Transformers-Tutorials VS anomalib

Most of this tutorial is based on Hugging Face course about Transformers and on Niels Rogge's Transformers tutorials: make sure to check their work and give them a star on GitHub, if you please ❤️

I'm pretraining on 850K grayscale spectrograms of birdsongs. I'm on epoch 400 out of 800 and the loss has declined from about 1.2 to 0.7. I don't really have a sense of what is "good enough" and I guess the only way I can judge is by looking at the reconstruction. I'm doing that using this notebook as a guide and right now it's doing quite badly.

More tutorials can be found at https://github.com/NielsRogge/Transformers-Tutorials.

Ive been looking for an off the shelf encoder-decoder document understanding model for key information extraction. I found a great Huggingface implementation with concise notebook examples. However, the token classification model outputs a list of token labels corresponding bounding boxes for the token, but, not the text contained within the labeled bounding boxes themselves. Am I missing something? LayoutLMv2 describes itself as being capable of information extraction but without extracting the text I feel like it's fallen short of that ambition.

Example Notebooks

Then, when it comes to semi-supervised learning for anomaly detection, I had positive experiences with Anomalib which offers a robust library dedicated to deep learning anomaly detection algorithms. They implemented the latest models with PyTorch and offer tools to benchmark their performance.

Anomalib is an open-source library for unsupervised anomaly detection in images. It offers a collection of state-of-the-art models that can be trained on your specific images.

Anomalib is an open-source deep learning library developed by Intel that makes it easy to benchmark different anomaly detection algorithms on both public and custom datasets, all by simply modifying a config file. As the largest public collection of anomaly detection algorithms and datasets, it has a strong focus on image-based anomaly detection. It’s a comprehensive, end-to-end solution that includes cutting-edge algorithms, relevant evaluation methods, prediction visualizations, hyperparameter optimization, and inference deployment code with Intel’s OpenVINO Toolkit.

try https://github.com/openvinotoolkit/anomalib it's primarily aimed at vision applications but might provide some inspiration

Anomaly detection typically refers to the task of finding unusual or rare items that deviate significantly from what is considered to be the "normal" majority. In this blogpost, we look at image anomalies using PatchCore. Next to indicating which images are anomalous, PatchCore also identifies the most anomalous pixel regions within each image. One big advantage of PatchCore is that it only requires normal images for training, making it attractive for many use cases where abnormal images are rare or expensive to acquire. In some cases, we don't even know all the unusual patterns that we might encounter and training a supervised model is not an option. One example use case is the detection of defects in industrial manufacturing, where most defects are rare by definition as production lines are optimised to produce as few of them as possible. Recent approaches have made significant progress on anomaly detection in images, as demonstrated on the MVTec industrial benchmark dataset. PatchCore, presented at CVPR 2022, is one of the frontrunners in this field. In this blog post we first dive into the inner workings of PatchCore. Next, we apply it to an example in medical imaging to gauge its applicability outside of industrial examples. We use the anomalib library, which was developed by Intel and offers ready-to-use implementations of many recent image anomaly detection methods.

Link to the github repo: https://github.com/openvinotoolkit/anomalib