d(152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(152, 152, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn3): BatchNorm2d(152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(shortcut): Sequential(
(0): Conv2d(56, 152, kernel_size=(1, 1), stride=(2, 2), bias=False)
(1): BatchNorm2d(152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(1): Block(
(conv1): Conv2d(152, 152, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(152, 152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=19, bias=False)
(bn2): BatchNorm2d(152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(152, 152, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn3): BatchNorm2d(152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(shortcut): Sequential()
)
(2): Block(
(conv1): Conv2d(152, 152, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(152, 152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=19, bias=False)
(bn2): BatchNorm2d(152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(152, 152, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn3): BatchNorm2d(152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(shortcut): Sequential()
)
(3): Block(
(conv1): Conv2d(152, 152, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(152, 152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=19, bias=False)
(bn2): BatchNorm2d(152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(152, 152, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn3): BatchNorm2d(152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(shortcut): Sequential()
)
)
(layer4): Sequential(
(0): Block(
(conv1): Conv2d(152, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(368, 368, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=46, bias=False)
(bn2): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(368, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn3): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(shortcut): Sequential(
(0): Conv2d(152, 368, kernel_size=(1, 1), stride=(2, 2), bias=False)
(1): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(1): Block(
(conv1): Conv2d(368, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(368, 368, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=46, bias=False)
(bn2): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(368, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn3): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(shortcut): Sequential()
)
(2): Block(
(conv1): Conv2d(368, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(368, 368, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=46, bias=False)
(bn2): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(368, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn3): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(shortcut): Sequential()
)
(3): Block(
(conv1): Conv2d(368, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(368, 368, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=46, bias=False)
(bn2): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(368, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn3): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(shortcut): Sequential()
)
(4): Block(
(conv1): Conv2d(368, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(368, 368, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=46, bias=False)
(bn2): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(368, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn3): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(shortcut): Sequential()
)
(5): Block(
(conv1): Conv2d(368, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(368, 368, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=46, bias=False)
(bn2): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(368, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn3): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(shortcut): Sequential()
)
(6): Block(
(conv1): Conv2d(368, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(368, 368, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=46, bias=False)
(bn2): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(368, 368, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn3): BatchNorm2d(368, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(shortcut): Sequential()
)
)
(linear): Linear(in_features=368, out_features=100, bias=True)
)
torch.Size([1, 100])
from torchsummary import summary
summary(RegNetX_200MF().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, 24, 32, 32] 1,536
BatchNorm2d-4 [-1, 24, 32, 32] 48
Conv2d-5 [-1, 24, 32, 32] 1,728
BatchNorm2d-6 [-1, 24, 32, 32] 48
Conv2d-7 [-1, 24, 32, 32] 576
BatchNorm2d-8 [-1, 24, 32, 32] 48
Conv2d-9 [-1, 24, 32, 32] 1,536
BatchNorm2d-10 [-1, 24, 32, 32] 48
Block-11 [-1, 24, 32, 32] 0
Conv2d-12 [-1, 56, 32, 32] 1,344
BatchNorm2d-13 [-1, 56, 32, 32] 112
Conv2d-14 [-1, 56, 32, 32] 4,032
