Synthetic data

The Synthetic data tab lets you generate image, audio, and time-series data and add it to your dataset.

​

How generation works

Every entry in the Synthetic data tab is a block. To run one:

1. Select a block for the kind of data you need.
2. Configure the parameters: how much data to generate, how to label it, and whatever the model needs, such as a text prompt.
3. Add credentials if the block calls an external provider (OpenAI, Replicate, and similar). The API key is stored once as a secret and reused on later runs. Developer accounts manage these under Account settings -> Secrets; enterprise users manage them in organization secrets.
4. Run the generation. Blocks that call an external provider run against your own account there, so the data you request has a cost attached.
5. Review the result. Generated samples land in your dataset with the label you specified, and from then on are ordinary samples you can filter, edit, relabel, or remove.

​

Using synthetic data well

Synthetic data is most useful for classes that are hard to collect in real life: rare faults, unsafe conditions, or scenarios you cannot stage.

• Keep your test set real. A model tested only on generated data shows how well it learned the generator’s output, which might not match real conditions.
• Tag what you generated, using metadata such as source: synthetic, so you can filter it in or out later when comparing experiments.
• Watch the balance. Generating thousands of samples for one class can turn a small real dataset into a mostly-synthetic one.

See the synthetic data concept page for when synthetic data is appropriate.

​

Built-in blocks

The blocks list below are available directly within Studio. You can also create your own custom synthetic data blocks if you have specific needs.

• DALL-E Image Generation Block
• FLUX Pro Image Generation Block
• Time-series Data Augmentation Block
• Whisper Keyword Spotting Generation Block

To use these blocks, navigate to the Synthetic data tab under the Data acquisition page in your project. From there, you can select the desired block and follow the prompts to generate synthetic data for your datasets.

​

DALL-E image generation block

DALL-E is a generative model that creates images from text descriptions. With the DALL-E block in the Synthetic data tab, you can generate image datasets for your projects. For more information, see the DALL-E tutorial. Select the DALL-E image generation block from the dropdown menu. Enter your OpenAI API key and a prompt, such as “A photo of a factory worker wearing a hard hat”, or “aerial view images of deserted streets”. Then, complete the rest of the fields as appropriate. Click Generate data to create the images.

​

Flux Pro image generation block

Flux Pro is a generative model that can create high-quality images based on textual descriptions. With the Flux Pro block in the Synthetic data tab, you can generate image datasets for your projects. The model runs on Replicate, so you need a Replicate account and API key. Select the Flux Pro image generation block from the dropdown menu. Enter your Replicate API key and a prompt, such as “A realistic top-down image of bunch of small and fully assembled printed circuit boards on a flat conveyor belt with long depth of field”. Then, complete the rest of the fields as appropriate. Click Generate data to create the images.

​

Time-series data augmentation block

Time-series data augmentation is a technique developed by Edge Impulse to create new time-series data from your existing datasets. With the time-series data augmentation block, you can easily generate additional time-series data to enhance your datasets and improve model performance. Select the time-series data augmentation block from the dropdown menu. Then configure each parameter as needed. Click Generate data to create the time-series samples. For a detailed explanation of each parameter, please refer to the Time-series data augmentation transformation block documentation.

​

Whisper keyword spotting generation block

The Whisper keyword spotting block uses the OpenAI text-to-speech API to generate spoken samples of a keyword. Use it to build audio datasets for voice recognition and command-and-control applications. Select the Whisper keyword spotting generation block from the dropdown menu. Enter your OpenAI API key and a keyword, such as “Hello Edge!”. Then, complete the rest of the fields as appropriate. Click Generate data to create the voice samples.

​

Custom synthetic data blocks

If none of the blocks from Edge Impulse fit your needs, you can modify them or develop from scratch to create a custom synthetic data block. This allows you to to integrate your custom generative models for unique project requirements. See the Custom synthetic data blocks page for more information.

​

Additional resources

• DALL-E tutorial
• Time-series data augmentation transformation block
• Custom synthetic data blocks