Model Exporting

Introduction

Export mode is used to convert the trained TensorFlow model into a format that can be used for deployment onto Ambiq’s family of SoCs. Currently, the command will convert the TensorFlow model into both TensorFlow Lite (TFL) and TensorFlow Lite for micro-controller (TFLM) variants. The command will also verify the models’ outputs match. The activations and weights can be quantized by configuring the quantization section in the configuration file or by setting the quantization parameter in the code.

1. Load the configuration data (e.g. configuration.json)
2. Load the test data (e.g. test.pkl)
3. Load the trained model (e.g. model.keras)
4. Quantize the model (e.g. 16x8)
5. Convert the model (e.g. TFL, TFLM)
6. Verify the models’ outputs match
7. Save artifacts (e.g. model.tflite)

Example configuration

modes-export-1.json150 lines · JSONDownload

Show full exampleCollapse example

modes-export-1.json

{
    "name": "arr-2-eff-sm",
    "project": "hk-rhythm-2",
    "job_dir": "./results/arr-2-eff-sm",
    "verbose": 2,
    "datasets": [
        {
            "name": "ptbxl",
            "params": {
                "path": "./datasets/ptbxl"
            }
        }
    ],
    "num_classes": 2,
    "class_map": {
        "0": 0,
        "7": 1,
        "8": 1
    },
    "class_names": [
        "NORMAL",
        "AFIB/AFL"
    ],
    "class_weights": "balanced",
    "sampling_rate": 100,
    "frame_size": 512,
    "samples_per_patient": [
        10,
        10
    ],
    "val_samples_per_patient": [
        5,
        5
    ],
    "test_samples_per_patient": [
        5,
        5
    ],
    "val_patients": 0.2,
    "val_size": 20000,
    "test_size": 20000,
    "batch_size": 256,
    "buffer_size": 20000,
    "epochs": 100,
    "steps_per_epoch": 50,
    "val_metric": "loss",
    "lr_rate": 0.001,
    "lr_cycles": 1,
    "threshold": 0.75,
    "val_metric_threshold": 0.98,
    "tflm_var_name": "g_rhythm_model",
    "tflm_file": "rhythm_model_buffer.h",
    "backend": "pc",
    "demo_size": 896,
    "display_report": true,
    "quantization": {
        "qat": false,
        "format": "INT8",
        "io_type": "int8",
        "conversion": "CONCRETE",
        "debug": false
    },
    "preprocesses": [
        {
            "name": "layer_norm",
            "params": {
                "epsilon": 0.01,
                "name": "znorm"
            }
        }
    ],
    "augmentations": [],
    "model_file": "model.keras",
    "use_logits": false,
    "architecture": {
        "name": "efficientnetv2",
        "params": {
            "input_filters": 16,
            "input_kernel_size": [
                1,
                9
            ],
            "input_strides": [
                1,
                2
            ],
            "blocks": [
                {
                    "filters": 24,
                    "depth": 2,
                    "kernel_size": [
                        1,
                        9
                    ],
                    "strides": [
                        1,
                        2
                    ],
                    "ex_ratio": 1,
                    "se_ratio": 2
                },
                {
                    "filters": 32,
                    "depth": 2,
                    "kernel_size": [
                        1,
                        9
                    ],
                    "strides": [
                        1,
                        2
                    ],
                    "ex_ratio": 1,
                    "se_ratio": 2
                },
                {
                    "filters": 40,
                    "depth": 2,
                    "kernel_size": [
                        1,
                        9
                    ],
                    "strides": [
                        1,
                        2
                    ],
                    "ex_ratio": 1,
                    "se_ratio": 2
                },
                {
                    "filters": 48,
                    "depth": 1,
                    "kernel_size": [
                        1,
                        9
                    ],
                    "strides": [
                        1,
                        2
                    ],
                    "ex_ratio": 1,
                    "se_ratio": 2
                }
            ],
            "output_filters": 0,
            "include_top": true,
            "use_logits": true
        }
    }
}

{
    "name": "arr-2-eff-sm",
    "project": "hk-rhythm-2",
    "job_dir": "./results/arr-2-eff-sm",
    "verbose": 2,
    "datasets": [
        {
            "name": "ptbxl",

… 142 more lines

---

Usage

CLI

The following command will export a rhythm model using the reference configuration.

Terminal

heartkit --task rhythm --mode export --config ./configuration.json

Python

The model can be evaluated using the following snippet:

Python example

task = hk.TaskFactory.get("rhythm")

params = hk.HKTaskParams(...)

task.export(params)

Example configuration

modes-export-4.py82 lines · PYTHONDownload

Show full exampleCollapse example

modes-export-4.py

hk.HKTaskParams(
    name="arr-2-eff-sm",
    project="hk-rhythm-2",
    job_dir="./results/arr-2-eff-sm",
    verbose=2,
    datasets=[hk.NamedParams(
        name="ptbxl",
        params=dict(
            path="./datasets/ptbxl"
        )
    )],
    num_classes=2,
    class_map={
        "0": 0,
        "7": 1,
        "8": 1
    },
    class_names=[
        "NORMAL", "AFIB/AFL"
    ],
    class_weights="balanced",
    sampling_rate=100,
    frame_size=512,
    samples_per_patient=[10, 10],
    val_samples_per_patient=[5, 5],
    test_samples_per_patient=[5, 5],
    val_patients=0.20,
    val_size=20000,
    test_size=20000,
    batch_size=256,
    buffer_size=20000,
    epochs=100,
    steps_per_epoch=50,
    val_metric="loss",
    lr_rate=1e-3,
    lr_cycles=1,
    threshold=0.75,
    val_metric_threshold=0.98,
    tflm_var_name="g_rhythm_model",
    tflm_file="rhythm_model_buffer.h",
    backend="pc",
    demo_size=896,
    display_report=True,
    quantization=hk.QuantizationParams(
        qat=False,
        format="INT8",
        io_type="int8",
        conversion="CONCRETE",
        debug=False
    ),
    preprocesses=[
        hk.NamedParams(
            name="layer_norm",
            params=dict(
                epsilon=0.01,
                name="znorm"
            )
        )
    ],
    augmentations=[
    ],
    model_file="model.keras",
    use_logits=False,
    architecture=hk.NamedParams(
        name="efficientnetv2",
        params=dict(
            input_filters=16,
            input_kernel_size=[1, 9],
            input_strides=[1, 2],
            blocks=[
                {"filters": 24, "depth": 2, "kernel_size": [1, 9], "strides": [1, 2], "ex_ratio": 1,  "se_ratio": 2},
                {"filters": 32, "depth": 2, "kernel_size": [1, 9], "strides": [1, 2], "ex_ratio": 1,  "se_ratio": 2},
                {"filters": 40, "depth": 2, "kernel_size": [1, 9], "strides": [1, 2], "ex_ratio": 1,  "se_ratio": 2},
                {"filters": 48, "depth": 1, "kernel_size": [1, 9], "strides": [1, 2], "ex_ratio": 1,  "se_ratio": 2}
            ],
            output_filters=0,
            include_top=True,
            use_logits=True
        )
    }
)

hk.HKTaskParams(
    name="arr-2-eff-sm",
    project="hk-rhythm-2",
    job_dir="./results/arr-2-eff-sm",
    verbose=2,
    datasets=[hk.NamedParams(
        name="ptbxl",

… 74 more lines

---

Arguments

Please refer to HKTaskParams for the list of arguments that can be used with the export command.