A3
alibi-detect
A3 | alibi-detect | |
---|---|---|
1 | 9 | |
9 | 2,085 | |
- | 1.6% | |
0.0 | 7.6 | |
almost 2 years ago | 10 days ago | |
Python | Python | |
GNU General Public License v3.0 or later | GNU General Public License v3.0 or later |
Stars - the number of stars that a project has on GitHub. Growth - month over month growth in stars.
Activity is a relative number indicating how actively a project is being developed. Recent commits have higher weight than older ones.
For example, an activity of 9.0 indicates that a project is amongst the top 10% of the most actively developed projects that we are tracking.
A3
-
[P] Looking for Resources on Anomaly Detection
As the code is available as well, you can directly test the approach and see if it is a viable solution for the telemetry data: https://github.com/Fraunhofer-AISEC/A3/
alibi-detect
-
Exploring Open-Source Alternatives to Landing AI for Robust MLOps
Numerous tools exist for detecting anomalies in time series data, but Alibi Detect stood out to me, particularly for its capabilities and its compatibility with both TensorFlow and PyTorch backends.
- Looking for recommendations to monitor / detect data drifts over time
-
[D] Distributions to represent an Image Dataset
That is, to see whether a test image belongs in the distribution of the training images and to provide a routine for special cases. After a bit of reading Ive found that this is related to the field of drift detection in which I tried out alibi-detect . Whereby the training images are trained by an autoencoder and any subsequent drift will be flagged by the AE.
-
[D] Which statistical test would you use to detect drift in a dataset of images?
Wasserstein distance is not very suitable for drift detection on most problems given that the sample complexity (and estimation error) scales with O(n^(-1/d)) with n the number of instances (100k-10m in your case) and d the feature dimension (192 in your case). More interesting will be to use for instance a detector based on the maximum mean discrepancy (MMD) with estimation error of O(n^(-1/2)). Notice the absence of the feature dimension here. You can find scalable implementations in Alibi Detect (disclosure: I am a contributor): MMD docs, image example. We just added the KeOps backend for the MMD detector to scale and speed up the drift detector further, so if you install from master, you can leverage this backend and easily scale the detector to 1mn instances on e.g. 1 RTX2080Ti GPU. Check this example for more info.
-
Ask HN: Who is hiring? (January 2022)
Seldon | Multiple positions | London/Cambridge UK | Onsite/Remote | Full time | seldon.io
At Seldon we are building industry leading solutions for deploying, monitoring, and explaining machine learning models. We are an open-core company with several successful open source projects like:
* https://github.com/SeldonIO/seldon-core
* https://github.com/SeldonIO/mlserver
* https://github.com/SeldonIO/alibi
* https://github.com/SeldonIO/alibi-detect
* https://github.com/SeldonIO/tempo
We are hiring for a range of positions, including software engineers(go, k8s), ml engineers (python, go), frontend engineers (js), UX designer, and product managers. All open positions can be found at https://www.seldon.io/careers/
- What Machine Learning model monitoring tools can you recommend?
- Ask HN: Who is hiring? (December 2021)
-
[D] How do you deal with covariate shift and concept drift in production?
I work in this area and also contribute to outlier/drift detection library https://github.com/SeldonIO/alibi-detect. To tackle this type of problem, I would strongly encourage following a more principled, fundamentally (statistically) sound approach. So for instance measuring metrics such as the KL-divergence (or many other f-divergences) will not be that informative since it has a lot of undesirable properties for the problem at hand (in order to be informative requires already overlapping distributions P and Q, it is asymmetric, not a real distance metric, will not scale well with data dimensionality etc). So you should probably look at Integral Probability Metrics (IPMs) such as the Maximum Mean Discrepancy (MMD) instead which have much nicer behaviour to monitor drift. I highly recommend the Interpretable Comparison of Distributions and Models NeurIPS workshop talks for more in-depth background.
-
[D] Is this a reasonable assumption in machine learning?
All of the above functionality and more can be easily used under a simple API in https://github.com/SeldonIO/alibi-detect.
What are some alternatives?
awesome-TS-anomaly-detection - List of tools & datasets for anomaly detection on time-series data.
pytorch-widedeep - A flexible package for multimodal-deep-learning to combine tabular data with text and images using Wide and Deep models in Pytorch
NAB - The Numenta Anomaly Benchmark
cleanlab - The standard data-centric AI package for data quality and machine learning with messy, real-world data and labels.
pyod - A Comprehensive and Scalable Python Library for Outlier Detection (Anomaly Detection)
seldon-core - An MLOps framework to package, deploy, monitor and manage thousands of production machine learning models
river - 🌊 Online machine learning in Python
Anomaly_Detection_Tuto - Anomaly detection tutorial on univariate time series with an auto-encoder
conductor - Conductor is a microservices orchestration engine.
susi - SuSi: Python package for unsupervised, supervised and semi-supervised self-organizing maps (SOM)
Lean and Mean Docker containers - Slim(toolkit): Don't change anything in your container image and minify it by up to 30x (and for compiled languages even more) making it secure too! (free and open source)
alibi - Algorithms for explaining machine learning models