author	sudo-0x2a
completion date	04/28/2025

Fine-tuning Inception-v4 with CIFAR-10 Dataset

A fine-tuning practice on a classical CNN model. The original Inception-v4 was trained on ImageNet, which outputs 1000 features on the final layer. By modifying the final layer and adjusting all weights to make the model suitable for the CIFAR-10 classification task.
CIFAR-10
Inception-v4

Results:

Validate both models on the same CIFAR-10 validation set.

	Before	After
Accuracy	0.02%	95.77%
Loss	9.47	0.1227

*Before the fine-tuning, the original pre-trained model was basically unusable.

The Project Tested With the Following Setup:

• Hardware: Macbook Pro 14 M4 pro - 12C16G with 24GB RAM
• Python Version: 3.11.12