pplonski86

Good for back: https://www.lifespanfitness.com/workplace/treadmill-desks

I have this one with an electric-height desk (custom built) and switch between chair and treadmill. I would recommend it. However, the treadmill, when not used, is taking a lot of place.

What data types do you plan to support? Have you considered supporting datasets with images+labels used in computer vision? How would you like to handle them?

Are you going to support data labeling tasks?

Thank you Amit! I would like to add much more to the application:

1. I would like to add scheduling. 2. I would like to define different types of notebooks: notebooks for exploratory analysis (default one), notebooks for workflow automation, notebooks for packages development (with test driven development), notebooks as REST API 3. I would like to allow to connect different data sources: databases and third party APIs 4. I would like to support Machine Learning APIs

I've seen your work, especially reviewNB - impressive!

In 2016 I've created web service for building Machine Learning models in the cloud. However, I wasn't happy with it because it only allows building ML models on already prepared data. So I've started searching for a nice way for preparing data in a low code manner.

At first, I was considering GUI with drag and drop nodes, where each node represents an operation on data. I wasn't happy with that solution because it can easily get messy (spaghetti of nodes) and it is untestable. So I keep searching ...

It came to me that the best low-code solution will be a code itself, with GUI to generate the code. I was looking for the best way to execute the arbitrary code on the web and I came to Jupyter Notebook.

As a result, I've started working on the MLJAR Studio. It has frontend created in react and redux. It connects to Jupyter Notebook (for computation) and Django server (store and manage the project).

Users can generate code by clicking. All code creates a script that can be executed.

All results are represented as widgets and can be easily shared with others by sending a link.

There is still a lot of work to do: more operations covered with GUI, the interaction between widgets and code, periodic runs of scripts, ability to connect with databases and APIs, and of course support for ML.

In the future, I would like to have two versions of MLJAR Studio: open-source version and pro-version with additional features.

Demo address: http://204.48.28.64:8000/ (you don't need to set account, be careful, it is one user only, so everyone can see your activities and data, and it uses HTTP!, the address will be active for 1 week).

Thank you! This is exactly what I'm looking for. Versions for Windows and Linux https://github.com/SelfControlApp/selfcontrol/wiki/FAQ

Music helps for me. Try to listen music that makes you feel better. I can listen one track in the loop for the whole day and work.

I have a similar problem with my product which offers machine learning as a service. It offers complex data science features (building ML models) for non-coders. From my experience, I can tell that it is hard to find users that don't have enough technical background to train the model by themself and enough background to understand what to do with ML model.

Who is your target user? I think that if user is able to install jupyter notebook, pandas and load data with python, there is a high chance that user can also search pandas documentation and write few lines of code.

Anyway, I like the idea of making UI for Pandas, but I think that there should be more comprehensive software to make data science easier for non-coders.

The source code from the tutorial is available at https://github.com/pplonski/my_ml_service

This tutorial shows how to create web service in Python and Django to serve multiple Machine Learning models. It is different (more advanced) from most of the tutorials available on the internet:

- it keeps information about many ML models in the web service. There can be several ML models available at the same endpoint with different versions. What is more, there can be many endpoint addresses defined.

- it stores information about requests sent to the ML models, this can be used later for model testing and audit.

- it has tests included for ML code and server code.

- it can run A/B tests between different versions of ML models.

What hardware did you use to produce such photos?

MLJAR | Łapy, Białystok, Poland or remote | full stack | at least 50% time | https://mljar.com

We are building tool for data analysis without coding.

Looking for full stack developer (python, django, channels, celery, redis, postgres, react, plotlyjs) to join us. My email is piotr at mljar.com

Congratulations on launching! I had a similar feeling that there is a missing tool, which will allow to code and see data at the same time. I have started doing some sketches of the app very similar to yours.

In my opinion Spreadsheets and Jupyter have many advantages and disadvantages - by combining both, there can be a nice tool for data science.

It will be nice to add more examples to your app.

What is your business model?

It depends on what algorithms you use in AutoML. If you decide to use simple algorithms: logistic regression, decision tree, random forest then you will have a simple model very quickly. Using Neural Networks in AutoML requires much more computational resources.

Random Forest doesn't increase overfit error when adding more trees. I did the experiment on toy dataset to check it https://mljar.com/blog/random-forest-overfitting/

What is more, Leo Breimain wrote on his website: "Random forests does not overfit. You can run as many trees as you want" https://www.stat.berkeley.edu/~breiman/RandomForests/cc_home...

