A model of image classification based on Yolov8 architecture using pytorch. Here, i use a custom dataset ** of 500 bird species containing about ** 80,000 images for training, validation and testing.

• Python3
• Pytorch

pip instal pytorch       # pytorch library
pip install torchsummary # summary
pip install torchvision  # pytorch for vision

NB: Update the libraries to their latest versions before training.

⬇️⬇️Download and extract training dataset on Kaggle: 500 bird species dataset

⬇️⬇️Download pretrained model: Model

Run the following scripts for training and/or testing

python train.py # For training the model 

Run the following scripts for visual result of model:

1. Download Docker

Open CMD

2. Download my image

docker pull vvduc1803/500bird_cls:latest                                  # Pull image

3. Copy and paste

docker run -it -d --name 500_bird_cls -p 1234:1234 vvduc1803/500bird_cls  # Run container

4. Copy and paste

docker logs -f 500_bird_cls                                               # Run visual result

Batch size: 64, GPU: RTX 3050 4G

Model:

Finetuning the model.

Sample classification results

Accuracy of the network on the 2500 test images: 76.38%

1. The MODEL with 38M params has a very large size i.e 0.4 GB, compared to other models like Resnet18(40 MB)
2. Adjusting parameters like **batch size, number of workers, pin_memory, ** etc. may help you reduce training time, especially if you have a big dataset and a high-end machine(hardware).
3. Adjusting parameters like learning rate, weight decay etc maybe can help you improve model.

1. Experiments with different learning-rate and optimizers.
2. Converting and optimizing pytorch models for mobile deployment.

Van Duc

• "https://blog.roboflow.com/whats-new-in-yolov8/"