
# License: BSD
# Author: Sasank Chilamkurthy

from __future__ import print_function, division

import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
import numpy as np
import torchvision
from torchvision import datasets, models, transforms
import matplotlib.pyplot as plt
import time
import os
import copy


# 訓練データ用のデータ拡張と正規化
# 検証データ用には正規化のみ実施
data_transforms = {
    'train': transforms.Compose([
        transforms.RandomResizedCrop(224),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
    'val': transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
}

data_dir = 'data/hymenoptera_data'
image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x),data_transforms[x])
                  for x in ['train', 'val']}
dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=4,
                                             shuffle=True, num_workers=4)
              for x in ['train', 'val']}
dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']}
class_names = image_datasets['train'].classes

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")


def imshow(inp, title=None, filename='a.png'):
    """Imshow for Tensor."""
    inp = inp.numpy().transpose((1, 2, 0))
    mean = np.array([0.485, 0.456, 0.406])
    std = np.array([0.229, 0.224, 0.225])
    inp = std * inp + mean
    inp = np.clip(inp, 0, 1)
    plt.imshow(inp)
    if title is not None:
        plt.title(title)
    plt.pause(0.001)  # プロット図が更新されるように少しだけ一時停止
    plt.savefig(filename)


# 訓練データのバッチを取得する
for inputs, classes in dataloaders['train']:
    break

# バッチからグリッドを作成する
out = torchvision.utils.make_grid(inputs)

imshow(out, title=[class_names[x] for x in classes], filename='data_samples.png')


def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
    since = time.time()

    best_model_wts = copy.deepcopy(model.state_dict())
    best_acc = 0.0

    for epoch in range(num_epochs):
        print('Epoch {}/{}'.format(epoch, num_epochs - 1))
        print('-' * 10)

        # 各エポックには訓練フェーズと検証フェーズがあります
        for phase in ['train', 'val']:
            if phase == 'train':
                model.train()  # モデルを訓練モードに設定します
            else:
                model.eval()   # モードを評価するモデルを設定します

            running_loss = 0.0
            running_corrects = 0

            # データをイレテートします
            for inputs, labels in dataloaders[phase]:
                inputs = inputs.to(device)
                labels = labels.to(device)

                # パラメータの勾配をゼロにします
                optimizer.zero_grad()

                # 順伝播
                # 訓練の時だけ、履歴を保持します
                with torch.set_grad_enabled(phase == 'train'):
                    outputs = model(inputs)
                    _, preds = torch.max(outputs, 1)
                    loss = criterion(outputs, labels)

                    # 訓練の時だけ逆伝播＋オプティマイズを行います
                    if phase == 'train':
                        loss.backward()
                        optimizer.step()

                # 損失を計算します
                running_loss += loss.item() * inputs.size(0)
                running_corrects += torch.sum(preds == labels.data)
            if phase == 'train':
                scheduler.step()

            epoch_loss = running_loss / dataset_sizes[phase]
            epoch_acc = running_corrects.double() / dataset_sizes[phase]

            print('{} Loss: {:.4f} Acc: {:.4f}'.format(
                phase, epoch_loss, epoch_acc))

            # モデルをディープ・コピーします
            if phase == 'val' and epoch_acc > best_acc:
                best_acc = epoch_acc
                best_model_wts = copy.deepcopy(model.state_dict())

        print()

    time_elapsed = time.time() - since
    print('Training complete in {:.0f}m {:.0f}s'.format(
        time_elapsed // 60, time_elapsed % 60))
    print('Best val Acc: {:4f}'.format(best_acc))

    # ベストモデルの重みをロードします
    model.load_state_dict(best_model_wts)
    return model


def visualize_model(model, num_images=6, filename_base='im'):
    was_training = model.training
    model.eval()
    images_so_far = 0
    fig = plt.figure()

    with torch.no_grad():
        for i, (inputs, labels) in enumerate(dataloaders['val']):
            inputs = inputs.to(device)
            labels = labels.to(device)

            outputs = model(inputs)
            _, preds = torch.max(outputs, 1)

            for j in range(inputs.size()[0]):
                images_so_far += 1
                ax = plt.subplot(num_images//2, 2, images_so_far)
                ax.axis('off')
                ax.set_title('predicted: {}'.format(class_names[preds[j]]))

                imshow(inputs.cpu().data[j], filename=filename_base + '.png')

                if images_so_far == num_images:
                    model.train(mode=was_training)
                    return


        model.train(mode=was_training)


model_ft = models.resnet18(pretrained=True)
num_ftrs = model_ft.fc.in_features
# ここでは，各出力サンプルのサイズは2に設定されています
# なお、NN.Linear(num_ftrs, len(class_names))という書き方で一般化することもできます。
model_ft.fc = nn.Linear(num_ftrs, 2)

model_ft = model_ft.to(device)

criterion = nn.CrossEntropyLoss()

