GitHub - miyamotok0105/pytorch_handbook: pytorch_handbook
先日の CNN(AlexNet) エントリの続きで、やはり上記urlの写経。
前回は、CIFAR-10 https://www.cs.toronto.edu/~kriz/cifar.html というバイナリ?で用意されたデータを使用しましたが、 今回は、 https://download.pytorch.org/tutorial/hymenoptera_data.zip にある アリとハチの画像を分類します。
画像のサンプルは以下。
#!/usr/local/python3/bin/python3 # -*- coding: utf-8 -*- import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler from torch.autograd import Variable import numpy as np import torchvision from torchvision import datasets, models, transforms # import matplotlib.pyplot as plt import matplotlib matplotlib.use('Agg') import matplotlib.pylab as plt from PIL import Image import time import os import cv2 from PIL import Image DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu' root = 'hymenoptera_data' num_classes = 2 TRAIN_NUM_EPOCHS = 500 fc_size = 9216 def main(): print("DEVICE:", DEVICE) (data_transforms, to_tensor_transforms) = get_pre_process() # 定義したDatasetとDataLoaderを使います。 custom_train_dataset = CustomDataset(root, data_transforms["train"], train=True) train_loader = torch.utils.data.DataLoader(dataset=custom_train_dataset, batch_size=5, shuffle=True) custom_test_dataset = CustomDataset(root, data_transforms["val"]) test_loader = torch.utils.data.DataLoader(dataset=custom_test_dataset, batch_size=5, shuffle=False) # for i, (images, labels) in enumerate(train_loader): # print(images.size()) # print(images[0].size()) # print(labels[0].item()) # #ここに訓練などの処理をきます。 # break global fc_size # batch_size=10, channel=3, size=224x224 fc_size = get_fc_size( torch.FloatTensor(10, 3, 224, 224) ) print("fc_size:",fc_size) net = AlexNet(num_classes, fc_size).to(DEVICE) criterion = nn.CrossEntropyLoss() optimizer = optim.SGD(net.parameters(), lr=0.01, momentum=0.9, weight_decay=5e-4) (train_loss_list, train_acc_list, val_loss_list, val_acc_list) = \ train_epochs(train_loader, test_loader, net, criterion, optimizer) output_to_file(train_loss_list,train_acc_list,val_loss_list,val_acc_list) def train_epochs(train_loader, test_loader, net, criterion, optimizer): train_loss_list = [] train_acc_list = [] val_loss_list = [] val_acc_list = [] for epoch in range(TRAIN_NUM_EPOCHS): train_loss = 0 train_acc = 0 val_loss = 0 val_acc = 0 #train net.train() for i, (images, labels) in enumerate(train_loader): images, labels = images.to(DEVICE), labels.to(DEVICE) optimizer.zero_grad() outputs = net(images) loss = criterion(outputs, labels) train_loss += loss.item() train_acc += (outputs.max(1)[1] == labels).sum().item() loss.backward() optimizer.step() avg_train_loss = train_loss / len(train_loader.dataset) avg_train_acc = train_acc / len(train_loader.dataset) #val net.eval() with torch.no_grad(): for images, labels in test_loader: images = images.to(DEVICE) labels = labels.to(DEVICE) outputs = net(images) loss = criterion(outputs, labels) val_loss += loss.item() val_acc += (outputs.max(1)[1] == labels).sum().item() avg_val_loss = val_loss / len(test_loader.dataset) avg_val_acc = val_acc / len(test_loader.dataset) output_str_format = ', '.join(['Epoch [{}/{}]','Loss: {loss:.4f}', 'val_loss: {val_loss:.4f}','val_acc: {val_acc:.4f}']) print (output_str_format.format(epoch+1, TRAIN_NUM_EPOCHS, i+1, loss=avg_train_loss, val_loss=avg_val_loss, val_acc=avg_val_acc)) train_loss_list.append(avg_train_loss) train_acc_list.append(avg_train_acc) val_loss_list.append(avg_val_loss) val_acc_list.append(avg_val_acc) return train_loss_list, train_acc_list, val_loss_list, val_acc_list def show_img(img): npimg = img.numpy() # ↓この行は理解できていません # plt.imshow(np.transpose(npimg, (1,2,0)), interpolation='nearest') def get_pre_process(): #画像の前処理 data_transforms = { 'train': transforms.Compose([ transforms.RandomResizedCrop(224), # re-size transforms.RandomHorizontalFlip(), # 反転 transforms.ToTensor(), transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]) #正規化 ]), 'val': transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]) ]), } #正規化をしない前処理 to_tensor_transforms = transforms.Compose([ transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor() ]) return