This project offers funtionality for training a top model over a pre-loaded keras model, fine tuning some layers of a pre-loaded model and retraining a pre-loaded model. There is also an option to train a simple CNN from scratch. The keras models supported for fine-tuning/retraing are:

• VGG16
• VGG19
• ResNet-50
• Inception-v3

The TransferModel class in src/keras_transfer_learning.py wraps model functionality and src/train_helper.py offers a training suite.

It is recommended to keep data specific code in seperate scripts. See branch planet-amazon-kaggle-src (currently outdated, sorry) for an example of using the retraining functionality on data of the Planet: Understanding the Amazon from Space kaggle competition.

Make sure to use python 3, some code is not compatible with python 2

This project was inspired from:
https://github.com/EKami/planet-amazon-deforestation
I recommend it for a wonderful jupyter notebook breakdown of training a CNN on competition data.

matplotlib==2.0.2
Keras==2.0.6
numpy==1.13.1
tensorflow_gpu==1.2.1
pandas==0.20.1
h5py==2.7.0
kaggle_data==0.1
tqdm==4.14.0
Pillow==4.2.1
seaborn==0.8
scikit_learn==0.19b2