ML → Deep Learning via PyTorch — the Garment Classifier
30 min
Day 39: Training the Garment Classifier: overfitting countermeasures
Training for real, with the overfitting toolkit
Now you train the actual Garment Classifier: pretrained ResNet backbone, your 8+ class dataset, the training loop from Day 29, validation from Day 30. The enemy is overfitting — with a few thousand images and millions of parameters, the model can memorize. You have a full toolkit of countermeasures, and stating them is a Stage 1 exit criterion.
- Data augmentation (Day 38) — the strongest single defense; each epoch sees slightly different images.
- Dropout — randomly zeroing activations during training so the model can't rely on any single neuron.
- Weight decay (Day 32) — penalizing large weights toward simpler solutions.
- Early stopping (Day 30) — halt at the best validation epoch.
- Transfer learning + freezing (Day 37) — fewer trainable parameters means less to overfit.
Adding dropout to the classifier head, with weight decay in the optimizer
import torch.nn as nn
model.fc = nn.Sequential(
nn.Dropout(0.5), # regularize the head
nn.Linear(model.fc.in_features, 8),
)
optimizer = torch.optim.Adam(
model.parameters(), lr=1e-4, weight_decay=1e-4 # L2 regularization
)Add countermeasures one at a time
Don't throw every regularizer in at once — you won't know what helped. Establish a baseline, then add augmentation, measure; add dropout, measure. This is the Continuous-tracks discipline the roadmap enforces from Stage 3: every change re-runs the evaluation, and improvement is a measured number, not a hunch.
Key terms
- Dropout
- Randomly zeroing a fraction of activations during training so the network cannot depend on any single neuron, reducing overfitting.
- Regularization
- Any technique that constrains a model to prefer simpler solutions that generalize better.
- Overfitting countermeasures
- The toolkit — augmentation, dropout, weight decay, early stopping, transfer learning — for closing the train/val gap.
What does a dropout layer do during training?
