{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "arduino_tinyml_workshop.ipynb",
      "provenance": [],
      "collapsed_sections": [],
      "toc_visible": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "f92-4Hjy7kA8",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://www.arduino.cc/\"><img src=\"https://raw.githubusercontent.com/sandeepmistry/aimldevfest-workshop-2019/master/images/Arduino_logo_R_highquality.png\" width=200/></a>\n",
        "# Tiny ML on Arduino\n",
        "## Gesture recognition tutorial\n",
        " * Sandeep Mistry - Arduino\n",
        " * Don Coleman - Chariot Solutions\n",
        "\n",
        " \n",
        "https://github.com/arduino/ArduinoTensorFlowLiteTutorials/"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "uvDA8AK7QOq-",
        "colab_type": "text"
      },
      "source": [
        "## Setup Python Environment \n",
        "\n",
        "Install up the Python libraries and Linux tools for the code in the notebook."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Y2gs-PL4xDkZ",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "# Setup environment\n",
        "!apt-get -qq install xxd\n",
        "!pip install pandas numpy matplotlib\n",
        "%tensorflow_version 2.x\n",
        "!pip install tensorflow"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "9lwkeshJk7dg",
        "colab_type": "text"
      },
      "source": [
        "# Upload Data\n",
        "\n",
        "1. If necessary, open the panel on the left side of Colab by clicking on the __>__\n",
        "1. Select the files tab in the left panel\n",
        "1. Drag the `punch.csv` and `flex.csv` files from your computer to the tab to upload them into colab."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Eh9yve14gUyD",
        "colab_type": "text"
      },
      "source": [
        "# Graph Data (optional)\n",
        "\n",
        "Plot the CSV data on two separate graphs, acceleration and gyroscope, because each data set has different units and scale."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "I65ukChEgyNp",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "import matplotlib.pyplot as plt\n",
        "import numpy as np\n",
        "import pandas as pd\n",
        "\n",
        "filename = \"punch.csv\"\n",
        "\n",
        "df = pd.read_csv(\"/content/\" + filename)\n",
        "\n",
        "index = range(1, len(df['aX']) + 1)\n",
        "\n",
        "plt.rcParams[\"figure.figsize\"] = (20,10)\n",
        "\n",
        "plt.plot(index, df['aX'], 'g.', label='x', linestyle='solid', marker=',')\n",
        "plt.plot(index, df['aY'], 'b.', label='y', linestyle='solid', marker=',')\n",
        "plt.plot(index, df['aZ'], 'r.', label='z', linestyle='solid', marker=',')\n",
        "plt.title(\"Acceleration\")\n",
        "plt.xlabel(\"Sample #\")\n",
        "plt.ylabel(\"Acceleration (G)\")\n",
        "plt.legend()\n",
        "plt.show()\n",
        "\n",
        "plt.plot(index, df['gX'], 'g.', label='x', linestyle='solid', marker=',')\n",
        "plt.plot(index, df['gY'], 'b.', label='y', linestyle='solid', marker=',')\n",
        "plt.plot(index, df['gZ'], 'r.', label='z', linestyle='solid', marker=',')\n",
        "plt.title(\"Gyroscope\")\n",
        "plt.xlabel(\"Sample #\")\n",
        "plt.ylabel(\"Gyroscope (deg/sec)\")\n",
        "plt.legend()\n",
        "plt.show()\n"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "kSxUeYPNQbOg",
        "colab_type": "text"
      },
      "source": [
        "# Train Neural Network\n",
        "\n",
        "\n",
        "\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Gxk414PU3oy3",
        "colab_type": "text"
      },
      "source": [
        "## Parse and prepare the data\n",
        "\n",
        "Parse the CSV files and transforms them to a format that can be used to train the fully connected neural network.\n",
        "\n",
        "If you've recorded additional gestures, update the `GESTURES` list with the names of the additional CSV files.\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "AGChd1FAk5_j",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "import matplotlib.pyplot as plt\n",
        "import numpy as np\n",
        "import pandas as pd\n",
        "import tensorflow as tf\n",
        "\n",
        "print(f\"TensorFlow version = {tf.__version__}\\n\")\n",
        "\n",
        "# Set a fixed random seed value, for reproducibility, this will allow us to get\n",
        "# the same random numbers each time the notebook is run\n",
        "SEED = 1337\n",
        "np.random.seed(SEED)\n",
        "tf.random.set_seed(SEED)\n",
        "\n",
        "# the list of gestures \n",
        "GESTURES = [\n",
        "    \"punch\",\n",
        "    \"flex\"\n",
        "]\n",
        "\n",
        "SAMPLES_PER_GESTURE = 119\n",
        "\n",
        "NUM_GESTURES = len(GESTURES)\n",
        "\n",
        "# create a one-hot encoded matrix that is used in the output\n",
        "ONE_HOT_ENCODED_GESTURES = np.eye(NUM_GESTURES)\n",
        "\n",
        "inputs = []\n",
        "outputs = []\n",
        "\n",
        "# read each csv file and push an input and output\n",
        "for gesture_index in range(NUM_GESTURES):\n",
        "  gesture = GESTURES[gesture_index]\n",
        "  print(f\"Processing index {gesture_index} for gesture '{gesture}'.\")\n",
