Inception v3 vs v4

This repository compares Inception v3 and Inception v4 models for image classification. The models are tested on the CIFAR-10, CIFAR-100, and Tiny ImageNet datasets. The goal is to understand how differences in their architectures influence training, accuracy, and stability on datasets of different sizes and difficulty.

Cloning the Repository

This repository uses Git LFS to store pretrained .pth model files. To clone the repository correctly, follow these steps:

1. If Git LFS is not installed, install it depending on your operating system:

   • macOS (Homebrew):

     brew install git-lfs

   • Linux (Debian / Ubuntu):

     sudo apt install git-lfs

   • Windows: download and install from https://git-lfs.com/

2. After installation, initialize Git LFS:

   git lfs install

3. Clone the repository:

   git clone https://github.com/thejvdev/dl-inception-models.git
   cd dl-inception-models

4. Download the actual pretrained .pth models from the Git LFS remote storage:

   git lfs pull

Models

The Inception architecture uses several convolution paths in parallel with different filter sizes. This helps the network learn visual features at multiple scales in the same layer, while staying efficient. By combining these parallel paths and keeping the model size under control, Inception models balance accuracy and speed well. In this project, Inception v3 and Inception v4 are studied. They have about 23 million and 43 million parameters, respectively.

Note

For a fair evaluation of both architectures, the auxiliary (AUX) output was disabled in Inception v3, since Inception v4 does not include an auxiliary classifier at all.

Training Methodology

The training methodology follows a standardized image classification pipeline based on Inception models. The dataset is partitioned into training, validation, and test subsets, with 70% of the data used for training and the remaining 30% evenly split between validation and testing. Data augmentation and regularization techniques are employed to improve generalization, while optimization is carried out using the Adam optimizer with weight decay and a One-Cycle learning rate schedule to ensure stable and efficient convergence.

Results

Model	CIFAR-10	CIFAR-100	Tiny ImageNet
Inception v3	95.41%	69.81%	44.65%
Inception v4	95.03%	80.79%	54.35%

Authors

Created by Denys Bondarchuk, Maksym Kuznietsov, Ruslan Pylypiv and Volodymyr Kolesnikov.