# ネットワークのすべてのパラメータが最適化対象です
optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)

# 7エポックごとに学習率を1/10ずつ減衰させます
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)

Downloading: "https://download.pytorch.org/models/resnet18-5c106cde.pth" to /root/.cache/torch/hub/checkpoints/resnet18-5c106cde.pth

HBox(children=(FloatProgress(value=0.0, max=46827520.0), HTML(value='')))


model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler,
                       num_epochs=25)

Epoch 0/24
----------
train Loss: 0.6579 Acc: 0.6598
val Loss: 0.3647 Acc: 0.8693

Epoch 1/24
----------
train Loss: 0.4436 Acc: 0.8115
val Loss: 0.2602 Acc: 0.8889

Epoch 2/24
----------
train Loss: 0.3758 Acc: 0.8525
val Loss: 0.3663 Acc: 0.8627

Epoch 3/24
----------
train Loss: 0.4801 Acc: 0.7746
val Loss: 0.8678 Acc: 0.7190

Epoch 4/24
----------
train Loss: 0.4152 Acc: 0.8484
val Loss: 0.3685 Acc: 0.8758

Epoch 5/24
----------
train Loss: 0.5817 Acc: 0.7992
val Loss: 0.5917 Acc: 0.7974

Epoch 6/24
----------
train Loss: 0.4582 Acc: 0.8443
val Loss: 0.3075 Acc: 0.9150

Epoch 7/24
----------
train Loss: 0.3692 Acc: 0.8443
val Loss: 0.2574 Acc: 0.9216

Epoch 8/24
----------
train Loss: 0.3042 Acc: 0.8648
val Loss: 0.2448 Acc: 0.9085

Epoch 9/24
----------
train Loss: 0.2838 Acc: 0.8811
val Loss: 0.2392 Acc: 0.9281

Epoch 10/24
----------
train Loss: 0.2779 Acc: 0.8811
val Loss: 0.2316 Acc: 0.9216

Epoch 11/24
----------
train Loss: 0.4327 Acc: 0.8361
val Loss: 0.2830 Acc: 0.8954

Epoch 12/24
----------
train Loss: 0.3345 Acc: 0.8566
val Loss: 0.2187 Acc: 0.9150

Epoch 13/24
----------
train Loss: 0.3179 Acc: 0.8607
val Loss: 0.2028 Acc: 0.9216

Epoch 14/24
----------
train Loss: 0.3340 Acc: 0.8484
val Loss: 0.2109 Acc: 0.9281

Epoch 15/24
----------
train Loss: 0.3628 Acc: 0.8238
val Loss: 0.2308 Acc: 0.9085

Epoch 16/24
----------
train Loss: 0.2692 Acc: 0.8566
val Loss: 0.2209 Acc: 0.9216

Epoch 17/24
----------
train Loss: 0.3819 Acc: 0.8361
val Loss: 0.2053 Acc: 0.9085

Epoch 18/24
----------
train Loss: 0.2927 Acc: 0.8770
val Loss: 0.2768 Acc: 0.9020

Epoch 19/24
----------
train Loss: 0.2942 Acc: 0.8566
val Loss: 0.2131 Acc: 0.9216

Epoch 20/24
----------
train Loss: 0.2187 Acc: 0.9139
val Loss: 0.2271 Acc: 0.9150

Epoch 21/24
----------
train Loss: 0.2382 Acc: 0.8852
val Loss: 0.2285 Acc: 0.9216

Epoch 22/24
----------
train Loss: 0.2750 Acc: 0.8484
val Loss: 0.2348 Acc: 0.9281

Epoch 23/24
----------
train Loss: 0.2746 Acc: 0.8730
val Loss: 0.2103 Acc: 0.9216

Epoch 24/24
----------
train Loss: 0.2915 Acc: 0.8730
val Loss: 0.2610 Acc: 0.9020

Training complete in 1m 37s
Best val Acc: 0.928105


visualize_model(model_ft, filename_base='a')


model_conv = torchvision.models.resnet18(pretrained=True)
for param in model_conv.parameters():
    param.requires_grad = False

