Train ECG Denosier¶

Date created: 2024/08/13

Last Modified: 2024/07/17

Description: Train, evaluate, and export ECG denoiser model from scratch

Overview¶

In this guide, we will train an ECG denoiser to remove noise and artifacts from raw ECG signals. Once trained, we demonstrate how to evaluate the model and export it for inference for both TF Lite and TF Lite for Micro.

Input

• Sensor: ECG
• Location: Wrist
• Sampling Rate: 100 Hz
• Frame Size: 2.56 seconds

Datasets

• Synthetic: Synthetic ECG signals from PhysioKit
• PTB-XL: The PTB-XL is a large publicly available electrocardiography dataset. It contains 21837 clinical 12-lead ECGs from 18885 patients of 10 second length. The ECGs are sampled at 500 Hz and are annotated by up to two cardiologists.

In [1]:

!pip install -q --disable-pip-version-check heartkit

!pip install -q --disable-pip-version-check heartkit

Setup¶

In [2]:

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
import contextlib
from pathlib import Path
import tempfile
import keras
import heartkit as hk
import numpy as np
import neuralspot_edge as nse
import matplotlib.pyplot as plt

import os os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' import contextlib from pathlib import Path import tempfile import keras import heartkit as hk import numpy as np import neuralspot_edge as nse import matplotlib.pyplot as plt

In [3]:

# Be sure to set the dataset path to the correct location
datasets_dir = Path(os.getenv('HK_DATASET_PATH', './datasets'))

plot_theme = hk.utils.dark_theme
nse.utils.silence_tensorflow()
hk.utils.setup_plotting(plot_theme)
logger = nse.utils.setup_logger(__name__)

# Be sure to set the dataset path to the correct location datasets_dir = Path(os.getenv('HK_DATASET_PATH', './datasets')) plot_theme = hk.utils.dark_theme nse.utils.silence_tensorflow() hk.utils.setup_plotting(plot_theme) logger = nse.utils.setup_logger(__name__)

Create preprocess/augmentation pipeline¶

Since our goal is to denoise ECG signals, we need to create an augmentation pipeline to generate noisy samples.

We will leverage neuralspot-edge preprocessing layers to create the following augmentations:

• Baseline wander: Simulate baseline wander by adding a low frequency sine signal
• Powerline noise: Simulate powerline noise by adding a 50 Hz sinusoidal signal
• Amplitude warp: Simulate amplitude warp by randomly scaling along a low frequency sine wave
• Gaussian noise: Simulate lead noise by adding random noise following a Gaussian distribution
• Background noise: Add real noise captured from NSTDB dataset

In [4]:

preprocesses = [hk.NamedParams(
    name="layer_norm",
    params=dict(
        epsilon=0.01
    )
)]

augmentations = [hk.NamedParams(
    name="random_noise_distortion",
    params=dict(
        amplitude=[0.1, 1.5],
        frequency=[0.5, 1.5],
        name="baseline_wander"
    )
), hk.NamedParams(
    name="random_sine_wave",
    params=dict(
        amplitude=[0, 0.05],
        frequency=[45, 50],
        auto_vectorize=False,
        name="powerline_noise"
    )
), hk.NamedParams(
    name="amplitude_warp",
    params=dict(
        amplitude=[0.9, 1.1],
        frequency=[0.5, 1.5],
        name="amplitude_warp"
    )
), hk.NamedParams(
    name="random_noise",
    params=dict(
        factor=[0.1, 0.5],
        name="random_noise"
    )
), hk.NamedParams(
    name="random_background_noise",
    params=dict(
        amplitude=[0.1, 0.5],
        num_noises=2,
        name="nstdb"
    )
)]

preprocesses = [hk.NamedParams( name="layer_norm", params=dict( epsilon=0.01 ) )] augmentations = [hk.NamedParams( name="random_noise_distortion", params=dict( amplitude=[0.1, 1.5], frequency=[0.5, 1.5], name="baseline_wander" ) ), hk.NamedParams( name="random_sine_wave", params=dict( amplitude=[0, 0.05], frequency=[45, 50], auto_vectorize=False, name="powerline_noise" ) ), hk.NamedParams( name="amplitude_warp", params=dict( amplitude=[0.9, 1.1], frequency=[0.5, 1.5], name="amplitude_warp" ) ), hk.NamedParams( name="random_noise", params=dict( factor=[0.1, 0.5], name="random_noise" ) ), hk.NamedParams( name="random_background_noise", params=dict( amplitude=[0.1, 0.5], num_noises=2, name="nstdb" ) )]

Define TCN model architecture¶

For this task, we are going to leverage a customized TCN model architecture that is smaller and can handle 1D signals. The model consists of 5 TCN blocks with a depth of 1. Each block leverages dilated depthwise-separable convolutions along with inverted expansion and squeeze and excitation layers. The model is followed by a 1D convolutional layer.

In [5]:

mbconv_blocks = [
    dict(depth=1, branch=1, filters=16, kernel=(1, 7), dilation=(1, 1), dropout=0, ex_ratio=1, se_ratio=0, norm="batch"),
    dict(depth=1, branch=1, filters=24, kernel=(1, 7), dilation=(1, 1), dropout=0, ex_ratio=1, se_ratio=2, norm="batch"),
    dict(depth=1, branch=1, filters=32, kernel=(1, 7), dilation=(1, 2), dropout=0, ex_ratio=1, se_ratio=2, norm="batch"),
    dict(depth=1, branch=1, filters=40, kernel=(1, 7), dilation=(1, 4), dropout=0, ex_ratio=1, se_ratio=2, norm="batch"),
    dict(depth=1, branch=1, filters=48, kernel=(1, 7), dilation=(1, 8), dropout=0, ex_ratio=1, se_ratio=2, norm="batch")
]

