Skip to content

Latest commit

 

History

History
98 lines (62 loc) · 5.26 KB

README.md

File metadata and controls

98 lines (62 loc) · 5.26 KB

birds-classification

based on kaggle BIRDS 450 SPECIES- IMAGE CLASSIFICATION

GRANDALA alt text

ABBOTTS BABBLER alt text

docker-flask-postgres

This is a project about docker-flask-postgres-image classification. To run this on your computer you must first install docker.

If you are a windows user, download and install docker desktop, open it on windows, register and log in to docker. The following steps will work correctly.

Running

The project is about 110 MB, I upload the whole project on google drive, you can download it at this link [https://drive.google.com/drive/folders/1-oq7-q45IX-ThQvTvYjrUjd4levQ3r5b?usp=sharing]

After download it and unzip, and then run this command.

cd project-folder-path

docker-compose up --build -d   # Run the container.

The site will be available to you at localhost:5000 or your IP:5000.

As long as your computer builds this image and runs the container, You can browse the site by accessing the site from your mobile phone or other device(under the same wifi), enjoy it!

Here is a demo: alt text

After prediction, you can click the More Info button to search for more information about this bird. alt text

Cautions! This model only has good accuracy in predicting bird photos, since it has only 450 categories, if you upload out of categories photos, it will give funny results, it will find the most similar one out of 450 species of birds ! Although there are ways to solve this problem, the results are more interesting now, so I leave it. Here are some interesting results.

alt text alt text

How it works

I used the python:3.8 image for construction, and then installed the necessary libraries through Dockerfile, such as flask, torch, psycopg2 and other necessary libraries. The role of flask is to allow users to access index.html, and users can access index.html and Upload a picture, which must be a bird picture, and then the model predicts the name of the bird in the backend and displays the result on the page. The role of postgres here is to save the picture path and the prediction to the database.

Dataset

This application is to determine the category of bird image uploaded by users. Based on kaggle's BIRDS 450 SPECIES- IMAGE CLASSIFICATION Dataset [https://www.kaggle.com/datasets/gpiosenka/100-bird-species] ~ 2 GB, this dataset is not balanced, so f1_score is used as the metric.

Model Architecture

Target model is based on the ConvNext tiny model ~ 110 MB without pre-training, and knowledge distillation is used to improve the training effect.

Teacher model: Pre-trained ConvNext xlarge model ~ 1,300 MB, using nn.CrossEntropyLoss() to calculate loss, no Image augmentation due to long training time (>18h), 5 rounds of training to get f1_score = 0.938, according to the trend of the training data, the F1 score should be improved even higher.

Loss & Metrics

Classification Loss - nn.CrossEntropyLoss()
Classification metric - f1_score

Teacher model: Convnext-xlarge-224-22k alt text

Student model: Convnext-tiny-224 alt text

Result of training Teacher model

  • The model has no obvious overfitting, but it took long time(18 hours) to train for 5 rounds, so the 5th round(epoch=5) is used as the best model
  • The f1 score of validating set is about 0.933
  • It seems that the xlarge model could get higher f1 score if it trains more epochs.

Result of training Student model

  • Here we can find that the student model is still making progress. Because of the time, there are only 5 rounds of training here, and the results of the 5th round are the best. the val_f1_score = 0.925

Summary

  • teacher model: val_f1_score = 0.933, test_f1_score = 0.942
  • student model: val_f1_score = 0.925, test_f1_score = 0.931
  • the size of xlarge model is 1,363MB, and the size of tiny model is 110MB, Effectively reduces the size by 92%
  • compare with teacher model, student model is almost the same performance!

Dockerfile

  • a. docker-compose for full architecture: synchronous projects
  • b. client: HTML Frontend
  • c. model: teacher model: transfer learning by ConvNext-xlarge model, student model: trained from scratched ConvNext-tiny model.

Summary

  • I use the xlarge model to teach the tiny model, the f1_score on testing set is about 0.928, the performance dropped by 1%

  • the size of xlarge model is 1,300 MB, and the size of tiny model is 110 MB, Effectively reduces the size by 91.5%