# 新しく構築されたモジュールのパラメータは、デフォルトでは requires_grad=True になっています。
num_ftrs = model_conv.fc.in_features
model_conv.fc = nn.Linear(num_ftrs, 2)

model_conv = model_conv.to(device)

criterion = nn.CrossEntropyLoss()

 # ファインチューニングとは違い、最終層のパラメータのみが最適化されていることを確認してください
optimizer_conv = optim.SGD(model_conv.fc.parameters(), lr=0.001, momentum=0.9)

# 7エポックごとに学習率を0.1倍ずつ減衰させる
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_conv, step_size=7, gamma=0.1)


model_conv = train_model(model_conv, criterion, optimizer_conv,
                         exp_lr_scheduler, num_epochs=25)

Epoch 0/24
----------
train Loss: 0.5942 Acc: 0.7049
val Loss: 0.2298 Acc: 0.9281

Epoch 1/24
----------
train Loss: 0.3885 Acc: 0.8238
val Loss: 0.2044 Acc: 0.9346

Epoch 2/24
----------
train Loss: 0.3899 Acc: 0.8238
val Loss: 0.2545 Acc: 0.9020

Epoch 3/24
----------
train Loss: 0.5229 Acc: 0.7582
val Loss: 0.2526 Acc: 0.8954

Epoch 4/24
----------
train Loss: 0.4906 Acc: 0.8115
val Loss: 0.1795 Acc: 0.9477

Epoch 5/24
----------
train Loss: 0.3060 Acc: 0.8852
val Loss: 0.2763 Acc: 0.8824

Epoch 6/24
----------
train Loss: 0.4984 Acc: 0.7377
val Loss: 0.2357 Acc: 0.8954

Epoch 7/24
----------
train Loss: 0.4188 Acc: 0.8033
val Loss: 0.1752 Acc: 0.9412

Epoch 8/24
----------
train Loss: 0.3272 Acc: 0.8648
val Loss: 0.2000 Acc: 0.9150

Epoch 9/24
----------
train Loss: 0.2764 Acc: 0.8934
val Loss: 0.1667 Acc: 0.9608

Epoch 10/24
----------
train Loss: 0.3596 Acc: 0.8402
val Loss: 0.2018 Acc: 0.9216

Epoch 11/24
----------
train Loss: 0.3567 Acc: 0.8730
val Loss: 0.1905 Acc: 0.9150

Epoch 12/24
----------
train Loss: 0.3074 Acc: 0.8525
val Loss: 0.1652 Acc: 0.9542

Epoch 13/24
----------
train Loss: 0.4013 Acc: 0.8279
val Loss: 0.1610 Acc: 0.9542

Epoch 14/24
----------
train Loss: 0.2848 Acc: 0.8730
val Loss: 0.1622 Acc: 0.9477

Epoch 15/24
----------
train Loss: 0.3529 Acc: 0.8648
val Loss: 0.1757 Acc: 0.9477

Epoch 16/24
----------
train Loss: 0.3388 Acc: 0.8279
val Loss: 0.1842 Acc: 0.9542

Epoch 17/24
----------
train Loss: 0.3896 Acc: 0.8443
val Loss: 0.1590 Acc: 0.9608

Epoch 18/24
----------
train Loss: 0.4212 Acc: 0.8279
val Loss: 0.2018 Acc: 0.9216

Epoch 19/24
----------
train Loss: 0.3185 Acc: 0.8402
val Loss: 0.1690 Acc: 0.9412

Epoch 20/24
----------
train Loss: 0.3104 Acc: 0.8525
val Loss: 0.1745 Acc: 0.9542

Epoch 21/24
----------
train Loss: 0.3066 Acc: 0.8607
val Loss: 0.1720 Acc: 0.9542

Epoch 22/24
----------
train Loss: 0.3096 Acc: 0.8648
val Loss: 0.1673 Acc: 0.9542

Epoch 23/24
----------
train Loss: 0.3565 Acc: 0.8238
val Loss: 0.1631 Acc: 0.9542

Epoch 24/24
----------
train Loss: 0.3169 Acc: 0.8689
val Loss: 0.1737 Acc: 0.9608

Training complete in 1m 23s
Best val Acc: 0.960784


visualize_model(model_conv, filename_base='b')

「画像分類タスクに対する転移学習の方法」¶

データの読み込み¶

まず，ターミナルで以下のコマンドを実行してデータをダウンロード・解凍

画像の可視化¶

モデルの訓練¶

モデル予測値の可視化¶

ConvNetをファインチューニングする方法¶

訓練と評価¶

Conv Netを特徴抽出器として使う方法¶

訓練と評価¶

さらなる学習のために¶