When generating images, pixel colors aren't the most useful way to describe an image! You need ways to capture the patterns of an image so they can be modified to generate new images. 🖼️
In deep learning, this is done with an Encoder: a neural network that takes an input image and outputs statistics about the higher-level features found. It's useful for a variety of downstream applications for high-quality image synthesis. 📊
This repository contains a collection of deep encoders implemented in PyTorch that computes features of the image at different scales. You can integrate these into your own Python applications! ⚙️
The pretrained models are automatically downloaded and available under the Creative Commons BY-NC-SA License. They have the following properties:
- Normalized activations with mean
1.0, based on the statistics of 50k samples from ImageNet's validation set. - Only convolution layers are included at 16-bit data format for more compact downloads than a full classifier.
These are the models currently available:
| Model | Version | Size |
|---|---|---|
| VGG11 | bbd30ac9 | 17Mb |
| VGG13 | c768596a | 18Mb |
| VGG16 | 397923af | 28Mb |
| VGG19 | dcbb9e9d | 38Mb |
| Thinet-Small | scratch-B | 15Mb |
| Thinet-Tiny | scratch-B | 2Mb |
The VGG family of networks are based on versions from the official torchvision library, and the Thinet family is based on the Rethinking Network Pruning repository.