Habr.com / TowardsDataScience post for our dataset

In addition to a github release and a medium post, we also made habr.com post:
- https://habr.com/ru/post/450760/

Also our post was accepted to an editor's pick part of TDS:
- http://bit.ly/ru_open_stt

Share / give us a star / clap if you have not already!

Original release
https://github.com/snakers4/open_stt/

#deep_learning
#data_science
#dataset

New in our Open STT dataset

https://github.com/snakers4/open_stt#updates

- An mp3 version of the dataset;
- A torrent for mp3 dataset;
- A torrent for the original wav dataset;
- Benchmarks on the public dataset / files with "poor" annotation marked;

#deep_learning
#data_science
#dataset

New version of our open STT dataset - 0.5, now in beta

Please share and repost!

https://github.com/snakers4/open_stt/releases/tag/v0.5-beta

What is new?
- A new domain - radio (1000+ new hours);
- A larger YouTube dataset with 1000+ additional hours;
- A small (300 hours) YouTube dataset downloaded in maximum quality;
- Ground truth validation sets for YouTube / books / public calls manually annotated;
- Now we will start to focus on actually cleaning and distilling the dataset. We have published a second list of "bad" data;

I'm back from vacation)

#deep_learning
#data_science
#dataset

Trying to migrate to JupyterLab from Jupyter Notebook?

Some time ago I noticed that the Jupyter extensions project was more or less frozen => JupyterLab obviously is trying to shift community attention to npm / nodejs plugins.

Again (like 6-12 months ago) I tried to do this.

This time Lab is more mature:
- Now at version >1;
- Now they have built-in package manager;
- They have some of the most necessary extensions (i.e. git, toc, google drive, etc);
- UI got polished a bit, but window in a window still produces a bit of mental friction. Only the most popular file formats are supported. Text editor inherited the best features, but it is still a bit rudimentary;
- Full screen width by default;
- Some useful things (like codefolding) are now turned on in settings json file;
- Using these extensions is a bit of a chore in edge cases (i.e. some user permission problems / you have to re-build an app each time you add an extensions);

But I could not switch mostly for one reason - this one
- https://github.com/jupyterlab/jupyterlab/issues/2275#issuecomment-498323475

If you have a Jupyter environment it is very easy to switch. For me, before it was:

# 5.6 because otherwise I have a bug with installing extensions
RUN conda install notebook=5.6

RUN pip install git+https://github.com/ipython-contrib/jupyter_contrib_nbextensions && \
    jupyter contrib nbextension install --user

CMD jupyter notebook --port=8888 --ip=0.0.0.0 --no-browser

And it just became:

RUN conda install -c conda-forge jupyterlab

CMD jupyter lab --port=8888 --ip=0.0.0.0 --no-browser

#data_science

Full IDE in a browser?

Almost)

You all know all the pros and cons of:
- IDEs (PyCharm);
- Advanced text editors (Atom, Sublime Text);
- Interactive environments (notebook / lab, Atom + Hydrogen);

I personally dislike local IDEs - not because connecting to a remote / remote kernel / remote interpreter is a bit of a chore. Setting up is easy, but always thinking about what is synced and what is not - is just pain. Also when your daily driver machine is on Windows, using Linux subsystem all the time with Windows paths is just pain. (Also I diskile bulky interfaces, but this is just a habit and it depends).

But what if I told you there is a third option? =)
If you work as a team on a remote machine / set of machines?

TLDR - you can run a modern web "IDE" (it is something between Atom and real IDE - less bulky, but less functions) in a browser.
Now you can just run it with one command.

Pros:
- It is open source (though shipped as a part of some enterprise packages like Eclipse Che);
- Pre-built images available;
- It is extendible - new modules get released - you can build yourself or just find a build;
- It has extensive linting, python language server (just a standard library though);
- It has full text search ... kind of;
- Follow definition in your code;
- Docstrings and auto-complete work for your modules and standard library (not for you packages);

Looks cool af!
If they ship a build with a remote python kernel, then it will be a perfect option for teams!

