Masood Krohy

CEO at PatternedScience | Ph.D. | Data Science Platform Architect (Big Data & ML) | Advanced Solutions for Quantitative Finance, Telecom & Big Data Geoscience

[Masood Krohy @MasoodKrohy]

My talk/demo of performing 12K+ #MachineLearning experiments on an @ApacheSpark cluster and using a Parallel Coordinates plot to find the best model. Stay tuned for future demos and more advanced use cases in #DataMining, #BigData & #DistributedComputing

https://t.co/iipkgsjivu

[Masood Krohy @MasoodKrohy]

@Love2Code It would be turtles all the way down 😉

[Chris Holdgraf @choldgraf]

✨big news: we are starting a non-profit!✨

It is called 2i2c, which stands for "The International Interactive Computing Collaboration".

We think it will do a lot of good for the world. Here's a short blog post about our hopes, plans, and next-steps:

https://t.co/PB35VY1FO1

[Masood Krohy @MasoodKrohy]

@choldgraf Nice write-up, Chris! The 100% Python definitely is on my wish-list; I am not too interested in dealing with Ruby gems, etc., but Jekyll is still gonna rule my sites for some time 🤷‍♂️

[Masood Krohy @MasoodKrohy]

2/2

[Masood Krohy @MasoodKrohy]

If you know your data well, Join type hints in Spark 3 can speed things up pretty well. Good summary slides from a presentation by the @databricks team.
https://t.co/XGNU3ERbJ6

1/2
#DataScience #ApacheSpark #BigData #DistributedComputing #DataEngineering

[Masood Krohy @MasoodKrohy]

@zacharylipton Now, that's some sophisticated elaboration!

[Masood Krohy @MasoodKrohy]

Good review of over a dozen ML platforms by Ian Hellström of @D2iQ.

A Tour of End-to-End Machine Learning Platforms
https://t.co/SrN7xF5RxG

#ArtificialIntelligence #MachineLearning #DeepLearning #DataScience

[Sean Law 🇨🇦 @seanmylaw]

I am super excited for this release as it is packed full of unexpected new features and a huge performance boost that should serve most of your day-to-day time series analysis needs! 💪 Check it out and let me know what you think! Please r/t for reach.

[Masood Krohy @MasoodKrohy]

Great talk covering tokenizers and predictions/transformers, among others.

Building a Pipeline for State-of-the-Art Natural Language Processing Using Hugging Face Tools: https://t.co/PVOYi7kFj7

2/2

[Masood Krohy @MasoodKrohy]

High-level API of the @huggingface's NLP library has really made some common tasks a breeze. This is a sentiment analysis example from a talk by their team at the recent Spark+AI summit.

#MachineLearning #DeepLearning #DataScience #NaturalLanguageProcessing #NLP
1/

[Masood Krohy @MasoodKrohy]

Site: https://t.co/gBULP6SLHI

How to convert your existing model: https://t.co/W1odyl4E0x

I'm curious to learn about applications where this has been used in areas other than image classification (and w/ frameworks other than PyTorch).

2/2

[Masood Krohy @MasoodKrohy]

Neural-Backed Decision Trees (NBDT) mixes the power of #deeplearning w/ the interpretability of decision trees, with minimal impact on accuracy. Best of both worlds? Interesting discovery.

Paper: https://t.co/xDo2ZP1JP8

#ArtificialIntelligence #MachineLearning #DataScience
1/

[Masood Krohy @MasoodKrohy]

JAX (& its Flax neural library) is the front runner as a next-gen framework. Numpy-like API w/ automatic parallelization, GPU/TPU execution, JIT compilation & differentiation.

Good intro by @jakevdp @SciPyConf:
https://t.co/PK7OPcKcbs

#ArtificialIntelligence #MachineLearning

[Masood Krohy @MasoodKrohy]

Partial use of FP16 gets you a 2-4x speedup in the aggregation portion, w/o losing model precision. Other 2 tricks involve NCCL which currently need some manual fine-turning.

Scaling TensorFlow 2 models to multi-worker GPUs (TF Dev Summit '20)
https://t.co/vmWr2dN5oR

2/2

[Masood Krohy @MasoodKrohy]

ML training in multi-worker, multi-GPU setup got a good boost in #TensorFlow 2.2

Storing model variables in FP32, but doing the gradient aggregation between different workers in FP16 is the easiest trick.

#ArtificialIntelligence #MachineLearning #DeepLearning #DataScience

1/2

[Masood Krohy @MasoodKrohy]

@jeremyjkun You don't have to necessarily label your data. If you're looking for a recurring pattern/anomaly, even if it's very rare, take a look at @stumpy_dev
You don't have to know the pattern beforehand. It's not ML.

[Masood Krohy @MasoodKrohy]

The plot is also used in MLFlow.

A year ago, I also gave a talk showcasing how I was using the plot to optimize a model with Spark: https://t.co/tFdByuqlRb
Glad to see major players caught up; it's a super useful technique.

3/3

[Masood Krohy @MasoodKrohy]

Interesting to observe that ML model hyperparameter tuning with a Parallel Coordinates plot is becoming standard.

TensorBoard guide: Tuning with the HParams Dashboard
https://t.co/BlKQsbAS0Z

#ArtificialIntelligence #MachineLearning #DeepLearning #DataScience #TensorFlow

1/3

[Masood Krohy @MasoodKrohy]

TensorBoard has also nice features for uploading logs to the interweb:
Collaborative ML with https://t.co/QrxWEiWK4F (TF Dev Summit '20)
https://t.co/Iat8NDhaph

2/3

[Masood Krohy @MasoodKrohy]

@random_forests Nice initiative!
I had also done some interpretability work for OD in images with Class Activation Maps, some 3 years ago, which you might find useful. TF 1.x code, but pretty results still :)
https://t.co/ZX6JzTmcLG