https://github.com/g-eoj/cv-tl-keras

Use the cross validation functions from scikit-learn to evaluate image classification transfer learning with Keras models.
https://github.com/g-eoj/cv-tl-keras

cross-validation keras numpy scikit-learn transfer-learning

Last synced: about 2 months ago
JSON representation

Use the cross validation functions from scikit-learn to evaluate image classification transfer learning with Keras models.

• Host: GitHub
• URL: https://github.com/g-eoj/cv-tl-keras
• Owner: g-eoj
• Created: 2017-07-04T20:59:07.000Z (almost 9 years ago)
• Default Branch: master
• Last Pushed: 2018-11-19T06:46:03.000Z (over 7 years ago)
• Last Synced: 2025-03-15T04:09:23.421Z (about 1 year ago)
• Topics: cross-validation, keras, numpy, scikit-learn, transfer-learning
• Language: Python
• Homepage:
• Size: 107 KB
• Stars: 1
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Cross Validate Transfer Learning

Use the cross validation functions from scikit-learn to evaluate image classification transfer learning with Keras models.

The system has the following features:

- Transfer learning bottlenecks stored in the HDF5 format for increased speed using h5py.

- Stratified k-fold, group k-fold, and leave one group out cross validation.

- Groups can be defined in csvs so that cross validation splits are done on group membership.

- Combine or exclude classes (so you don't have to copy or delete images on disk).

- Ways to deal with unbalanced classes:

    - Oversampling to balance classes during training.

    - Use class balance to scale the loss function during training.

- Logger to save all terminal output to a log file.

Results (particularly individual class scores for each fold) are saved as csvs for further analysis in your platform of choice.

Training status, accuracy, F-scores, and confusion matrices are output to terminal while the tests are running.

Additional info, such as model or data summaries, can also be output to terminal.

## Quick Intro

[CIFAR-10 Example Notebook](./examples/cifar10-example.ipynb)

## Setup

Install [miniconda](http://conda.pydata.org/miniconda.html).

Create a conda environment:

    conda create -n Keras python=3 numpy scipy yaml h5py scikit-learn pillow matplotlib tensorflow keras

    source activate Keras 

## Usage

See the example notebook and script. Most functions have detailed docstrings.

### Preparing Images

Put the images in folders named with the image class label. Then put these folders in a parent directory so the directory structure looks something like:

    images

    ├── airplane

    ├── automobile

    ├── bird

    ├── cat

    ├── deer

    ├── dog

    ├── frog

    ├── horse

    ├── ship

    └── truck