I don't like such brute-force approach. Even if you have low cost of computing the number of possible combinations of hyperparameters is huge! Google Cloud AutoML Tables solution cost is 20 USD per 1 hour of computing (I guess that's because of inefficient Neural Architecture Search algorithm). Running few ML experiment can easily end with huge bill.

You are right! The more trees in the RF the more stable the model is. What is more, you got feature importance for free!

But still, NN has much more hyperparameters than RF and training good NN model requires far more experience.

You are right, CatBoost is an amazing algorithm. However, you will be shocked when you will talk with many older senior data scientist, that never heard of Xgboost or LightGBM. CatBoost for them is far too new.

Google has released AutoML service for structured datasets (AutoML Tables, currently in Beta). I've compared my open-source AutoML python package mljar-supervised with Google AutoML Tables on 6 datasets from openml.org. The results of Google AutoML Tables are poor (I was using 1-hour training). I think that the reason for poor results is a short training time. I think that the 1-hour of training on 92 machines in parallel is not enough for Google optimization algorithm. I believe they are using some Neural Architecture Search algorithm which is not very efficient. Maybe training for a longer time will give better results but the cost of 1-hour training is ~20USD so it can easily skyrocket.

Google has released AutoML service for structured datasets (AutoML Tables, currently in Beta). I've compared my open-source AutoML python package mljar-supervised with Google AutoML Tables on 6 datasets from openml.org. The results of Google AutoML Tables are poor (I was using 1-hour training). I think that the reason for poor results is a short training time. I think that the 1-hour of training on 92 machines in parallel is not enough for Google optimization algorithm. I believe they are using some Neural Architecture Search algorithm which is not very efficient. Maybe training for a longer time will give better results but the cost of 1-hour training is ~20USD so it can easily skyrocket.

Recently I did a list of AutoML open and close source packages https://mljar.com/blog/automl-software-list/

The first thing is, the Neural Networks are not golden ML algorithm that is working always the best. I'm not surprised to see GBDT methods working better than NN. The NNs are very powerful because they can accept the wide-range of data types (tabular, images, voice) which is not possible with GBDT (maybe not yet).

The second, I think architecture search fo NNs is very inefficient right now. Most of the methods are training new NN in every attempt. There is a lack for method that can start with any NN architecture and grow/prune it dynamically. Take a look at this poster, with dynamic topology adjustemnt for MLP http://www.ire.pw.edu.pl/~rsulej/NetMaker/icaisc/icaisc_post...

Anthony, can Kaggle make this dataset public or make competition public and enable post-competition submission? It will be beneficial for AutoML research.

I'm working on AutoML solution, that is available in the cloud (https://mljar.com). What is more, the core of my AutoML is open source (https://github.com/mljar/mljar-supervised) - both are easy to use. The cloud version has user interface so you need just to upload data and do few clicks. You don't need to have programming knowledge. For python package, you need to know how to manipulate data in python (basic numpy and pandas).

I've made many tests of my AutoML solution and I observe that Neural Networks doesn't work the best on tabular dataset (maybe not training long enough, but I dont have 2500CPU hours). I really prefer gradient boosting methods (xgboost, catboost, lightgbm) on tabular data. They are much faster than NN and require less preprocessing (no feature scaling).

I like the service. I've seen that you are going to ofer paid version. How are you going to keep two repositories (open source and pro) in sync?

Hi Gerardo!

Congratulations! The product looks very neat. I got questions:

1. What does the logo mean?

2. Do you have two servers, one for REST and one for web sockets? (I'm trying to use similar architecture in my product)

Thanks! Sure, piotr at mljar.com :)

I've built mljar for myself. When I was starting working on it (2016) there were no such solutions or they were very expensive (over 50k/year).

I want to have a service where I can train many models and be able to check every model (for example check learning curves). I want a solution that can train many models in parallel in the cloud, so I don't need to heat my laptop and dont need to wait a lot. I think I achieved this. Is it better than other solutions? Hmmm, it is very similar to others (at the end, they all train some ML models), but ...

After creating the AutoML solution I come to the conclusion that AutoML is broken: https://pplonski.github.io/automatic-machine-learning-is-bro... - even if you can easily train ML model (good or bad - doesn't matter), there is still hard to use/apply machine learning in real life.

Right now, I think that AutoML is just a brick in the solution that should offer automatization. There should be a service similar to Zappier but with machine learning - you can join your data and services with ML models which live in your data ecosystem and use ML for automatization.

DataRobot has over 220M in total funding. They have resources for sales. Though, I will think about it.