[PyTorch] - Intermediate Deep Learning with PyTorch - Images & Convolutional Neural Networks

Images & Convolutional Neural Networks

 

Data Augmentation

image data는 다른 데이터들에 비해서 개수가 부족한데, 인위적으로 flip, rotate 등을 통해 augmentataion을 줄 수 있다.

---

 

CNN(Convolutional Neural Network)

요런 필터가 이미지를 돌아다니면서 픽셀별로 정보를 수집함

입력의 공간 패턴을 보존하고, nn.Linear()보다 적은 parameter을 사용함

합성곱(Convolution)의 과정 : input patch와 Filter을 element-wise dot product를 해서 sum

Padding 추가를 통해 공간 차원 보존, 테두리 영역 보존 등의 역할 수행. nn.Conv2d()에 parameter padding=n으로 부여

class Net(nn.Module):
    def __init__(self, num_classes):
        super().__init__()
        # Define feature extractor, 보통 2번 수행
        self.feature_extractor = nn.Sequential(
            nn.Conv2d(3, 32, kernel_size=3, padding=1),
            nn.ELU(),
            nn.MaxPool2d(kernel_size=2),
            nn.Conv2d(32, 64, kernel_size=3, padding=1),
            nn.ELU(),
            nn.MaxPool2d(kernel_size=2),
            nn.Flatten(),
        )
        # Define classifier
        self.classifier = nn.Linear(64*16*16, num_classes)
    
    def forward(self, x):  
        # Pass input through feature extractor and classifier
        x = self.feature_extractor(x)
        x = self.classifier(x)
        return x

 

Image Classifier Training loop

# Define the model
net = Net(num_classes=7)
# Define the loss function
criterion = nn.CrossEntropyLoss()
# Define the optimizer
optimizer = optim.Adam(net.parameters(), lr=0.001)

for epoch in range(3):
    running_loss = 0.0
    # Iterate over training batches
    for images, labels in dataloader_train:
        optimizer.zero_grad()
        outputs = net(images)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()
        running_loss += loss.item()
    
    epoch_loss = running_loss / len(dataloader_train)
    print(f"Epoch {epoch+1}, Loss: {epoch_loss:.4f}")

---

Precision & Recall

Precision : positive로 예측한 것들 중 실제로 positive인 것들

= Fraction of correct positive predictions

Recall : 맞게 예측한 것들 중 positive인 것들

= Fraction of all positive examples correctly predicted

 

각 클래스별로 개별적으로 분석하는 방법

Macro average

Micro average

Weighted average

 

 

Multi-class model evaluation

# Define metrics
metric_precision = Precision(task="multiclass", num_classes=7, average="micro")
metric_recall = Recall(task="multiclass", num_classes=7, average="micro")

net.eval()
with torch.no_grad():
    for images, labels in dataloader_test:
        outputs = net(images)
        _, preds = torch.max(outputs, 1)
        metric_precision(preds, labels)
        metric_recall(preds, labels)

precision = metric_precision.compute()
recall = metric_recall.compute()
print(f"Precision: {precision}")
print(f"Recall: {recall}")

Analyzing metrics per class

# Define precision metric
metric_precision = Precision(
    task="multiclass", num_classes=7, average=None
)

net.eval()
with torch.no_grad():
    for images, labels in dataloader_test:
        outputs = net(images)
        _, preds = torch.max(outputs, 1)
        metric_precision(preds, labels)
precision = metric_precision.compute()

# Get precision per class
precision_per_class = {
    k: precision[v].item()
    for k, v 
    in dataset_test.class_to_idx.items()
}
print(precision_per_class)