I’m currently working on a multi-label image classification problem where my dataset is significantly imbalanced. Each image in my dataset can have multiple labels. The labels are provided in a one-hot encoded format.
For example: [1,0,0,0,0,1,0] etc.

My train df:

Image Index Finding Labels
0   00005504_002.png    Pleural_Thickening
1   00003527_002.png    Atelectasis|Pneumonia
2   00018285_000.png    Effusion|Mass
3   00016971_007.png    Emphysema|Mass
4   00014022_071.png    Atelectasis|Consolidation|Pleural_Thickening

To balance the dataset, I considered undersampling the overrepresented classes. However, I encountered a challenges that it might lose effectiveness of model.

I am looking to implement class weights/Sample weights in PyTorch. How can I effectively implement in PyTorch for this multi-label classification problem? I have read online that class weights might not work well with one-hot encoded labels and that using sample weights or a custom loss function might be necessary. How can i implement custom loss with weighted sampling ?. Any particular advice would be appreciated.

Below is my code:

`resnet50 = ResNet101(input_shape=(256, 256, 3), weights='imagenet', include_top=False)

for layer in resnet50.layers[:-3]:
    layer.trainable = False

x = Flatten()(resnet50.output)
x = Dense(512, activation='relu')(x)
prediction = Dense(13, activation='sigmoid')(x)  

model = Model(inputs=resnet50.input, outputs=prediction)
learning_rate = 0.001
adam_optimizer = Adam(learning_rate=learning_rate)

model.compile(optimizer=adam_optimizer, loss='binary_crossentropy', metrics=['accuracy', AUC(multi_label=True)])
early_stopping = EarlyStopping(monitor='val_auc', patience=5, restore_best_weights=True)
history = model.fit(train_dataset, epochs=100, validation_data=val_dataset, callbacks=[early_stopping])`