Elevator pitch

Every year ten thousand papers are published to top deep learning conferences. The state-of-the-art models trained as part of the research paper experiments get lost in the github repos and never see the light of day.

We built an App Store for centralizing and serving these pre-trained deep learning models.

  • The Research Scientists upload their pretrained models on our platform.
  • We provide the infrastructure to serve these models to end-users.

Monetization:

  • Users pay $0.1 to $10 for 1000 API calls
  • Researchers earn money when their model gets used.

At the end of day, Researchers have an incentive to get their models used in real world to make an impact, while earning income. And deep learning enthusiasts would love to access state-of-the-art models (trained by top researchers in the world) in their applications. Our platform bridges this gap.

Inspiration

Each year, thousands of scientific papers on deep learning are published, but most of these state-of-the-art models are just gathering dust in GitHub repositories. The headache of creating and maintaining the infrastructure to deploy these models deter most researchers from bringing their work to the public. Meanwhile, technologists in all fields are hoping to apply these models but just cannot find readily consumable models.

What it does

Deeplify provides a convenient way for researchers to serve their trained models via an API by taking care of deployment, upkeep, outreach, and monetization. These APIs are presented in the central platform for developers to easily discover and consume.

How we built it

After a Brainstorming session, we started prototyping the front end and implementing it with React.

For the backend, we developed a cloud layer that can run deep learning jobs in a scalable fashion using kubernetes cluster. The cluster can autoscale (scale up/down) the number of nodes based on the assigned workload.

We have a script that can create a kubernetes cluster to meet specific needs - GPU type (K80, P100, V100), Number of GPUs per node (1, 2, 4, 8). We also built a generic docker image for PyTorch training that can load user specified model and data from firebase and publish the results back to firebase.

Backend workflow:

  • User model is stored on firebase storage
  • Batch of images submitted through API endpoint also get uploaded to firebase storage
  • After the data is available on firebase storage, a kubernetes pod is spun up.
  • Kubernetes pod loads the model and images from firebase storage. It passes the images through the model and gets the predicted output.
  • The predicted output is written back to firebase database

Coming up with a design for the cloud layer such that it can autoscale (up/down) based on demand was non-trivial.

Challenges we ran into

Agreeing on the clear product definition and creating a final solution that combines the feature of an individual API for each model together with a platform for searching for different pre-trained models.

Coming up with a design for the cloud layer such that it can autoscale (up/down) based on demand was non-trivial.

Accomplishments that we're proud of

Our approach of tackling a field with increasing importance and still finding a niece solution that hasn't been published before, which gives scholars and data scientists an incentive to publish their studies.

What we learned

Brainstorming and full understanding and agreement in the team about the idea is the most important part of starting such a project

What's next for deeplify

Getting the product known by the desired userbase by advertising especially at universities and gathering information from actual users of our service.

Share this project:

Updates

posted an update

Elevator Pitch: Every year ten thousand papers are published to top deep learning conferences. The state-of-the-art models trained as part of the research paper experiments get lost in the github repos and never see the light of day.

We built an App Store for pre-trained deep learning models.

  • The Research Scientists upload their pretrained models on our platform.
  • We provide the infrastructure to serve these models to end-users.

Monetization:

  • Users pay $0.1 to $10 for 1000 API calls
  • Researchers earn money when their model gets used.

At the end of day, Researchers have an incentive to get their models used in real world to make an impact, while earning income. And deep learning enthusiasts would love to access state-of-the-art models (trained by top researchers in the world) in their applications. Our platform bridges this gap.

Log in or sign up for Devpost to join the conversation.