TensorImage

TensorImage is an open source library designed to make training and deploying image classification models easy, with many built-in and useful features:

• Cluster training: automatically compare the performance of multiple trainers, speeding up the process of hyperparameter tuning and feature engineering, as there is no need to do it manually

• Multithreaded training: by default, all training operations are run in 10 threads to make training models faster

• Built-in image data augmentation operations for feature engineering:

  • Image flipping
  • Salt-pepper noise
  • Brightness adjusting
  • Random brightness
  • Contrast adjusting
  • Random contrast
  • Hue adjusting
  • Random hue
  • Saturation adjusting
  • Random saturation
  • Gamma adjusting
  • Gaussian blur
  • Color filtering
  • Grayscale filter
  • Median blur
  • Motion blur
  • Bilateral blur
  • Translation
  • Affine transformations
  • Perspective transformations
  • Piecewise affine transformations

• Workspace organization: all datasets, trained models, and internal metadata files are stored automatically inside a workspace directory, where you can quickly find any files you need

• Large-scale image classification: deploy trained models on thousands of images, with predictions for all images being stored in your workspace directory

• Model inference for individual/batches of images for real-time prediction without writing on disk

• Real-time training from individual/batches of images without reading from disk

Upcoming features

• More data augmentation operations to apply on images:
  • Image scaling
  • Image cropping
  • Image generation with GANs (Generative Adversarial Networks)
• Option to apply different data augmentation operations at once, e.g: instead of only applying gaussian blur, to be able to apply gaussian blur, pepper salt noise and random contrast at once, not uniquely separately
• Deployment of real-time image classification trainers and classifiers in a localhost server
• Automatic hyperparameter tuning using genetic algorithms
• Real-time object detection model training
• Real-time object detection model inference
• Deployment of real-time object detection trainers and classifiers in a localhost server

Installation

From the terminal:

$ pip3 install tensorimage

Documentation

You can view TensorImage's documentation here.

Support

If you are experiencing any errors or bugs, please report them in the issues section or contact us at tensor.image2@gmail.com

Contributing

If you have any ideas for features that should be added to TensorImage, please feel free to fork TensorImage and open a pull request.

License

TensorImage is licensed under the MIT license.