Multi-class U-net segmentation: target_var got float instead of int/long values after converting to tensors

I’m trying to adjust a binary segmentation U-net model, to be able to train a multi-class U-net on the German Asfalt Pavement Distress (GAPs) dataset.

The code files, and the dataset are available through the following link:
https://drive.google.com/drive/folders/14NQdtMXokIixBJ5XizexVECn23Jh9aTM?usp=sharing

The following is the code in “train_unet_GAPs.py”:

import torch
from torch import nn
from unet.unet_transfer import UNet16, UNetResNet
from pathlib import Path
import torchvision.transforms as transforms
from torch.utils.data import DataLoader, Dataset, random_split
import torch.nn.functional as F
from torch.autograd import Variable
import shutil
from data_loader import ImgDataSet
import os
import argparse
import tqdm
import numpy as np
import scipy.ndimage as ndimage

class AverageMeter(object):
    """Computes and stores the average and current value"""
    def __init__(self):
        self.reset()

    def reset(self):
        self.val = 0
        self.avg = 0
        self.sum = 0
        self.count = 0

    def update(self, val, n=1):
        self.val = val
        self.sum += val * n
        self.count += n
        self.avg = self.sum / self.count

def create_model(device, type ='vgg16'):
    if type == 'vgg16':
        print('create vgg16 model')
        model = UNet16(pretrained=True)
    elif type == 'resnet101':
        encoder_depth = 101
        num_classes = 8
        print('create resnet101 model')
        model = UNetResNet(encoder_depth=encoder_depth, num_classes=num_classes, pretrained=True)
    elif type == 'resnet34':
        encoder_depth = 34
        num_classes = 8
        print('create resnet34 model')
        model = UNetResNet(encoder_depth=encoder_depth, num_classes=num_classes, pretrained=True)
    else:
        assert False
    model.eval()
    return model.to(device)

def adjust_learning_rate(optimizer, epoch, lr):
    """Sets the learning rate to the initial LR decayed by 10 every 30 epochs"""
    lr = lr * (0.1 ** (epoch // 30))
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

def find_latest_model_path(dir):
    model_paths = []
    epochs = []
    for path in Path(dir).glob('*.pt'):
        if 'epoch' not in path.stem:
            continue
        model_paths.append(path)
        parts = path.stem.split('_')
        epoch = int(parts[-1])
        epochs.append(epoch)

    if len(epochs) > 0:
        epochs = np.array(epochs)
        max_idx = np.argmax(epochs)
        return model_paths[max_idx]
    else:
        return None

def train(train_loader, model, criterion, optimizer, validation, args):

    latest_model_path = find_latest_model_path(args.model_dir)

    best_model_path = os.path.join(*[args.model_dir, 'model_best.pt'])

    if latest_model_path is not None:
        state = torch.load(latest_model_path)
        epoch = state['epoch']
        model.load_state_dict(state['model'])
        epoch = epoch

        #if latest model path does exist, best_model_path should exists as well
        assert Path(best_model_path).exists() == True, f'best model path {best_model_path} does not exist'
        #load the min loss so far
        best_state = torch.load(latest_model_path)
        min_val_los = best_state['valid_loss']

        print(f'Restored model at epoch {epoch}. Min validation loss so far is : {min_val_los}')
        epoch += 1
        print(f'Started training model from epoch {epoch}')
    else:
        print('Started training model from epoch 0')
        epoch = 0
        min_val_los = 9999

    valid_losses = []
    for epoch in range(epoch, args.n_epoch + 1):

        adjust_learning_rate(optimizer, epoch, args.lr)

        tq = tqdm.tqdm(total=(len(train_loader) * args.batch_size))
        tq.set_description(f'Epoch {epoch}')

        losses = AverageMeter()

        model.train()
        for i, (input, target) in enumerate(train_loader):
            #input_var  = Variable(input).cuda()
            input_var  = input.cuda()
            #target_var = Variable(target).type(torch.LongTensor).cuda()
            target_var = target.type(torch.LongTensor).cuda()
            #target_var = Variable(target).long().cuda()

            masks_pred = model(input_var)
            #print(masks_pred.shape, target_var.shape)
            
            #masks_probs_flat = masks_pred.view(-1)
            #true_masks_flat  = target_var.view(-1)
            #print(masks_probs_flat.shape, true_masks_flat.shape)

            #loss = criterion(masks_probs_flat, true_masks_flat)
            loss = criterion(masks_pred, target_var.squeeze())
            losses.update(loss)
            tq.set_postfix(loss='{:.5f}'.format(losses.avg))
            tq.update(args.batch_size)

            # compute gradient and do SGD step
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

        valid_metrics = validation(model, valid_loader, criterion)
        valid_loss = valid_metrics['valid_loss']
        valid_losses.append(valid_loss)
        print(f'\tvalid_loss = {valid_loss:.5f}')
        tq.close()