        "  \n",
        "  output = ONE_HOT_ENCODED_GESTURES[gesture_index]\n",
        "  \n",
        "  df = pd.read_csv(\"/content/\" + gesture + \".csv\")\n",
        "\n",
        "  # get rid of pesky empty value lines of csv which cause NaN inputs to TensorFlow\n",
        "  df = df.dropna()\n",
        "  df = df.reset_index(drop=True)\n",
        "  \n",
        "  # calculate the number of gesture recordings in the file\n",
        "  num_recordings = int(df.shape[0] / SAMPLES_PER_GESTURE)\n",
        "  \n",
        "  print(f\"\\tThere are {num_recordings} recordings of the {gesture} gesture.\")\n",
        "  \n",
        "  for i in range(num_recordings):\n",
        "    tensor = []\n",
        "    for j in range(SAMPLES_PER_GESTURE):\n",
        "      index = i * SAMPLES_PER_GESTURE + j\n",
        "      # normalize the input data, between 0 to 1:\n",
        "      # - acceleration is between: -4 to +4\n",
        "      # - gyroscope is between: -2000 to +2000\n",
        "      tensor += [\n",
        "          (df['aX'][index] + 4) / 8,\n",
        "          (df['aY'][index] + 4) / 8,\n",
        "          (df['aZ'][index] + 4) / 8,\n",
        "          (df['gX'][index] + 2000) / 4000,\n",
        "          (df['gY'][index] + 2000) / 4000,\n",
        "          (df['gZ'][index] + 2000) / 4000\n",
        "      ]\n",
        "\n",
        "    inputs.append(tensor)\n",
        "    outputs.append(output)\n",
        "\n",
        "# convert the list to numpy array\n",
        "inputs = np.array(inputs)\n",
        "outputs = np.array(outputs)\n",
        "\n",
        "print(\"Data set parsing and preparation complete.\")"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "d5_61831d5AM",
        "colab_type": "text"
      },
      "source": [
        "## Randomize and split the input and output pairs for training\n",
        "\n",
        "Randomly split input and output pairs into sets of data: 60% for training, 20% for validation, and 20% for testing.\n",
        "\n",
        "  - the training set is used to train the model\n",
        "  - the validation set is used to measure how well the model is performing during training\n",
        "  - the testing set is used to test the model after training"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "QfNEmUZMeIEx",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "# Randomize the order of the inputs, so they can be evenly distributed for training, testing, and validation\n",
        "# https://stackoverflow.com/a/37710486/2020087\n",
        "num_inputs = len(inputs)\n",
        "randomize = np.arange(num_inputs)\n",
        "np.random.shuffle(randomize)\n",
        "\n",
        "# Swap the consecutive indexes (0, 1, 2, etc) with the randomized indexes\n",
        "inputs = inputs[randomize]\n",
        "outputs = outputs[randomize]\n",
        "\n",
        "# Split the recordings (group of samples) into three sets: training, testing and validation\n",
        "TRAIN_SPLIT = int(0.6 * num_inputs)\n",
        "TEST_SPLIT = int(0.2 * num_inputs + TRAIN_SPLIT)\n",
        "\n",
        "inputs_train, inputs_test, inputs_validate = np.split(inputs, [TRAIN_SPLIT, TEST_SPLIT])\n",
        "outputs_train, outputs_test, outputs_validate = np.split(outputs, [TRAIN_SPLIT, TEST_SPLIT])\n",
        "\n",
        "print(\"Data set randomization and splitting complete.\")"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "a9g2n41p24nR",
        "colab_type": "text"
      },
      "source": [
        "## Build & Train the Model\n",
        "\n",
        "Build and train a [TensorFlow](https://www.tensorflow.org) model using the high-level [Keras](https://www.tensorflow.org/guide/keras) API."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "kGNFa-lX24Qo",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "# build the model and train it\n",
        "model = tf.keras.Sequential()\n",
        "model.add(tf.keras.layers.Dense(50, activation='relu')) # relu is used for performance\n",
        "model.add(tf.keras.layers.Dense(15, activation='relu'))\n",
        "# the final layer is softmax because we only expect one gesture to occur per input\n",
        "model.add(tf.keras.layers.Dense(NUM_GESTURES, activation='softmax'))\n",
        "model.compile(optimizer='rmsprop', loss='mse', metrics=['mae'])\n",
        "history = model.fit(inputs_train, outputs_train, epochs=600, batch_size=1, validation_data=(inputs_validate, outputs_validate))\n",
        "\n"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "NUDPvaJE1wRE",
        "colab_type": "text"
      },
      "source": [
        "## Verify \n",
        "\n",
        "Graph the models performance vs validation.\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "kxA0zCOaS35v",
        "colab_type": "text"
      },
      "source": [
        "### Graph the loss\n",
        "\n",
        "Graph the loss to see when the model stops improving."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "bvFNHXoQzmcM",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "# increase the size of the graphs. The default size is (6,4).\n",
        "plt.rcParams[\"figure.figsize\"] = (20,10)\n",
        "\n",
        "# graph the loss, the model above is configure to use \"mean squared error\" as the loss function\n",
