Image recognition app for weather phenomenona images. The model is trained on this Kaggle's dataset.

I tested different models and found that the best model for this application is the ResNet50. Model training and selection can be found in the JupyterNotebook.

Install the dependencies with the following command:

pip install -r requirements.txt

or build Docker image from Dockerfile with the following command:

docker build -t cnn_recognition_weather .

First, you need to start the API service.

Run the API service with the following command:

python api.py

and open the browser at http://localhost:5000/. You should see a message that confirms the service is running.

Run the recently built image with the following command:

docker run -p 5000:5000 --rm cnn_recognition_weather

and open the browser at http://localhost:5000/. Port 5000 is the port that the service is listening on, but you can change it to any port you want.

To see documentation for the API endpoints, open the browser at http://localhost:5000/docs

At this moment, there is only one endpoint at /predict. It takes an image url as a parameter and returns the weather phenomenon label and prediction score. For example,

localhost:5000/predict?image_url=https://media.istockphoto.com/photos/water-is-life-picture-id165981483?k=20&m=165981483&s=612x612&w=0&h=4IXRF_i9xnCwpVeqZuqYYANZ5TfO6alaVQvZ1hZFIMA=

Should return the following JSON:

{
  "prediction": "dew",
  "probability": 100
}

A bunch of convolutional neural networks pretrained on imagenet were trained on this dataset. The process is described in the JupyterNotebook. At the end, the best model was selected and saved as a Keras model. Only this model was uploaded to this repository and Heroku due to space limitations.

You can train and generate new models on your own using the JupyterNotebook.

The whole app is deployed on Heroku. You can test it on the following link: https://weather-recognition.herokuapp.com/docs