uTensor is an extreme light-weight Deep-Learning Inference framework built on mbed and Tensorflow.

This project is under going constant development. The development is running on develop branch.

• Mbed CLI
• Tensorflow
• tf-node-viewer (Optional, for graph-weight extraction)
• Mbed-os 5.6+ compatiable boards with at least 256kb of RAM
• SD Card (Must be LESS than 32 GB)
• SD Card reader for the board (Optional if built into the board)

mbed detect to see which target is connect to the board

mbedls -l to list all supported targets

See mbed_app.json

1. Clone the repository
2. Run mbed deploy to download all referenced libraries
3. Insert the prepared SD card to the board (see SD Card Preparation Section)
4. Use mbed compile -t GCC_ARM -m NUCLEO_F767ZI --profile=./build_profile/release.json to build for ST NUCLEO F767ZI. Or, mbed compile -t GCC_ARM -m NUCLEO_F767ZI --profile=./build_profile/release.json -f to compile and flash

The test data has to be loaded to the SD card for the default binary to run:

1. Install python dependencies pip install -r requirements.txt (Note: may have to use pip3)
2. Go to the [project]\TESTS\scripts folder
3. Run python3 compileTestData.py. This will create [project]\TESTS\scripts\testData directory.
4. Copy [project]\TESTS\scripts\testData to the root of your SD card.

The quantized weight and input data are stored in the SD. Setting the serial baud rate to 115200, here is what you should see:

Deep MLP on Mbed (Trained with Tensorflow)

running deep-mlp...
PASSED 0.00000000

prediction: 7

Currently, the binary runs the first sample of the MNIST dataset which contains a handwritten digit of number 7. Ths network architecture is a 3-layer Relu based MLP, as shown below:

The related Tensorflow training script please refer to the node-viewer project.