        "loss = history.history['loss']\n",
        "val_loss = history.history['val_loss']\n",
        "epochs = range(1, len(loss) + 1)\n",
        "plt.plot(epochs, loss, 'g.', label='Training loss')\n",
        "plt.plot(epochs, val_loss, 'b', label='Validation loss')\n",
        "plt.title('Training and validation loss')\n",
        "plt.xlabel('Epochs')\n",
        "plt.ylabel('Loss')\n",
        "plt.legend()\n",
        "plt.show()\n",
        "\n",
        "print(plt.rcParams[\"figure.figsize\"])"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "DG3m-VpE1zOd",
        "colab_type": "text"
      },
      "source": [
        "### Graph the loss again, skipping a bit of the start\n",
        "\n",
        "We'll graph the same data as the previous code cell, but start at index 100 so we can further zoom in once the model starts to converge."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "c3xT7ue2zovd",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "# graph the loss again skipping a bit of the start\n",
        "SKIP = 100\n",
        "plt.plot(epochs[SKIP:], loss[SKIP:], 'g.', label='Training loss')\n",
        "plt.plot(epochs[SKIP:], val_loss[SKIP:], 'b.', label='Validation loss')\n",
        "plt.title('Training and validation loss')\n",
        "plt.xlabel('Epochs')\n",
        "plt.ylabel('Loss')\n",
        "plt.legend()\n",
        "plt.show()"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "CRjvkFQy2RgS",
        "colab_type": "text"
      },
      "source": [
        "### Graph the mean absolute error\n",
        "\n",
        "[Mean absolute error](https://en.wikipedia.org/wiki/Mean_absolute_error) is another metric to judge the performance of the model.\n",
        "\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "mBjCf1-2zx9C",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "# graph of mean absolute error\n",
        "mae = history.history['mae']\n",
        "val_mae = history.history['val_mae']\n",
        "plt.plot(epochs[SKIP:], mae[SKIP:], 'g.', label='Training MAE')\n",
        "plt.plot(epochs[SKIP:], val_mae[SKIP:], 'b.', label='Validation MAE')\n",
        "plt.title('Training and validation mean absolute error')\n",
        "plt.xlabel('Epochs')\n",
        "plt.ylabel('MAE')\n",
        "plt.legend()\n",
        "plt.show()\n"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "guMjtfa42ahM",
        "colab_type": "text"
      },
      "source": [
        "### Run with Test Data\n",
        "Put our test data into the model and compare the predictions vs actual output\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "V3Y0CCWJz2EK",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "# use the model to predict the test inputs\n",
        "predictions = model.predict(inputs_test)\n",
        "\n",
        "# print the predictions and the expected ouputs\n",
        "print(\"predictions =\\n\", np.round(predictions, decimals=3))\n",
        "print(\"actual =\\n\", outputs_test)"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "j7DO6xxXVCym",
        "colab_type": "text"
      },
      "source": [
        "# Convert the Trained Model to Tensor Flow Lite\n",
        "\n",
        "The next cell converts the model to TFlite format. The size in bytes of the model is also printed out."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "0Xn1-Rn9Cp_8",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "# Convert the model to the TensorFlow Lite format without quantization\n",
        "converter = tf.lite.TFLiteConverter.from_keras_model(model)\n",
        "tflite_model = converter.convert()\n",
        "\n",
        "# Save the model to disk\n",
        "open(\"gesture_model.tflite\", \"wb\").write(tflite_model)\n",
        "  \n",
        "import os\n",
        "basic_model_size = os.path.getsize(\"gesture_model.tflite\")\n",
        "print(\"Model is %d bytes\" % basic_model_size)\n",
        "  \n",
        "  "
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "ykccQn7SXrUX",
        "colab_type": "text"
      },
      "source": [
        "## Encode the Model in an Arduino Header File \n",
        "\n",
        "The next cell creates a constant byte array that contains the TFlite model. Import it as a tab with the sketch below."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "9J33uwpNtAku",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "!echo \"const unsigned char model[] __attribute__((aligned(4))) = {\"  > /content/model.h\n",
        "!cat gesture_model.tflite | xxd -i                                  >> /content/model.h\n",
        "!echo \"};\"                                                          >> /content/model.h\n",
        "\n",
        "import os\n",
        "model_h_size = os.path.getsize(\"model.h\")\n",
        "print(f\"Header file, model.h, is {model_h_size:,} bytes.\")\n",
        "print(\"\\nOpen the side panel (refresh if needed). Double click model.h to download the file.\")"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "1eSkHZaLzMId",
        "colab_type": "text"
      },
      "source": [
        "# Classifying IMU Data\n",
        "\n",
        "Now it's time to switch back to the tutorial instructions and run our new model on the Arduino Nano 33 BLE Sense to classify the accelerometer and gyroscope data.\n"
      ]
    }
  ]
}