Skip to content

Latest commit

 

History

History
86 lines (65 loc) · 3.13 KB

WORKFLOW.md

File metadata and controls

86 lines (65 loc) · 3.13 KB
Raw

MicroNets Workflow

Console Output

Almost all of the output from the code will be sent through the onboard radio to your computer with the CrazyFlie Dongle. You will need to download the CrazyFlie client and install it, which can be found here. Make sure to follow all the installation steps so you can run cfclient and open the window.

Then, to see console output:

  1. Turn on the CrazyFlie normally (not in pairing mode).
  2. Press Scan to find the radio device and then press Connect.
  3. Go to the Console tab and you should see the debug output from the Crazyflie.

To print to this console, we can use DEBUG_PRINT or consolePrintf in the code. This is an easy way to see debug output from our TF-Micro model.

Training a Model

Training the Model

First, train your model in tensorflow. You can use Microsoft AirSim with a custom environment or OpenAI gym for creating an environment and training a DQN for the crazyflie. Currently, the only layers that I have tested are fully connected and convolutional layers. I don't think any other layers are supported on the crazyflie.

Quantization

After training the tensorflow model the next steps are freezing the tensorflow graph and quantizing to 8-bit. The process depends on how you defined your tensorflow graph, but examples of freezing can be found in tfmicro/train/freeze_simple.py and tfmicro/train/freeze_conv.py. Then, you need to clone the tensorflow repository, compile it, and quantize with the following command:

cd ${TF_HOME}

bazel run tensorflow/lite/toco:toco -- \
    --input_file=${CHECKPT_DIR}/${MODEL_NAME}.pb \
    --output_file=$(pwd)/${MODEL_NAME}.tflite \
    --input_shapes=1,${NUM_INPUTS} \
    --input_arrays=input \
    --output_arrays='labels_softmax' \
    --inference_type=QUANTIZED_UINT8 \
    --change_concat_input_ranges=false

replacing the variable as necessary to match the parameters of your model.

Deploying the Model

Once the model has been quantized, turn it into a C array and load it in using the TF-Micro API.

xxd -i [tf_micro_model.tflite] >> tfmicro_models.cc

Then, you can load in the model using TF-Micro - again, see the example code in src/deck/drivers/src to see how everything is loaded in.

Writing Deployment Code

To deploy the model, the easiest way is to create a .c file in the src/deck/drivers/src folder. Essentially, we write a code that pretends to be a Crazyflie Extension deck, and when compiling we force the Crazyflie to load that code in first. See tfmicrodemo.c inside this folder for a template. Furthermore, you must edit the Makefile - see all of the comments in the Makefile that are labeled as TF Micro Compilation, you'll have to edit those areas with names for your own C file.

Wrapping TF Micro

All of the original Crazyflie code was compiled in C and TF-Micro requires some C++ extensions, so there are wrappers needed to call the TF-Micro library. The wrappers can be extended - they are at tfmicro/machinelearning.h and tfmicro/machinelearning.cc.