Release 1.1.2

The release includes the following fixes:

• Excluded extra visualization in some methods;
• Added check for invalid face crop: during the inference on the stream from the web camera rarely some of the face box coordinates turned out to be negative. This usually happened when part of the face was out of the frame. This led to invalid crop of the face and error during the inference. Therefore, additional check was added: when such situation happens, the prediction method returns empty list and the warning is shown.

Release 1.1.1

Added the following fixes:

• Fixed the visualization for the case when there is more than one person on the image;
• Fixed the output video saving option in analyze_video method in the FER class;
• Refactored the tests;
• Updated the README file.

Release 1.1.0

Key updates

• New pretrained weights that correspond to MobileNetv2 model has been added;
• Some minor fixes made to FER inference class for convenience purposes;
• Updated README file: the information about the web application has been added.

MobileNetv2 weights

Small fixes to FER class (#20)

* Add no_grad to inference

* Make preprocess_output_list method available from outside

* Implement visualization method in the fer class for convenience

Release 1.0.0

Key updates

• Config files are set using Hydra library.
• Updated README file.

Release 0.5.0

Key updates

• Moved training to Pytorch Lightning package.
• Added possibility to choose model architecture for inference.

Release 0.1.1

First release of fer-pytorch package for facial expression recognition implemented in Pytorch.

Key features:

• Possibility to train the wide variety of classification neural net architectures on FER+ dataset using the famous timm library;
• Inference class for applying the fer model on different types of data (images, list of images, video files and e.t.c.);
• Output images and video files can be saved;
• Convenient interface for downloading the ready-to-use pretrained on FER+ dataset weights from the internet, as well as loading custom ones stored locally;
• Convenient tool for downloading the FER+ dataset and storing it in such a way, that the training can be launched immediately.

FER+ data

FER+ processed data ready for training and testing.

Resnet34 weights

Weights that correspond to the best experiment so far.