        #save the model of the current epoch
        epoch_model_path = os.path.join(*[args.model_dir, f'model_epoch_{epoch}.pt'])
        torch.save({
            'model': model.state_dict(),
            'epoch': epoch,
            'valid_loss': valid_loss,
            'train_loss': losses.avg
        }, epoch_model_path)

        if valid_loss < min_val_los:
            min_val_los = valid_loss

            torch.save({
                'model': model.state_dict(),
                'epoch': epoch,
                'valid_loss': valid_loss,
                'train_loss': losses.avg
            }, best_model_path)

def validate(model, val_loader, criterion):
    losses = AverageMeter()
    model.eval()
    with torch.no_grad():

        for i, (input, target) in enumerate(val_loader):
            input_var = Variable(input).cuda()
            target_var = Variable(target).long().cuda()
            

            output = model(input_var)
            loss = criterion(output, target_var.squeeze())

            losses.update(loss.item(), input_var.size(0))

    return {'valid_loss': losses.avg}

def save_check_point(state, is_best, file_name = 'checkpoint.pth.tar'):
    torch.save(state, file_name)
    if is_best:
        shutil.copy(file_name, 'model_best.pth.tar')

def calc_crack_pixel_weight(mask_dir):
    avg_w = 0.0
    n_files = 0
    for path in Path(mask_dir).glob('*.*'):
        n_files += 1
        m = ndimage.imread(path)
        ncrack = np.sum((m > 0)[:])
        w = float(ncrack)/(m.shape[0]*m.shape[1])
        avg_w = avg_w + (1-w)

    avg_w /= float(n_files)

    return avg_w / (1.0 - avg_w)

if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='PyTorch ImageNet Training')
    parser.add_argument('-n_epoch', default=10, type=int, metavar='N', help='number of total epochs to run')
    parser.add_argument('-lr', default=0.001, type=float, metavar='LR', help='initial learning rate')
    parser.add_argument('-momentum', default=0.9, type=float, metavar='M', help='momentum')
    parser.add_argument('-print_freq', default=20, type=int, metavar='N', help='print frequency (default: 10)')
    parser.add_argument('-weight_decay', default=1e-4, type=float, metavar='W', help='weight decay (default: 1e-4)')
    #parser.add_argument('-batch_size',  default=4, type=int,  help='weight decay (default: 1e-4)')
    parser.add_argument('-batch_size',  default=2, type=int,  help='weight decay (default: 1e-4)')
    #parser.add_argument('-num_workers', default=4, type=int, help='output dataset directory')
    parser.add_argument('-num_workers', default=2, type=int, help='output dataset directory')

    parser.add_argument('-data_dir',type=str, help='input dataset directory')
    parser.add_argument('-model_dir', type=str, help='output dataset directory')
    parser.add_argument('-model_type', type=str, required=False, default='resnet101', choices=['vgg16', 'resnet101', 'resnet34'])

    args = parser.parse_args()
    os.makedirs(args.model_dir, exist_ok=True)

    DIR_IMG  = os.path.join(args.data_dir, 'images')
    DIR_MASK = os.path.join(args.data_dir, 'masks')

    img_names  = [path.name for path in Path(DIR_IMG).glob('*.jpg')]
    mask_names = [path.name for path in Path(DIR_MASK).glob('*.png')]

    print(f'total images = {len(img_names)}')

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

    model = create_model(device, args.model_type)

    optimizer = torch.optim.SGD(model.parameters(), args.lr,
                                momentum=args.momentum,
                                weight_decay=args.weight_decay)

    #crack_weight = 0.4*calc_crack_pixel_weight(DIR_MASK)
    #print(f'positive weight: {crack_weight}')
    #criterion = nn.BCEWithLogitsLoss(pos_weight=torch.Tensor([crack_weight]).to('cuda'))
    #criterion = nn.BCEWithLogitsLoss().to('cuda')
    criterion = nn.CrossEntropyLoss().to('cuda')

    #channel_means = [0.485, 0.456, 0.406]
    #channel_stds  = [0.229, 0.224, 0.225]
    channel_means = [0.5]
    channel_stds  = [0.5]
    #train_tfms = transforms.Compose([transforms.ToTensor()])
    train_tfms = transforms.Compose([transforms.Resize((448,448)), transforms.ToTensor()])
                                     #,transforms.Normalize(channel_means, channel_stds)])

    #val_tfms = transforms.Compose([transforms.ToTensor()])
    val_tfms = transforms.Compose([transforms.Resize((448,448)), transforms.ToTensor()])
                                   #,transforms.Normalize(channel_means, channel_stds)])