I hope it will not follow a path taken by another crowd favourite similar web editor (it was purhcased by Amazon).

Links
- Website;
- Pre-built apps for python;
- Language server they are using;

#data_science

An ideal remote IDE?

Joking?
No, looks like VScode recently got its remote development extensions (it was only in insiders build a couple of months ago) working just right.

I tried remote-ssh extension and it looks quite polished. No syncing your large data folders and loading all python dependencies locally for hours.

The problem? It took me an hour just to open ssh session under Windows properly (permissions and Linux folder path substitution is hell on Windows). When I opened it - it worked like a charm.

So for now (it is personal) - best tools are in my opinion:
- Notebooks - for exploration and testing;
- VScode for codebase;
- Atom - for local scripts;

#data_science

Managing your DS / ML environment neatly and in style

If you have a sophisticated environment that you need to do DS / ML / DL, then using a set of Docker images may be a good idea.
You can also tap into a vast community of amazing and well-maintained Dockerhub repositories (e.g. nvidia, pytorch).

But what you have to do this for several people? And use it with a proper IDE via ssh?
A well-known features of Docker include copy on write and user "forwarding". If you approach naively, each user will store his own images, which take quite some space.
And also you have to make your ssh daemon works inside of a container as a second service.

So I solved these "challenges" and created 2 public layers so far:
- Basic DS / ML layer - FROM aveysov/ml_images:layer-0 - from dockerfile;
- DS / ML libraries - FROM aveysov/ml_images:layer-0- from dockerfile;

Your final dockerfile may look something like this just pulling from any of those layers.
Note that when building this, you will need to pass your UID as a variable, e.g.:

docker build --build-arg NB_UID=1000 -t av_final_layer -f Layer_final.dockerfile .

When launched, this launched a notebook with extensions. You can just exec into the machine itself to run scripts or use an ssh daemon inside (do not forget to add your ssh key and service ssh start).


#deep_learning
#data_science

Extreme NLP network miniaturization

Tried some plain RNNs on a custom in the wild NER task.
The dataset is huge - literally infinite, but manually generated to mimick in-the-wild data.

I use EmbeddingBag + 1m n-grams (an optimal cut-off). Yeah, on NER / classification it is a handy trick that makes your pipeline totally misprint / error / OOV agnostic. Also FAIR themselves just guessed this too. Very cool! Just add PyTorch and you are golden.

What is interesting:
- Model works with embedding sizes 300, 100, 50 and even 5! 5 is dangerously close to OHE, but doing OHE on 1m n-grams kind-of does not make sense;
- Model works with various hidden sizes
- Naturally all of the models run on CPU very fast, but the smallest model also is very light in terms of its weights;
- The only difference is - convergence time. It kind of scales as a log of model size, i.e. model with 5 takes 5-7x more time to converge compared to model with 50. I wonder what if I use embedding size of 1?;

As added bonus - you can just store such miniature model in git w/o lfs.
What is with training transformers on US$250k worth of compute credits you say?)

#nlp
#data_science
#deep_learning

My foray into the STT Dark Forest

My tongue-in-cheek article on ML in general, and how to make your STT model train 3-4x faster with 4-5x less weights with the same quality

https://spark-in.me/post/stt-dark-forest

#data_science
#deep_learning
#stt

Poor man's ensembling techniques

So you want to improve your model's performance a bit.
Ensembling helps. But as is ... it's useful only on Kaggle competitions, where people stack over9000 networks trained on 100MB of data.

But for real life usage / production, there exist ensembling techniques, that do not require significant computation cost increase (!).
All of this is not mainstream yet, but it may work on you dataset!
Especially if your task is easy and the dataset is small.

- SWA (proven to work, usually used as a last stage when training a model);
- Lookahead optimizer (kind of new, not thoroughly tested);
- Multi-Sample Dropout (seems like a cheap ensemble, should work for classification);

Applicability will vary with your task.
Plain vanilla classification can use all of these, s2s networks probably only partially.

