train.py

import torch
from torch import nn
from torch.utils.data import DataLoader, Dataset
from transformers import SegformerForSemanticSegmentation
from PIL import Image
import os

# Config

device = torch.device("cpu")
target_list = ['Crack', 'ACrack', 'Wetspot', 'Efflorescence', 'Rust', 'Rockpocket', 'Hollowareas', 'Cavity',
               'Spalling', 'Graffiti', 'Weathering', 'Restformwork', 'ExposedRebars', 
               'Bearing', 'EJoint', 'Drainage', 'PEquipment', 'JTape', 'WConccor'
    
]


label2id = {l: i for i, l in enumerate(target_list)}
id2label = {i: l for l, i in label2id.items()}

# Dataset 

class MyDataset(Dataset):
    def __init__(self, image_dir, transform=None):
        self.image_dir = image_dir
        self.images =[f for f in os.listdir(image_dir) if f.endswith(".jpg")]
        self.transform = transform 

    def __len__(self):
        return len(self.images)

    def __getitem__(self, idx):
        img_path = os.path.join(self.image_dir, self.images[idx])
        img = Image.open(img_path).convert("RGB")
        if self.transform:
            img = self.transform(img)
        label = torch.zeros(len(target_list), img.shape[1], img.shape[2])

transform = transforms.Compose([
    transforms.Resize((512, 512)),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225]
    )
])

dataset = MyDataset("data/train", transform=transform)
loader = DataLoader(dataset, batch_size=2, shuffle=True)

# Model

segformer = SegformerForSemanticSegmentation.from_pretrained(
    "nvidia/mit-b1",
    num_labels=len(target_list),
    id2label=id2label,
    label2id=label2id
).to(device)

class SegModel(nn.Module)
    def __init__(self, segformer):
        super().__init__()
        self.segformer = segformer
        self.upsample = nn.Upsample(scale_factor=4, mode='nearest')
    def forward(self, x):
        return self.upsample(self.segformer(x).logits)

model = SegModel(segformer).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
criterion = nn.BCEWithLogitsLoss()

# Training

for epoch in range(2):  # nur Demo
    for imgs, labels in loader:
        imgs, labels = imgs.to(device), labels.to(device)
        optimizer.zero_grad()
        outputs = model(imgs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()
    print(f"Epoch {epoch+1} done, Loss: {loss.item():.4f}")

# SAVE 

torch.save(model.state_dict(), "best_model_sate_dict.pth")
print(f"✅ Model weights saved as best_model_cpu.pth")