One hot encode label for multi-label classification

vision
1

Dear all,
im try to prepare dataset for multi-label classification with pytorch, there is an example with pytorch (dataloader) for multi-label classification?
thanks :slight_smile:

2

It is for the classification of text, images …?

3

video classification

4

In this case you can proceed as follows (I am just making an illustration)

import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from sklearn.preprocessing import MultiLabelBinarizer

torch.manual_seed(0) 

class VideoDataset(Dataset):
    """ Dataset class for Video Classification"""
    def __init__(self, data_folder, in_memory : bool):
        super().__init__()

        # I guess here you go so file to extract the videos and labels .     
        # l = label, ptf = path_to_file
        files =  ["ptf1.mp4",    "ptf2.mp4", "ptf3.mp4",         "ptf4.mp4",   "ptf5.mp4"]
        labels = [ ["l1", "l2"],  ["l2"],    ["l1", "l2", "l3"], ["l2", "l3"], ["l2", "l3"]]

        mlb = MultiLabelBinarizer()
        labels = mlb.fit_transform(labels) #array([[1, 1, 0], [0, 1, 0], [1, 1, 1], [0, 1, 1], [0, 1, 1]])
        
        
        self.in_memory = in_memory
        if self.in_memory :
            # preprocess the data beforehand : costly in memory, especially if the data are large
            self.data = [(self.video_to_tensor(f), torch.tensor(l, dtype=torch.long)) for f, l in zip(files, labels)]
        else :
            # pre-processing of data during loading: time consuming, because at each time the same processing will be done on the same data
            self.data = [(f, torch.tensor(l, dtype=torch.long)) for f, l in zip(files, labels)]

    def __getitem__(self, index):
        if self.in_memory :
            return self.data[index]
        else :
            inst = self.data[index]
            return self.video_to_tensor(inst[0]), inst[1]

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

    def video_to_tensor(self, file) :
        """convert a video to a tensor : can be a feature extraction .... depend on what you classify"""
        return torch.empty(5).uniform_(-10, 10)

dataset = VideoDataset(data_folder="my date folder path", in_memory=True)
len(dataset) # 5

batch_size = 2
shuffle = True
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=shuffle, collate_fn = None)

Here is how the classification can be done

class Model(nn.Module) :
    def __init__(self, input_dim=5, output_dim=3) :
        super().__init__()
        self.linear_layer = nn.Linear(input_dim, output_dim) # suppose 3 labels
        
    def forward(self, x) :
        return self.linear_layer(x)  

model = Model()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-2)
criterion = nn.BCEWithLogitsLoss() # because of multi-label

for x, y in dataloader :
    optimizer.zero_grad()
    y_pred = model(x)
    loss = criterion(y, y_pred)
    loss.backward() 
    optimizer.step()
2 Likes