This past month, the source code of CRAL (CNN Research Abstraction Library) was made publically available in the spirit of the open-source movement and to promote open and collaborative method of development. Use cases for CRAL include medicine, remote sensing, manufacturing, Oil & Gas and other such critical sectors.
We have used CRAL to work on a wide array of deep learning based computer vision projects. Engineers and students from prominent universities, medical AI and drone enterprises have discovered CRAL to be the go-to solution for the development of better, accurate end models.
CRAL comes with state-of-the-art algorithms for image classification, object detection, semantic segmentation, instance segmentation and keypoint detection (currently under development). Built-in data loaders let you import data in multiple formats - ranging from standard COCO JSON format to XML based Pascal-VOC format and other such formats. With CRAL, you can now fast-track working on hyperparameter optimization for a range of SOTA algorithms. Use the collective results to perform statistical analysis for comparing models, accuracy, inference time and model size. With DLCV research being the vast domain that it is, deploying the CRAL library lets you follow a systematic approach to the reliable model you want to achieve, instead of the usual cumbersome methods required to get there.
We built CRAL on TensorFlow to leverage the existing large community that can help in updating and maintaining the library, which in-turn makes the usage of mature tool ecosystems such as TFX and Tensorboard easier for production-ready models. By letting you combine different modules seamlessly, CRAL is flexible and aids in easy custom pipeline construction. For example, you can now use multiple backbones for object detection or figure which algorithm works best for your dataset and application. Additionally, the multi-GPU optimized algorithms in this library enable you to build & train your algorithms faster in a cost-effective manner.
We have integrated Segmind Track with CRAL to help your automatically track your work. Segmind Track is an open-source experiment tracking platform built on top of MLFlow. You can track each run's metrics, parameters and artefacts including checkpoint files and Tensorboard files on a cloud dashboard.
Learn more about CRAL by reading the documentation, or check out the source code on Github. Sophisticated state of the art algorithms demand a methodical approach, and CRAL enables you to realize reliable & accurate end models faster. We will continue publishing our latest research/results as we develop and enhance CRAL further.