Facebook Quietly Releases a New Version of the PyTorch Library, Featuring Several …
Facebook quietly releases a new version of the PyTorch library, which includes several amazing features! AI developers, particularly image processors, are in for a treat. This library primarily competes for attention with TensorFlow. At the beginning of the year, the number of articles citing it increased by nearly 200%.
Facebook quietly releases a new version of the PyTorch library, which includes several amazing features! AI developers, particularly image processors, are in for a treat.
This library primarily competes for attention with TensorFlow. At the beginning of the year, the number of articles citing it increased by nearly 200%. Last week, a new amazing version, PyTorch 1.3, was quietly released. What’s new and interesting here?
- Secure Machine Learning Research with Crypten. This addresses the issue of data security and privacy using advanced cryptography. You can simply perform computations on encrypted data.
- Modular Object Detection with Detectron2. Detectron has been rewritten from Caffe2 directly into PyTorch. It additionally includes all the models that were available in the original Detectron, plus several new models including Cascade R-CNN, Panoptic FPN, and TensorMask. Furthermore, it enhances object detection capabilities and introduces new tasks such as semantic segmentation and panoptic segmentation.
- Improved Model Interpretability with Captum. Captum evaluates the contribution of each input feature to the model's output. It assesses the contribution of each neuron in a given layer to the model. It evaluates the contribution of each input element to the activation of a specific hidden neuron. Additionally, it includes a visualisation widget for interpretability, which facilitates understanding of the model. Captum's feature statistics work across images, text, and other features, helping users comprehend feature attribution.
Sources:
- https://pytorch.org/blog/pytorch-1-dot-3-adds-mobile-privacy-quantization-and-named-tensors/
- https://towardsdatascience.com/facebook-has-been-quietly-open-sourcing-some-amazing-deep-learning-capabilities-for-pytorch-a7ed5bc71f26
- https://www.oreilly.com/ideas/one-simple-graphic-researchers-love-pytorch-and-tensorflow?fbclid=IwAR3kYmlyD7zky37IYFu0cafQn7yemhl8P-7MNyB30z0q5RDzxcTOrP8kxDk
Originally published on Facebook — link to post
Původní zdroj: facebook