imbalanced data

[kike-0304]

Hi,

I am using DeepTab for a binary classification task with highly imbalanced data.

For example:

• Class 0: 90%
• Class 1: 10%

How can I apply class weighting during training?

In LightGBM, I would typically use scale_pos_weight, and in PyTorch I can use BCEWithLogitsLoss(pos_weight=...).

Does DeepTab provide a built-in way to handle class imbalance, such as:

• class weights
• sample weights
• weighted loss functions
• weighted sampling

If so, could you provide an example?

Thank you.

[mkumar73]

Hello @kike-0304,

Thank you for the question.

To answer directly: class weights, sample weights, weighted loss functions, and weighted sampling are not currently exposed through the public estimator API in the current v1.x release.

Current Workaround

For now, the recommended workaround is to handle imbalance before fitting the model. For numerical features, you can use resampling, for example:

from imblearn.over_sampling import SMOTE
from deeptab.models import MambularClassifier

X_res, y_res = SMOTE(random_state=42).fit_resample(X_train, y_train)

model = MambularClassifier()
model.fit(X_res, y_res, max_epochs=100)

What's Coming in v2.0.0

v2.0.0 will add first-class support for several of these. Based on the direction of the project:

Weighted loss (pos_weight, equivalent to BCEWithLogitsLoss(pos_weight=...)):

import torch
import torch.nn as nn
from deeptab.models import MambularClassifier

pos_weight = torch.tensor([(y_train == 0).sum() / (y_train == 1).sum()])

# Expected v2.0.0 API (not yet available)
model = MambularClassifier(
    loss_fct=nn.BCEWithLogitsLoss(pos_weight=pos_weight)
)
model.fit(X_train, y_train, max_epochs=100)

Configurable stratification via SplitConfig:

# Expected v2.0.0 API (not yet available)
from deeptab.configs import SplitConfig

model = MambularClassifier(
    split_config=SplitConfig(stratify=True, val_size=0.15, random_state=42)
)

Class weights, sample weights, and weighted sampling aren't in v2.0.0 yet. We will look into whether they can be included before the release, which is planned for later this month. Thank you!