utils.py

import torch
from torchvision.models.detection import FasterRCNN
from collections import defaultdict

def collate_fn(batch):
    return tuple(zip(*batch))

def train_one_epoch(model, optimizer, data_loader, device, epoch):
    model.train()
    metric_logger = defaultdict(float)
    for images, targets in data_loader:
        images = [image.to(device) for image in images]
        targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
        
        # Zero the gradients
        optimizer.zero_grad()
        
        # Forward pass
        loss_dict = model(images, targets)
        
        losses = sum(loss for loss in loss_dict.values())
        
        # Backpropagation
        losses.backward()
        optimizer.step()

        # Log training progress
        metric_logger["loss"] += losses.item()

def evaluate(model, data_loader, device):
    model.eval()
    for images, targets in data_loader:
        images = [image.to(device) for image in images]
        targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
        
        with torch.no_grad():
            predictions = model(images)

        # Evaluate predictions (e.g., IOU, detection accuracy)
        # Use suitable evaluation metrics for your dataset