Fighting overfitting

Introduzione al Deep Learning con PyTorch

Jasmin Ludolf

Senior Data Science Content Developer, DataCamp

Reasons for overfitting

Overfitting: the model does not generalize to unseen data
- Model memorizes training data
- Performs well on training data but poorly on validation data

Possible causes:

Problem	Solutions
Dataset is not large enough	Get more data / use data augmentation
Model has too much capacity	Reduce model size / add dropout
Weights are too large	Weight decay

Fighting overfitting

Strategies:

Reducing model size or adding dropout layer
Using weight decay to force parameters to remain small
Obtaining new data or augmenting data

"Regularization" using a dropout layer

Randomly zeroes out elements of the input tensor during training

model = nn.Sequential(nn.Linear(8, 4),
                      nn.ReLU(),
                      nn.Dropout(p=0.5))
features = torch.randn((1, 8))
print(model(features))

tensor([[1.4655, 0.0000, 0.0000, 0.8456]], grad_fn=<MulBackward0>)

Dropout is added after the activation function
Behaves differently in training vs. evaluation - use model.train() for training and model.eval() to disable dropout during evaluation

Regularization with weight decay

optimizer = optim.SGD(model.parameters(), lr=0.001, weight_decay=0.0001)

Controlled by the weight_decay parameter in the optimizer, typically set to a small value (e.g., 0.0001)
Weight decay encourages smaller weights by adding a penalty during optimization
Helps reduce overfitting, keeping weights smaller and improving generalization

Data augmentation

examples of data augmentation

Let's practice!

Introduzione al Deep Learning con PyTorch