FloydHub: Deep learning platform for productive data science teams

Deep learning platform for productive data science teams - train and deploy your models collaboratively.

Workspaces is a new IDE for DL from FloydHub - Think of Workspaces as your persistent, on-demand deep learning computer in the cloud. You can run Jupyter notebooks, Python scripts, access the terminal, and much more. You can seamlessly toggle between CPU and GPU machines while you're working, right when you're ready for that extra computing power.

Hi Everyone! 👋 Our team at FloydHub is excited to launch our new product - Workspaces(https://blog.floydhub.com/worksp...). Since launching FloydHub just over a year ago we have learned a lot about how data scientists and engineers are building deep learning and machine learning models. We have built a new cloud IDE from the ground up to solve all the major pain points in building your ML models in the cloud. - We are adopting *JupyterLab* as the core of workspaces. JupyterLab is a fully extensible interactive computing environment - you can run both notebooks and python scripts using the terminal. It is very cool and has a much better experience than vanilla Jupyter notebooks. - You can write your algorithm on a low cost machine and when you are ready to train your model, you can switch to a GPU machine at the click of a button. GPU can greatly reduce the training time and help you find the best model for your problem. - You have access to Tensorboard to track training metrics. Plus all the machine level metrics tracked are available from the web interface. - Finally, you can easily search among of thousands of public datasets already uploaded to FloydHub by our community. Then easily download any of them in to your workspace and start using them in your project. We believe that workspaces will greatly simplify your ML workflow and help you focus on building ML algorithms. Workspaces support both the latest versions of Tensorflow and PyTorch. We are also launching some cool starter projects - click link below to create your first workspace and try them out yourselves. I'll be online all day to answer any questions you may have. https://www.floydhub.com/explore...

Loving the new IDE!! 🚀🚀🚀This makes DL a lot more accessible. The 'mounting-data' feature is brilliant. Switching between instances is also a lot more intuitive. The github integration is sweet. I'm curious about the new direction and how it fits with other software/DL services. I think you used to compare yourself with Github and now you have a closer integration with them. Hence, are you moving away from version control and focusing more on the IDE? If so, are you planning on integrating the version control with Github? On the deployment side of things, are you looking into adding support for a scalable infrastructure such as Kubernetes? Ideally, my model runs on a CPU and when someone logs on to my website, it spins up a GPU and keeps adding GPUs based on requests. Last note, I just watched Karpathy's talk on Software 2.0 and IDE's for DL: https://www.figure-eight.com/bui... . He puts a lot of emphasis on tools for building datasets instead of models. What's your take on it and are you working on any products for datasets?

Thanks @emilwallner for the feedback! FloydHub will continue to provide by versioning for your training. i.e., the platform will continue to track the version of code, data and parameters you used to create your model. But the code can come from any source - including GitHub. We are seeing a big adoption of GitHub from the deep learning community - so we are making it very easy to start your FloydHub experience directly from a GitHub project. Stay tuned for more features in this area! A deep learning project goes through the following 5 stages: - Exploration and data collection - Build the first version of the model on sample data - Train on large datasets to identify the best performing model - Deploy the model and use it - Share the model with others We are tackling these problems from top-down. Workspaces will become the core functionality that ties all these stages together. We are planning to focus on model deployment and sharing features in the next few weeks. We are planning to design a lambda style serving option this time. From our users' feedback, data wrangling is definitely a big pain point. FloydHub provides data versioning for users who need to track every change they make to their dataset. But this could take time and resources - especially in the early stages of the project where the data changes a lot. So we are planning to build a more flexible data management solution again using workspaces. JupyterLab has one of the best interfaces to work with different kinds of data files (csv, images etc.) - so we believe this will be a big improvement for FloydHub users.

@narenst 💯 Thanks!!

Currently evaluating FloydHub in comparison to eg. Colab. There's a gravity well around Google ecosystem, but so many things they could do better that FH seems to be working on.

Pros:

7-day dedicated machines (vs. 12h for Colab)... great bootstrap experience w/ public datasets

Cons:

lots of collaboration features, but comment system could use love

FloydHub

Deep learning platform for productive data science teams

FloydHub Workspaces

FloydHub Launches

Do you use FloydHub?

Review FloydHub?

Reviews of FloydHub