data_transforms , to_tensor_transforms def get_nn_features(): features = nn.Sequential( nn.Conv2d(3, 64, kernel_size=11, stride=4, padding=2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), nn.Conv2d(64, 192, kernel_size=5, padding=2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), nn.Conv2d(192, 384, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.Conv2d(384, 256, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.Conv2d(256, 256, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) return features def get_fc_size( size_check ): features = get_nn_features() #バッチサイズ10, 6×6のフィルターが256枚 #10バッチは残して、6×6×256を１次元に落とす=>6×6×256=9216 print("size1:",features(size_check).size()) #バッチ１０の値を軸にして残りの次元を１次元へ落とした場合の #Tensorの形状をチェックすると9216。 print("size2:",features(size_check).view(size_check.size(0), -1).size()) #fc_sizeを全結合の形状として保持 global fc_size fc_size = features(size_check).view(size_check.size(0), -1).size()[1] print("size3:",fc_size) return fc_size def output_to_file(train_loss_list,train_acc_list,val_loss_list,val_acc_list): plt.figure() plt.plot(range(TRAIN_NUM_EPOCHS), train_loss_list, color='blue', linestyle='-', label='train_loss') plt.plot(range(TRAIN_NUM_EPOCHS), val_loss_list, color='green', linestyle='--', label='val_loss') # plt.ylim([0,0.04]) plt.legend() plt.xlabel('epoch') plt.ylabel('loss') plt.title('Training and validation loss') plt.grid() plt.savefig( 'test_4_1.png' ) plt.figure() plt.plot(range(TRAIN_NUM_EPOCHS), train_acc_list, color='blue', linestyle='-', label='train_acc') plt.plot(range(TRAIN_NUM_EPOCHS), val_acc_list, color='green', linestyle='--', label='val_acc') plt.ylim([0,1]) plt.legend() plt.xlabel('epoch') plt.ylabel('acc') plt.title('Training and validation accuracy') plt.grid() plt.savefig( 'test_4_2.png' ) class CustomDataset(torch.utils.data.Dataset): classes = ['ant', 'bee'] def __init__(self, root, transform=None, train=True): self.transform = transform # 指定する場合、前処理クラスを受取り self.images = [] # 画像とlabelの保持用 self.labels = [] # 訓練と検証で、読込むpathを取得 if train == True: root_ants_path = os.path.join(root, 'train', 'ants') root_bees_path = os.path.join(root, 'train', 'bees') else: root_ants_path = os.path.join(root, 'val', 'ants') root_bees_path = os.path.join(root, 'val', 'bees') ant_images = os.listdir(root_ants_path) # アリの画像一覧を取得 ant_labels = [0] * len(ant_images) # アリをlabel=0に指定 bee_images = os.listdir(root_bees_path) # ハチの画像一覧を取得 bee_labels = [1] * len(bee_images) # ハチをlabel=1に指定 # listのmerge for image, label in zip(ant_images, ant_labels): self.images.append(os.path.join(root_ants_path, image)) self.labels.append(label) for image, label in zip(bee_images, bee_labels): self.images.append(os.path.join(root_bees_path, image)) self.labels.append(label) def __getitem__(self, index): image = self.images[index] # indexを元に画像のpathとlabel取得 label = self.labels[index] with open(image, 'rb') as f: # pathから画像読込み image = Image.open(f) image = image.convert('RGB') if self.transform is not None: # 前処理がある場合は入れる image = self.transform(image) return image, label def __len__(self): # ここにはデータ数を指定します。 return len(self.images) class AlexNet(nn.Module): def __init__(self, num_classes, fc_size): super(AlexNet, self).__init__() self.features = get_nn_features() self.classifier = nn.Sequential( nn.Dropout(p=0.5), nn.Linear(fc_size, 4096), #fc_sizeで計算した形状を指定 nn.ReLU(inplace=True), nn.Dropout(p=0.5), nn.Linear(4096, 4096), nn.ReLU(inplace=True), nn.Linear(4096, num_classes) ) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.classifier(x) return x if __name__ == '__main__': main()
↑こう書くと、↓こう出力されます。
手元のcent os 8 でも実行しましたが、GPUがなく時間がかかる為、 google colaboratory でも実行しています。
$ wget https://download.pytorch.org/tutorial/hymenoptera_data.zip $ unzip hymenoptera_data.zip $ ./foo.py DEVICE: cuda size1: torch.Size([10, 256, 6, 6]) size2: torch.Size([10, 9216]) size3: 9216 fc_size: 9216 Epoch [1/500], Loss: 0.1392, val_loss: 0.1386, val_acc: 0.5061 Epoch [2/500], Loss: 0.1394, val_loss: 0.1386, val_acc: 0.5061 : Epoch [500/500], Loss: 0.0758, val_loss: 0.0612, val_acc: 0.8653
