Compile and execute micro_speech example on Redboard

[aczhang9]

micro_speech example contains a model that is trained to recognize verbal "Yes" and "No".

• [x] compile and execute micro_speech example on mac

  make -f tensorflow/lite/experimental/micro/tools/make/Makefile micro_speech
  tensorflow/lite/experimental/micro/tools/make/gen/osx_arm64/bin/micro_speech

• [x] compile and execute micro_speech example for Redboard
• [x] verify execution of micro_speech on Redboard matches execution on mac

[aczhang9]

Model on mac accurately detects "Yes," "No," and "Silence." Model is much more sensitive to "Yes" than "No." Also all scores are > 200.

To compile and execute micro_speech example on mac:

1. git clone git@github.com:robotastic/tensorflow.git
2. compile: (source code in https://github.com/robotastic/tensorflow/tree/master/tensorflow/lite/experimental/micro/examples/micro_speech)

   cd tensorflow
   make -f tensorflow/lite/experimental/micro/tools/make/Makefile micro_speech

3. execute:

   tensorflow/lite/experimental/micro/tools/make/gen/osx_arm64/bin/micro_speech

Model observations:

• model accurately detects "Yes," "No," and "Silence."
• model is much more sensitive to "Yes" than "No."
• all scores are >200
• after the first yes/no is heard, the model no longer detects "silence." During periods of silence, the terminal does not output any inferences

  
  Heard silence (205) @1696ms

...

Heard yes (202) @4312ms

Heard yes (201) @21768ms

[aczhang9]

For compilation and execution on the Redboard, I based example code from: https://github.com/robotastic/tensorflow/tree/sparkfun_artemis/tensorflow/lite/experimental/micro/examples/micro_speech

Significant changes in audio_provider.cpp are needed in order to implement functions that access hardware APIs specific to the Redboard.

Model can correctly detect "Yes" and "No" when they are spoken but during silent periods the model erroneously detects "No" and "Unknown."

[aczhang9]

Debugging attempted:

• re-writing audio_provider.cpp using PDM class from Apollo3 library: https://github.com/aczhang9/tflite-micro-arduino-examples/blob/2.4.0-ALPHA-precompiled/examples/micro_speech/audio_provider.cpp
• using model weights that are on current main of Arduino TF Lite library: https://github.com/tensorflow/tflite-micro-arduino-examples/blob/main/examples/micro_speech/micro_features_model.cpp but weights here are int8 and rest of source code and Arduino library is expecting weights of uint8.

Debugging to do:

• [x] compile this branch for the board directly https://github.com/robotastic/tensorflow/tree/sparkfun_artemis/tensorflow/lite/experimental/micro/examples/micro_speech but need to link cross compiler to makefile
• [ ] investigate why model erroneously outputs "no" and "unknown" during silence
• [x] run tests https://github.com/robotastic/tensorflow/tree/master/tensorflow/lite/experimental/micro/examples/micro_speech#build-the-tests
• [x] run mac example with same source files used to compile for Redboard

[aczhang9]

Accurately working micro_speech steps:

1. clone this repo: git@github.com:robotastic/tensorflow.git and checkout sparkfun_artemis branch
2. manually download https://developer.arm.com/-/media/Files/downloads/gnu-rm/7-2018q2/gcc-arm-none-eabi-7-2018-q2-update-mac.tar.bz2 to tensorflow/lite/experimental/micro/tools/make/downloads/
3. Compile: make -f tensorflow/lite/experimental/micro/tools/make/Makefile TARGET=sparkfun_redboard_artemis micro_speech_bin
4. Upload binary to board:

   - Use a jumper wire to connect 3.3V to the right side of the resistor above LED 13.
   - While connected, hit the reset button.
   - While connected, run the command below: 
   python3 tensorflow/lite/experimental/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/boards_sfe/common/tools_sfe/ambiq/ambiq_bin2board.py --bin tensorflow/lite/experimental/micro/tools/make/gen/sparkfun_redboard_artemis_cortex-m4/bin/micro_speech.bin --load-address-blob 0x20000 --magic-num 0xCB -o main_nonsecure_ota --version 0x0 --load-address-wired 0xC000 -i 6 --options 0x1 -b 115200 -port /dev/cu.usbserial-110 -r 2 -v

   If python script errors about module/package not found, check python3 --version and make sure module is installed under the same version pip3 show <package>.

5. Use screen as a serial monitor screen /dev/tty.usbserial-110 115200.

Observations:

• Model is much more sensitive to "yes" than "no." This seems to match other people's experience with this model.
• Line 379 in audio_provider.cc caused an Mbed OS fault when I compiled the example using PlatformIO but the program runs fine under this repo.

To do:

• [ ] build PlatformIO version using Apollo3 library downloaded from this Makefile Note: The method of uploading the binary to the board above uses the Ambiq Bootloader and the SparkFun Variable Loader was used for the PlatformIO projects. Read here to facilitate switching between the two bootloaders.