QuickDraw

Introduction :

Hello there, I am Thanh Minh Truong, this is one of my beginnings in the process of learning AI

Here is my python source code about Deep Learning - Image Classification with Quick Draw dataset - an online game developed by Google.

I used Pytorch framework for my project, this includes:

1. Preprocessing and validating input images and labels
2. Build CNN architecture
3. Model training
4. Model evaluation
5. Testing

In the future, i will use pen (or any object) with color red. When we are drawing some object (like star, moon,...), model will give a prediction of the object we draw

Dataset:

The dataset used for training my model could be found at [Quick Draw dataset] https://console.cloud.google.com/storage/browser/quickdraw_dataset/full/numpy_bitmap. Here I only picked up 20 files for 20 categories

Categories:

apple	banana	basketball	book	clock
cloud	eye	flip flops	flower	hat
key	moon	pillow	pizza	star
sun	t-shirt	table	underwear	zigzag

Model:

This is my model parameters ( use summary from torchsummary):

Training :

I downloaded .npy file corresponding to 20 classes and store them in folder Dataset

After that, I will run this scripts in the terminal:

python train.py -r 0.8 -s 1000 -b 8 -e 10 -o Adam -l 0.001

Checkpoint:

I save the checkpoint to the checkpoint/best.pth folder after the training process

Experiments :

For each class, I take the first 1000 images, and then split them to training and test sets with ratio 8:2.

This is result:

Confunsion matrix:

Test

Run test.py with images for testing in folder test_imaged

Requirements :

• python 3.10
• torch 2.1.1
• opencv-python 4.8.1
• matplotlib 3.8.2
• tqdm 4.66.1
• numpy 1.26.2
• scikit-learn 1.3.2
• tensorboard 2.15.1