#data_science
#deep_learning

Playing with name NER

Premise

So, I needed to separate street names that are actual name + surname. Do not ask me why.
Yeah I know that maybe 70% of streets are human names more or less.
So you need 99% precision and at least 30-40% recall.
Or you can imagine a creepy soviet name like Трактор.

So, today making a NER parser is easy, take out our favourite framework (plan PyTorch ofc) of choice.
Even use FastText or something even less true. Add data and boom you have it.

The pain

But not so fast. Turns our there is a reason why cutting out proper names is a pain.
For Russian there is the natasha library, but since it works on YARGY, it has some assumptions about data structure.
I.e. names should be capitalized, come in pairs (name - surname), etc etc - I did not look their rules under the hood, but I would write it like this.

So probably this would be a name - Иван Иванов
But this probably would not ванечка иванофф

Is it bad?
Ofc no, it just assumes some stuff that may not hold for your dataset.
And yeah it works for streets just fine.

Also recognizing a proper name without context does not really work. And good luck finding (or generating) corpora for that.

Why deep learning may not work

So I downloaded some free databases with names (VK.com respects your secutity lol - the 100M leaked database is available, but useless, too much noise) and surnames.
Got 700k surnames of different origin, around 100-200k male and female names. Used just random words from CC + wiki + taiga for hard negative mining.
Got 92% accuracy on 4 classes (just word, female name, male name, surname) with some naive models.

... and it works .... kind of. If you give it 10M unique word forms, it can distinguish name-like stuff in 90% of cases.
But for addresses it is useless more or less and heuristics from natasha work much better.

The moral

- A tool that works on one case may be 90% useless on another;
- Heuristics have very high precision, low recall and are fragile;
- Neural networks are superior, but you should match your artifically created dataset to the real data (it may take a month to pull off properly);
- In any case, properly cracking both approaches may take time, but both heuristics and NNs are very fast to create, but sometimes 3 plain rules give you 100% precision with 10% recall and sometimes generating a fake dataset that matches your domain is a no-brainer. It depends.

#data_science
#nlp
#deep_learning

Easiest solutions to manage configs for ML models

When you have a lot of experiments, you need to minimize your code bulk and manage model configs concisely.
(This also kind of can be done via CLI parameters, but usually these things complement each other)

I know 3 ways:

(0) dicts + kwargs + dotdicts
(1) [attr](https://github.com/python-attrs/attrs)
(2) new python 3.7 [DataClass](https://docs.python.org/3/library/dataclasses.html) (which is very similar to attr)

Which one do you use?

#data_science

Streamlit vs. viola vs. panel vs. dash vs. bokeh server

TLDR - make scientific web-apps via python only w/o any web-programming (i.e. django, tornado).

Dash
- Mostly for BI
- Also a paid product
- Looks like the new Tableau
- Serving and out-of-the-box scaling options

Bokeh server
- Mostly plotting (very flexible, unlimited capabilities)
- High entry cost, though bokeh is kind of easy to use
- Also should scale well

Panel
- A bokeh server wrapper with a lot of capabilities for geo + templates

Streamlit
- The nicest looking app for interactive ML apps (maybe even annotation)
- Has pre-built styles and grid
- Limited only to its pre-built widgets
- Built on tornado with a very specific data model incompatible with the majority of available widgets
- Supposed to scale well - built on top of tornado

Viola
- If it runs in a notebook - it will run in viola
- Just turns a notebook into a server
- The app with the most promise for DS / ML
- Scales kind of meh - you need to run a jupyter kernel for each user - also takes some time to spin up a kernel
- Fully benefits from a rich ecosystem of jupyter / python / widgets
- In theory has customizable grid and CSS, but does not come pre-built with this => higher barrier to entry

Also most of these apps have no authentication buil-in.

More details:

- A nice summary here;
- A very detailed pros and cons summary of Streamlit + Viola. Also a very in-depth detailed discussion;
- Also awesome streamlit boilerplate is awesome;

#data_science

Collapsible Headings now in JupyterLab

- https://github.com/aquirdTurtle/Collapsible_Headings

The only thing that kept me from switching!
Please share your favourite plugins for JupyterLab!

#data_science