Transfer Learning using Keras and EfficientNet

Hi! This is a demo of image classification using the transfer learning technique for our Computer Vision vs. Image recognition blogpost. In a Jupyter Notebook, we'll go step-by-step over how to load and analyse data, add simple image augmentation, prepare the base model with pre-trained weights, and train an image classifier using Keras.

Prerequisites: You should be familiar with basic concepts of Machine Learning and the Python programming language. Some familiarity with Jupyter Notebooks and a general understanding of Convolutional Neural Networks is a plus, but not required.

Try the notebook in Colab

You don't even need to clone the repository, just click on the icon below and run the notebook in Google Colab!

Running locally

Requirements: you must have Docker and Docker Compose installed.

To run the notebook:

1. clone the repo
2. run docker-compose up -d in the terminal from the repo's root directory
3. open the logs docker-compose logs and open the last link (starts with http://127.0.0.1:8888/)
4. the JupyterLab IDE should appear in your browser
5. open work/TransferLearning.ipynb in JupyterLab

Common issues

Invalid credentials. If you cannot log in to JupyterLab despite using the right token (double-check in the docker logs, a new token gets generated each time the server restarts), there is a chance that you already have an instance of Jupyter running locally, to which the requests get sent. To check it, run sudo lsof -PiTCP:8888 -sTCP:LISTEN to check which services use port 8888. If you have any non-docker services listed, it might be the problem. Check them using ps -p <PID> and kill <PID> if needed. In my case, it was a VS Code Python extension running another Jupyter server that caused the issue and I had to close VS Code for the docker service to work properly.