BatchNorm2d-15 [-1, 56, 32, 32] 112
Conv2d-16 [-1, 56, 32, 32] 3,136
BatchNorm2d-17 [-1, 56, 32, 32] 112
Conv2d-18 [-1, 56, 32, 32] 1,344
BatchNorm2d-19 [-1, 56, 32, 32] 112
Block-20 [-1, 56, 32, 32] 0
Conv2d-21 [-1, 152, 32, 32] 8,512
BatchNorm2d-22 [-1, 152, 32, 32] 304
Conv2d-23 [-1, 152, 16, 16] 10,944
BatchNorm2d-24 [-1, 152, 16, 16] 304
Conv2d-25 [-1, 152, 16, 16] 23,104
BatchNorm2d-26 [-1, 152, 16, 16] 304
Conv2d-27 [-1, 152, 16, 16] 8,512
BatchNorm2d-28 [-1, 152, 16, 16] 304
Block-29 [-1, 152, 16, 16] 0
Conv2d-30 [-1, 152, 16, 16] 23,104
BatchNorm2d-31 [-1, 152, 16, 16] 304
Conv2d-32 [-1, 152, 16, 16] 10,944
BatchNorm2d-33 [-1, 152, 16, 16] 304
Conv2d-34 [-1, 152, 16, 16] 23,104
BatchNorm2d-35 [-1, 152, 16, 16] 304
Block-36 [-1, 152, 16, 16] 0
Conv2d-37 [-1, 152, 16, 16] 23,104
BatchNorm2d-38 [-1, 152, 16, 16] 304
Conv2d-39 [-1, 152, 16, 16] 10,944
BatchNorm2d-40 [-1, 152, 16, 16] 304
Conv2d-41 [-1, 152, 16, 16] 23,104
BatchNorm2d-42 [-1, 152, 16, 16] 304
Block-43 [-1, 152, 16, 16] 0
Conv2d-44 [-1, 152, 16, 16] 23,104
BatchNorm2d-45 [-1, 152, 16, 16] 304
Conv2d-46 [-1, 152, 16, 16] 10,944
BatchNorm2d-47 [-1, 152, 16, 16] 304
Conv2d-48 [-1, 152, 16, 16] 23,104
BatchNorm2d-49 [-1, 152, 16, 16] 304
Block-50 [-1, 152, 16, 16] 0
Conv2d-51 [-1, 368, 16, 16] 55,936
BatchNorm2d-52 [-1, 368, 16, 16] 736
Conv2d-53 [-1, 368, 8, 8] 26,496
BatchNorm2d-54 [-1, 368, 8, 8] 736
Conv2d-55 [-1, 368, 8, 8] 135,424
BatchNorm2d-56 [-1, 368, 8, 8] 736
Conv2d-57 [-1, 368, 8, 8] 55,936
BatchNorm2d-58 [-1, 368, 8, 8] 736
Block-59 [-1, 368, 8, 8] 0
Conv2d-60 [-1, 368, 8, 8] 135,424
BatchNorm2d-61 [-1, 368, 8, 8] 736
Conv2d-62 [-1, 368, 8, 8] 26,496
BatchNorm2d-63 [-1, 368, 8, 8] 736
Conv2d-64 [-1, 368, 8, 8] 135,424
BatchNorm2d-65 [-1, 368, 8, 8] 736
Block-66 [-1, 368, 8, 8] 0
Conv2d-67 [-1, 368, 8, 8] 135,424
BatchNorm2d-68 [-1, 368, 8, 8] 736
Conv2d-69 [-1, 368, 8, 8] 26,496
BatchNorm2d-70 [-1, 368, 8, 8] 736
Conv2d-71 [-1, 368, 8, 8] 135,424
BatchNorm2d-72 [-1, 368, 8, 8] 736
Block-73 [-1, 368, 8, 8] 0
Conv2d-74 [-1, 368, 8, 8] 135,424
BatchNorm2d-75 [-1, 368, 8, 8] 736
Conv2d-76 [-1, 368, 8, 8] 26,496
BatchNorm2d-77 [-1, 368, 8, 8] 736
Conv2d-78 [-1, 368, 8, 8] 135,424
BatchNorm2d-79 [-1, 368, 8, 8] 736
Block-80 [-1, 368, 8, 8] 0
Conv2d-81 [-1, 368, 8, 8] 135,424
BatchNorm2d-82 [-1, 368, 8, 8] 736
Conv2d-83 [-1, 368, 8, 8] 26,496
BatchNorm2d-84 [-1, 368, 8, 8] 736
Conv2d-85 [-1, 368, 8, 8] 135,424
BatchNorm2d-86 [-1, 368, 8, 8] 736
Block-87 [-1, 368, 8, 8] 0
Conv2d-88 [-1, 368, 8, 8] 135,424
BatchNorm2d-89 [-1, 368, 8, 8] 736
Conv2d-90 [-1, 368, 8, 8] 26,496
BatchNorm2d-91 [-1, 368, 8, 8] 736
Conv2d-92 [-1, 368, 8, 8] 135,424
BatchNorm2d-93 [-1, 368, 8, 8] 736
Block-94 [-1, 368, 8, 8] 0
Conv2d-95 [-1, 368, 8, 8] 135,424
BatchNorm2d-96 [-1, 368, 8, 8] 736
Conv2d-97 [-1, 368, 8, 8] 26,496
BatchNorm2d-98 [-1, 368, 8, 8] 736
Conv2d-99 [-1, 368, 8, 8] 135,424
BatchNorm2d-100 [-1, 368, 8, 8] 736
Block-101 [-1, 368, 8, 8] 0
Linear-102 [-1, 100] 36,900
================================================================
Total params: 2,355,156
Trainable params: 2,355,156
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.01
Forward/backward pass size (MB): 27.56
Params size (MB): 8.98
Estimated Total Size (MB): 36.55
----------------------------------------------------------------
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 = 50
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, lr=0.05, factor=0.8):
super().__init__()
self.save_hyperparameters()
self.model = RegNetX_200MF()
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(lr=0.05, factor=0.5)
model.datamodule = cifar100_dm
trainer = pl.Trainer(
gpus=1,
max_epochs=100,
progress_bar_refresh_rate=100,
logger=pl.loggers.TensorBoardLogger('tblogs/', name='regnetx_200mf'),
callbacks=[LearningRateMonitor(logging_interval='step')],
)
trainer.fit(model, cifar100_dm)
trainer.test(model, datamodule=cifar100_dm);
| Name | Type | Params
---------------------------------
0 | model | RegNet | 2.4 M
---------------------------------
2.4 M Trainable params
0 Non-trainable params
2.4 M Total params
9.421 Total estimated model params size (MB)
(...)
Epoch 33: reducing learning rate of group 0 to 2.5000e-02.
Epoch 40: reducing learning rate of group 0 to 1.2500e-02.
Epoch 47: reducing learning rate of group 0 to 6.2500e-03.
Epoch 56: reducing learning rate of group 0 to 3.1250e-03.
Epoch 72: reducing learning rate of group 0 to 1.5625e-03.
Epoch 79: reducing learning rate of group 0 to 7.8125e-04.
Epoch 95: reducing learning rate of group 0 to 3.9063e-04.
(...)
-------------------------------------------------------------------------------
DATALOADER:0 TEST RESULTS
{'test_acc': 0.7339000105857849, 'test_loss': 1.152081847190857}
--------------------------------------------------------------------------------
CPU times: user 1h 46min 25s, sys: 18min 45s, total: 2h 5min 10s
Wall time: 2h 8min 30s
以上