Train Your Own Deep Net  --  How-to Video

Train deep nets that can work on low-end hardware, use only a few kB of memory while drawing power in the sub-µW range.

A download link to the trained net will be emailed to you about a day after you upload the data.

This portal lives on the Google Cloud and requires a Google or G Suite account for sign-in. If you use Gmail, you already have a Google account.

After you have signed in, click on Continue.

Application Domain

The Infxl net is ideal for energy-constrained IoT edge devices and wearables. It is targeted at data from temperature, pressure, vibration, acceleration, speed, angular velocity, tilt, flow, smell, taste, and environmental sensors.

Ultra-Low Power, Fast, Compact Deep Nets

The Infxl net runs well on low-end hardware, but it can provide 10x throughput of conventional nets on a higher-end MCU, DSP, or FPGA. It will be sent to you as optimized C code, free from floating-point ops and multiplications.

Trained nets in C:  MCU  FPGA  Multicore

Training Data Files

Data must be organized in the form of 4 CSV files: training-features (max size 14 MB), training-labels (4 MB), validation-features (7 MB), and validation-labels (2 MB). They must consist of integers only and not have headings or missing values.

Each label must have its own column, e.g. two columns for a two-class problem.

Binary columns must be mapped to the set {-127, 127}. Continuous-valued columns must be mapped to the range [-127, 127]. For example, if continuous values are in the range [0, 1], they are to be mapped using int(round(254 * value)) - 127

Sample pre-processed files:  Features  Labels

  .  
  .