How to create a custom multilabel dataset class?


(Perry Son) #1

Hello everyone,

I am interested in creating a custom multilabel dataset class. I am reading the data from a csv file.

class MyDataset(Dataset):

    def __init__(self, df_data, data_dir = './', img_ext='.png', transform=None):
        super().__init__()
        self.df = df_data.values
        self.data_dir = data_dir
        self.img_ext = img_ext
        self.transform = transform

    def __len__(self):
        return len(self.df)
    
    def __getitem__(self, index):
        img_name,label = self.df[index]
        img_path = os.path.join(self.data_dir, img_name+self.img_ext)
        image = cv2.imread(img_path)
        if self.transform is not None:
            image = self.transform(image)
        return image, label

How can this be implemented?


#2

If you would like to use nn.BCEWithLogitsLoss (or nn.BCELoss) for the multi-label classification task, you would have to create a multi-hot encoded target.
Assuming label contains the class indices, you could use the following code:

target = torch.zeros(nb_classes)
target[label] = 1.

(Perry Son) #3

The code threw an error in the training function.

TypeError: new(): invalid data type 'str’

nb_classes = (list(range(1103)))
class MyDataset(Dataset):

    def __init__(self, df_data, data_dir = './', img_ext='.png', transform=None):
        super().__init__()
        self.df = df_data.values
        self.data_dir = data_dir
        self.img_ext = img_ext
        self.transform = transform

    def __len__(self):
        return len(self.df)
    
    def __getitem__(self, index):
        img_name,label = self.df[index]
        img_path = os.path.join(self.data_dir, img_name+self.img_ext)
        image = cv2.imread(img_path)
        target = torch.zeros(nb_classes)
        target[label] = 1.
        if self.transform is not None:
            image = self.transform(image)
        return image, target

Training function

total_step = len(loader_train)
for epoch in range(num_epochs):
    for i, (images, labels) in enumerate(loader_train):
        images, labels = Variable(images), Variable(labels)

        images = images.to(device)
        labels = labels.to(device)
        
        # Forward pass
        outputs = resNet(images)
        loss = criterion(outputs, labels)
        
        # Backward and optimize
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        if (i+1) % 100 == 0:
            print ('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}' 
                   .format(epoch+1, num_epochs, i+1, total_step, loss.item()))

#4

nb_classes should be the number of classes given as an int.
Could you change that and try to run your code again?


(Perry Son) #5

My mistake there. I thought nb_classes was a list of the classes indices.

nb_classes = 1103

However, the training function threw another error.
IndexError: too many indices for tensor of dimension 1

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-77-d05ecce66b45> in <module>()
      1 total_step = len(loader_train)
      2 for epoch in range(num_epochs):
----> 3     for i, (images, labels) in enumerate(loader_train):
      4         images, labels = Variable(data), Variable(target)
      5 

/opt/conda/lib/python3.6/site-packages/torch/utils/data/dataloader.py in __next__(self)
    613         if self.num_workers == 0:  # same-process loading
    614             indices = next(self.sample_iter)  # may raise StopIteration
--> 615             batch = self.collate_fn([self.dataset[i] for i in indices])
    616             if self.pin_memory:
    617                 batch = pin_memory_batch(batch)

/opt/conda/lib/python3.6/site-packages/torch/utils/data/dataloader.py in <listcomp>(.0)
    613         if self.num_workers == 0:  # same-process loading
    614             indices = next(self.sample_iter)  # may raise StopIteration
--> 615             batch = self.collate_fn([self.dataset[i] for i in indices])
    616             if self.pin_memory:
    617                 batch = pin_memory_batch(batch)

<ipython-input-69-ac5ff07bb12a> in __getitem__(self, index)
     17         #label = list(map(int, label.split(' ')))
     18         target = torch.zeros(nb_classes)
---> 19         target[label] = 1.
     20         if self.transform is not None:
     21             image = self.transform(image)

IndexError: too many indices for tensor of dimension 1


#6

How did you create label?
Here is a small dummy example how I suggested it should work:

nb_classes = 5
label = [0, 3]
target = torch.zeros(nb_classes)
target[label] = 1

(Perry Son) #7

Oh I know how it works!

I convert each label to an int array of class indices.

def __getitem__(self, index):
       img_name,label = self.df[index]
       img_path = os.path.join(self.data_dir, img_name+self.img_ext)
       image = cv2.imread(img_path)
       label = list(map(int, label.split(' '))) #'1 40 56 813' --> [1, 40, 56, 813]
       target = torch.zeros(nb_classes)
       target[label] = 1.
       if self.transform is not None:
           image = self.transform(image)
       return image, target

Thanks!