Machine learning

Dog Breed Classification: CNNs vs. Vision Transformers

End-to-end PyTorch pipelines comparing custom CNNs, transfer learning, and ViTs.

Context: Machine Learning · University of Michigan
Role: Individual
When: 2026

Python
PyTorch
CNNs
Vision Transformers
Transfer learning

Constraint

Classify dog breeds (binary and multiclass) from a limited dataset without overfitting.

Decision

Compared custom CNNs, transfer learning with different layer-freezing setups, and ViT inference under one pipeline with early stopping and checkpointing.

Evidence

Test AUROC of 0.832 with frozen pretrained convolutional layers; best custom CNN reached 0.7652.

01 // Pipeline

▪Image standardization, CNN training, and ViT inference in PyTorch.
▪Early stopping, checkpointing, and AUROC-based evaluation for binary and multiclass tasks.

02 // Experiments

▪Transfer learning across multiple layer-freezing configurations.
▪CNNs compared against Vision Transformers.
▪Custom CNNs tuned with global average pooling, dropout, and weight decay.

03 // Results

▪Frozen pretrained conv layers: test AUROC 0.832.
▪Final custom CNN: test AUROC 0.7652.