    #mask_tfms = transforms.Compose([transforms.ToTensor()])
    mask_tfms = transforms.Compose([transforms.Resize((448,448)), transforms.ToTensor()])#,transforms.Normalize((0.5,), (0.5,))])
    '''
    mask_tfms = transforms.Compose([
    transforms.ToTensor(),
    transforms.Lambda(lambda x: x.repeat(3,1,1)),
    transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))])
    '''
    dataset = ImgDataSet(img_dir=DIR_IMG, img_fnames=img_names, img_transform=train_tfms, mask_dir=DIR_MASK, mask_fnames=mask_names, mask_transform=mask_tfms)
    train_size = int(0.85*len(dataset))
    valid_size = len(dataset) - train_size
    train_dataset, valid_dataset = random_split(dataset, [train_size, valid_size])

    train_loader = DataLoader(train_dataset, args.batch_size, shuffle=False, pin_memory=torch.cuda.is_available(), num_workers=args.num_workers)
    valid_loader = DataLoader(valid_dataset, args.batch_size, shuffle=False, pin_memory=torch.cuda.is_available(), num_workers=args.num_workers)

    #model.cuda()
    model.to(torch.device("cuda:0"))

    train(train_loader, model, criterion, optimizer, validate, args)


The following is the code in “data_loader.py”:

import os
import numpy as np
from torch.utils.data import DataLoader, Dataset
import random
from PIL import Image
import matplotlib.pyplot as plt

class ImgDataSet(Dataset):
    def __init__(self, img_dir, img_fnames, img_transform, mask_dir, mask_fnames, mask_transform):
        self.img_dir = img_dir
        self.img_fnames = img_fnames
        self.img_transform = img_transform

        self.mask_dir = mask_dir
        self.mask_fnames = mask_fnames
        self.mask_transform = mask_transform

        self.seed = np.random.randint(2147483647)

    def __getitem__(self, i):
        fname = self.img_fnames[i]
        fpath = os.path.join(self.img_dir, fname)
        img = Image.open(fpath)
        if self.img_transform is not None:
            random.seed(self.seed)
            img = self.img_transform(img)
            #print('image shape', img.shape)

        mname = self.mask_fnames[i]
        mpath = os.path.join(self.mask_dir, mname)
        mask = Image.open(mpath)
        #print('khanh1', np.min(test[:]), np.max(test[:]))
        if self.mask_transform is not None:
            print(np.array(mask).min(), np.array(mask).max())
            mask = self.mask_transform(mask)
            print(mask.min(), mask.max())
            #print('mask shape', mask.shape)
            #print('khanh2', np.min(test[:]), np.max(test[:]))

        return img, mask #torch.from_numpy(np.array(mask, dtype=np.int64))

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


class ImgDataSetJoint(Dataset):
    def __init__(self, img_dir, img_fnames, joint_transform, mask_dir, mask_fnames, img_transform = None, mask_transform = None):
        self.joint_transform = joint_transform

        self.img_dir = img_dir
        self.img_fnames = img_fnames
        self.img_transform = img_transform

        self.mask_dir = mask_dir
        self.mask_fnames = mask_fnames
        self.mask_transform = mask_transform

        self.seed = np.random.randint(2147483647)

    def __getitem__(self, i):
        fname = self.img_fnames[i]
        fpath = os.path.join(self.img_dir, fname)
        img = Image.open(fpath)

        mname = self.mask_fnames[i]
        mpath = os.path.join(self.mask_dir, mname)
        mask = Image.open(mpath)

        if self.joint_transform is not None:
            img, mask = self.joint_transform([img, mask])

        #debug
        # img = np.asarray(img)
        # mask = np.asarray(mask)
        # plt.subplot(121)
        # plt.imshow(img)
        # plt.subplot(122)
        # plt.imshow(img)
        # plt.imshow(mask, alpha=0.4)
        # plt.show()

        if self.img_transform is not None:
            img = self.img_transform(img)

        if self.mask_transform is not None:
            mask = self.mask_transform(mask)

        return img, mask #torch.from_numpy(np.array(mask, dtype=np.int64))

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

The above codes are modifications to the model at GitHub - khanhha/crack_segmentation: This repository contains code and dataset for the task crack segmentation using two architectures UNet_VGG16, UNet_Resnet and DenseNet-Tiramusu to train on that dataset. It was a model for crack segmentation (1 class).

So, when I print mask min and max values before and after the mask_transforms:

            print(np.array(mask).min(), np.array(mask).max())
            mask = self.mask_transform(mask)
            print(mask.min(), mask.max())

I get the following example values:

1 6
tensor(0.0039) tensor(0.0235)
0 7
tensor(0.) tensor(0.0275)

I don’t know why the mask values change into floats. The training is working, but the results after 11 epochs are not good at all. I don’t know what’s wrong with the codes that don’t make it work properly.

The transformation is most likely wrong as it will transform your class labels to float values in [0, 1]. The subsequent target = target.long() call will then create class indices in [0, 1] and will thus wipe all other classes.
No not use the standard data transformations on the segmentation maps, but transform it manually via torch.from_numpy() etc.
Also, in case you are resizing the target, make sure to use the “nearest neighbors” interpolation to avoid creating “new” labels in the target.

Thanks @ptrblck for your reply.
As I still don’t have experience with pytorch, could you pls suggest code revisions?