{
  "cells": [
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "%matplotlib inline"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "\n\n# microTVM Host-Driven AoT\n**Authors**:\n[Mehrdad Hessar](https://github.com/mehrdadh),\n[Alan MacDonald](https://github.com/alanmacd)\n\nThis tutorial is showcasing microTVM host-driven AoT compilation with\na TFLite model. AoTExecutor reduces the overhead of parsing graph at runtime \ncompared to GraphExecutor. Also, we can have better memory management using ahead \nof time compilation. This tutorial can be executed on a x86 CPU using C runtime (CRT)\nor on Zephyr platform on a microcontroller/board supported by Zephyr.\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "import numpy as np\nimport pathlib\nimport json\nimport os\n\nimport tvm\nfrom tvm import relay\nfrom tvm.relay.backend import Executor, Runtime\nfrom tvm.contrib.download import download_testdata"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "## Import a TFLite model\n\nTo begin with, download and import a Keyword Spotting TFLite model.\nThis model is originally from [MLPerf Tiny repository](https://github.com/mlcommons/tiny).\nTo test this model, we use samples from [KWS dataset provided by Google](https://ai.googleblog.com/2017/08/launching-speech-commands-dataset.html).\n\n**Note:** By default this tutorial runs on x86 CPU using CRT, if you would like to run on Zephyr platform\nyou need to export `TVM_MICRO_USE_HW` environment variable.\n\n\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "use_physical_hw = bool(os.getenv(\"TVM_MICRO_USE_HW\"))\nMODEL_URL = \"https://github.com/tlc-pack/web-data/raw/main/testdata/microTVM/model/keyword_spotting_quant.tflite\"\nMODEL_PATH = download_testdata(MODEL_URL, \"keyword_spotting_quant.tflite\", module=\"model\")\nSAMPLE_URL = \"https://github.com/tlc-pack/web-data/raw/main/testdata/microTVM/data/keyword_spotting_int8_6.pyc.npy\"\nSAMPLE_PATH = download_testdata(SAMPLE_URL, \"keyword_spotting_int8_6.pyc.npy\", module=\"data\")\n\ntflite_model_buf = open(MODEL_PATH, \"rb\").read()\ntry:\n    import tflite\n\n    tflite_model = tflite.Model.GetRootAsModel(tflite_model_buf, 0)\nexcept AttributeError:\n    import tflite.Model\n\n    tflite_model = tflite.Model.Model.GetRootAsModel(tflite_model_buf, 0)\n\ninput_shape = (1, 49, 10, 1)\nINPUT_NAME = \"input_1\"\nrelay_mod, params = relay.frontend.from_tflite(\n    tflite_model, shape_dict={INPUT_NAME: input_shape}, dtype_dict={INPUT_NAME: \"int8\"}\n)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "## Defining the target\n\nNow we need to define the target, runtime and executor. In this tutorial, we focused on\nusing AOT host driven executor. We use the host micro target which is for running a model\non x86 CPU using CRT runtime or running a model with Zephyr platform on qemu_x86 simulator\nboard. In the case of a physical microcontroller, we get the target model for the physical\nboard (E.g. nucleo_l4r5zi) and pass it to `tvm.target.target.micro` to create a full\nmicro target.\n\n\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "# Use the C runtime (crt) and enable static linking by setting system-lib to True\nRUNTIME = Runtime(\"crt\", {\"system-lib\": True})\n\n# Simulate a microcontroller on the host machine. Uses the main() from `src/runtime/crt/host/main.cc <https://github.com/apache/tvm/blob/main/src/runtime/crt/host/main.cc>`_.\n# To use physical hardware, replace \"host\" with something matching your hardware.\nTARGET = tvm.target.target.micro(\"host\")\n\n# Use the AOT executor rather than graph or vm executors. Don't use unpacked API or C calling style.\nEXECUTOR = Executor(\"aot\")\n\nif use_physical_hw:\n    boards_file = pathlib.Path(tvm.micro.get_microtvm_template_projects(\"zephyr\")) / \"boards.json\"\n    with open(boards_file) as f:\n        boards = json.load(f)\n    BOARD = os.getenv(\"TVM_MICRO_BOARD\", default=\"nucleo_l4r5zi\")\n    TARGET = tvm.target.target.micro(boards[BOARD][\"model\"])"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "## Compile the model\n\nNow, we compile the model for the target:\n\n\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "with tvm.transform.PassContext(opt_level=3, config={\"tir.disable_vectorize\": True}):\n    module = tvm.relay.build(\n        relay_mod, target=TARGET, params=params, runtime=RUNTIME, executor=EXECUTOR\n    )"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "## Create a microTVM project\n\nNow that we have the compiled model as an IRModule, we need to create a firmware project\nto use the compiled model with microTVM. To do this, we use Project API. We have defined\nCRT and Zephyr microTVM template projects which are used for x86 CPU and Zephyr boards\nrespectively.\n\n\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "template_project_path = pathlib.Path(tvm.micro.get_microtvm_template_projects(\"crt\"))\nproject_options = {}  # You can use options to provide platform-specific options through TVM.\n\nif use_physical_hw:\n    template_project_path = pathlib.Path(tvm.micro.get_microtvm_template_projects(\"zephyr\"))\n    project_options = {\"project_type\": \"host_driven\", \"zephyr_board\": BOARD}\n\ntemp_dir = tvm.contrib.utils.tempdir()\ngenerated_project_dir = temp_dir / \"project\"\nproject = tvm.micro.generate_project(\n    template_project_path, module, generated_project_dir, project_options\n)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "## Build, flash and execute the model\nNext, we build the microTVM project and flash it. Flash step is specific to\nphysical microcontrollers and it is skipped if it is simulating a microcontroller\nvia the host main.cc or if a Zephyr emulated board is selected as the target.\nNext, we define the labels for the model output and execute the model with a\nsample with expected value of 6 (label: left).\n\n\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "project.build()\nproject.flash()\n\nlabels = [\n    \"_silence_\",\n    \"_unknown_\",\n    \"yes\",\n    \"no\",\n    \"up\",\n    \"down\",\n    \"left\",\n    \"right\",\n    \"on\",\n    \"off\",\n    \"stop\",\n    \"go\",\n]\nwith tvm.micro.Session(project.transport()) as session:\n    aot_executor = tvm.runtime.executor.aot_executor.AotModule(session.create_aot_executor())\n    sample = np.load(SAMPLE_PATH)\n    aot_executor.get_input(INPUT_NAME).copyfrom(sample)\n    aot_executor.run()\n    result = aot_executor.get_output(0).numpy()\n    print(f\"Label is `{labels[np.argmax(result)]}` with index `{np.argmax(result)}`\")\n#\n# Output:\n# Label is `left` with index `6`\n#"
      ]
    }
  ],
  "metadata": {
    "kernelspec": {
      "display_name": "Python 3",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.7.5"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}