Quickstart: Introduction to Diffgram

Diffgram's mission is to make AI Accessible and Practical. We believe Artificial Intelligence (AI) should be in every system because it frees us from the drudgery of routine knowledge tasks — leading to more creative and varied work.

Diffgram is AI/ML data annotation for images and video. Diffgram supports annotation for many machine learning tasks including Object Detection and Semantic Segmentation. For advanced users Diffgram is also data orchestration, storage, and event driven updates.

Diffgram transforms a manual copy and paste your team is already doing (or soon will be) into a software driven one. Similar to the way CI/CD processes bring order to development.

Studies show that real world AI projects often iterate on datasets. Annotation Interfaces and Providers generally place the burden of managing this on the data science team. While this can work to get initial prototypes out the door, it creates a block - the data science team. Custom scripting, copy and pasting, and in house projects end up distracting from valuable work. Diffgram fills this gap by providing core data mechanisms.


See Quickstart: Installation of Diffgram Open Core

Diffgram provides multiple ways to access the platform:

  • Open Core Installation
  • Shared service @ diffgram.com
  • Enterprise (Including private deploy, fully managed, and completely self hosted)

A core idea in Diffgram is to assume the data will be continually updated, such as adding new data to an existing set, updating a known element, or adding new sets entirely. A concrete example of this is getting top level labels and spatial locations from a prediction (eg Road Sign), and then having humans annotate attributes. There are two core objects, the Dataset and the Task Template, in Diffgram. They relate to each other to create useful work.

Diffgram is complimentary to Model Training Platforms and you can use any Workforce.

1. Setup a Dataset

A Dataset in Diffgram can be watched by Task Templates. This means you can define a Dataset upfront in Diffgram, and then import data continually. Diffgram automatically manages syncing files and creating new Tasks.

2. Setup a Task Template

A Task Template watches datasets and outputs to a datasets. Multiple Task Templates may be strung together to create non-linear data pipelines. This can be done in advance and on demand as the need arises, such as changing Label templates.

3. Iterate and Learn

  • Try creating a task template that watches the output dataset of a prior task template.

Video Quickstart for Streaming Data Pipelines

Quickstart playlist



Need help?

If you have any questions feel free to email us [email protected]
And for enterprise sales requests please get in touch here.

Did this page help you?