Inception v3 vs v4 on CIFAR-100

This repo benchmarks Inception-v4 against Inception-v3 on CIFAR-100, focusing on efficiency and accuracy under the same training setup.

Models

• Inception-v3: classic Inception family model with factorized convolutions and auxiliary classifier support.
• Inception-v4: newer, deeper Inception variant, typically providing stronger feature extraction and accuracy.

Params (approx.)

• Inception-v3: ~23.9M
• Inception-v4: ~42.7M

Aux output

• Inception-v4 does not include an auxiliary classifier head, so to keep the comparison fair, the aux head was disabled for Inception-v3 as well.

Data & Setup

Split

• Train — 70%
• Validation — 15%
• Test — 15%

Augmentations

• Resize to 299×299 (item-wise)
• Random horizontal flip (no vertical flips)
• Random rotation up to ±15°
• Random zoom up to 1.2×
• Random perspective/affine warp (max_warp=0.2, p_affine=0.9)
• Random lighting changes (brightness/contrast, max_lighting=0.3, p_lighting=0.9)
• Random erasing (cutout-style, p=0.2)
• Normalization to ImageNet stats (mean/std)

Batch size: 256

Training

Training process features

• LR finder — used to pick max learning rate
• One-Cycle policy — stable + efficient convergence
• Optimizer — Adam with decoupled weight decay (wd=1e-2)
• Criterion — CrossEntropyLoss (via CrossEntropyLossFlat)
• MixUp — improves robustness / reduces overfitting
• Callbacks — EarlyStopping (val loss), SaveModel (best weights), CSVLogger (metrics per epoch)

Visualization: LR range + learning curves
Epochs: 100

Results (Test Top-1)

• Inception-v3: 0.6981
• Inception-v4: 0.8079

Hardware

Training was done on CUDA cores using NVIDIA GeForce RTX 5090.

Note

Deeper insights, intermediate experiments, and full result logs are saved inside the notebooks.

Conclusion

During the analysis, both models demonstrated strong image recognition on CIFAR-100, but Inception-v4 was clearly superior, delivering about a +10% absolute Top-1 improvement over Inception-v3 (0.8079 vs 0.6981); the most likely reason is that the deeper block depth in v4 (higher representational capacity and richer hierarchical features) made a noticeable difference.

Author

Created by Denys Bondarchuk. Feel free to reach out or contribute to the project!