Autoencoder Graph Classfication

Introduction

This is a two step image classification solution.
First I use unsuperviser method trained a AutoEncoder model,
in second step I training a Classifier to complete this task.

My environment

 Nvidia RTX 4060
 cuda version 12.2
 PyTorch version 1.81+cu111

Although I using Python3.9, but i guess if you can successful install PyTorch 1.81+cu111, this repo should be work for you.
Because I didn't use any unique function of Python3.9 or Syntactic sugar ：）

Dataset

Here I take the Dogs vs. Cats dataset from Kaggle as an example demonstration.

I also split dataset by ratio 8:2 (train : test).

Init your environment

Recommond use conda environment

conda create --name AGC python=3.9
conda activate AGC

Install requirement.txt

git clone https://github.com/wuyiulin/AGC.git
cd AGC
pip install -r requirements.txt

Prepare your dataset and follow this architecture:

├── Train dataset 
│   └── fold(include all class training image, w/o label.)
└── Test  dataset
    ├── Class 1 image fold(only include class1 images.)
    :
    :
    :
    └── Class N image fold(only include classN images.)

Replace your dataset path in AutoEncoder.py and Classifier.py

Start

Step1-1: Training AutoEncoder

python AutoEncoder.py train

Step1-2: Check AutoEncoder training loss

python VisionLoss.py

You can check training loss here, if loss curve too ridiculous to use?
Kick this model weight to the Death Star, it deserved.

Step1-2: Check AutoEncoder training loss

python AutoEncoder.py vis

Sometimes we can't only depend loss curve to judge a model,
because all that glitters is not gold.

So you must check out those lovely kitties<3
If, unfortunately, they generated something ugly, you know what to do ^Q^

Step2-1: Training Classifier

python Classifier.py train

Step2-2: Test Classifier

python Classifier.py test

Test Loss: 0.7738, Accuracy: 84.58%

You can check final accuracy of this task,
it seems not a perfect score in Dogs vs. Cats dataset.

But in another dataset with natural lighting interference and multiple noises, this model achieved very high scores.

Test Loss: 0.0009, Accuracy: 99.94%

So I have full confidence that this model has certain usability in optical inspection in industrial or simpler environmental settings.😤

For my convenience, the program will automatically clear the /checkpoints and /log directories where model weights and loss records are stored in step 1-1.
Please refrain from storing any treasure there
(e.g. exceptionally well-trained model weights). <3

If you wish to modify this behavior, please go to edit the train function in AutoEncoder.py and Classifier.py.

Contact

Further information please contact me.

[email protected]