architecture = dict(
    name="tcn",
    params=dict(
        input_kernel=(1, 7),
        input_norm="batch",
        blocks=mbconv_blocks,
        output_kernel=(1, 7),
        include_top=True,
        use_logits=True,
        model_name="tcn"
    )
)

mbconv_blocks = [ dict(depth=1, branch=1, filters=16, kernel=(1, 7), dilation=(1, 1), dropout=0, ex_ratio=1, se_ratio=0, norm="batch"), dict(depth=1, branch=1, filters=24, kernel=(1, 7), dilation=(1, 1), dropout=0, ex_ratio=1, se_ratio=2, norm="batch"), dict(depth=1, branch=1, filters=32, kernel=(1, 7), dilation=(1, 2), dropout=0, ex_ratio=1, se_ratio=2, norm="batch"), dict(depth=1, branch=1, filters=40, kernel=(1, 7), dilation=(1, 4), dropout=0, ex_ratio=1, se_ratio=2, norm="batch"), dict(depth=1, branch=1, filters=48, kernel=(1, 7), dilation=(1, 8), dropout=0, ex_ratio=1, se_ratio=2, norm="batch") ] architecture = dict( name="tcn", params=dict( input_kernel=(1, 7), input_norm="batch", blocks=mbconv_blocks, output_kernel=(1, 7), include_top=True, use_logits=True, model_name="tcn" ) )

Configure datasets¶

Capturing noise-free ECG signals is challenging due to the presence of various artifacts. Therefore, we use a combination of synthetic and controlled, real-world datasets as our training data. HeartKit exposes an ECG Synthetic dataset generator provided by PhysioKit.

In [6]:

datasets = [
    hk.NamedParams(
        name="ecg-synthetic",
        params=dict(
            num_pts=5000,
            params=dict(
                presets=["SR", "AFIB", "ant_STEMI", "LAHB", "LPHB", "high_take_off", "LBBB", "random_morphology"],
                preset_weights=[24, 8, 1, 1, 1, 1, 1, 0],
                duration=10,
                sample_rate=100,
                heart_rate=[40, 160],
                impedance=[1, 2],
                p_multiplier=[0.7, 1.3],
                t_multiplier=[0.7, 1.3],
                noise_multiplier=[0, 0.01],
                voltage_factor=[800, 1000]
            )
        )
    ),
    hk.NamedParams(
        name="ptbxl",
        params=dict(
            path=datasets_dir / "ptbxl",
        )
    )
]

datasets = [ hk.NamedParams( name="ecg-synthetic", params=dict( num_pts=5000, params=dict( presets=["SR", "AFIB", "ant_STEMI", "LAHB", "LPHB", "high_take_off", "LBBB", "random_morphology"], preset_weights=[24, 8, 1, 1, 1, 1, 1, 0], duration=10, sample_rate=100, heart_rate=[40, 160], impedance=[1, 2], p_multiplier=[0.7, 1.3], t_multiplier=[0.7, 1.3], noise_multiplier=[0, 0.01], voltage_factor=[800, 1000] ) ) ), hk.NamedParams( name="ptbxl", params=dict( path=datasets_dir / "ptbxl", ) ) ]

Task configuration¶

Here we provide the constants that we will use throughout the guide. For better performance, adjust parameters as needed such as BATCH_SIZE, EPOCHS, and LEARNING_RATE.

In [7]:

params = hk.HKTaskParams(
    # Common arguments
    name="hk-ecg-denoiser",
    job_dir=Path(tempfile.gettempdir()) / "hk-ecg-denoiser",
    # Dataset arguments
    datasets=datasets,
    # Signal arguments
    sampling_rate=100,
    frame_size=256,
    # Dataloader arguments
    samples_per_patient=5,
    val_samples_per_patient=10,
    test_samples_per_patient=10,
    # Preprocessing/Augmentation arguments
    preprocesses=preprocesses,
    augmentations=augmentations,
    # Class arguments
    num_classes=1,
    class_map={0: 0},
    class_names=["DENOISE"],
    # Split arguments
    val_patients=0.1,
    val_size=10000,
    test_size=10000,
    val_file="val.pkl",
    test_file="val.pkl",
    # Model arguments
    model_file="model.keras",
    architecture=architecture,
    # Training parameters
    lr_rate=1e-3,
    lr_cycles=1,
    batch_size=256,
    buffer_size=25000,
    epochs=100,
    steps_per_epoch=50,
    val_metric="loss",
    class_weights="balanced",
    # Evaluation arguments
    threshold=0.5,
    val_metric_threshold=0.98,
    # Export parameters
    tflm_var_name="ecg_denoise_flatbuffer",
    tflm_file="ecg_denoise_flatbuffer.h",
    # Demo params
    backend="pc",
    demo_size=800,
    display_report=True,
    # Extra arguments
    verbose=1,
    seed=42
)

params = hk.HKTaskParams( # Common arguments name="hk-ecg-denoiser", job_dir=Path(tempfile.gettempdir()) / "hk-ecg-denoiser", # Dataset arguments datasets=datasets, # Signal arguments sampling_rate=100, frame_size=256, # Dataloader arguments samples_per_patient=5, val_samples_per_patient=10, test_samples_per_patient=10, # Preprocessing/Augmentation arguments preprocesses=preprocesses, augmentations=augmentations, # Class arguments num_classes=1, class_map={0: 0}, class_names=["DENOISE"], # Split arguments val_patients=0.1, val_size=10000, test_size=10000, val_file="val.pkl", test_file="val.pkl", # Model arguments model_file="model.keras", architecture=architecture, # Training parameters lr_rate=1e-3, lr_cycles=1, batch_size=256, buffer_size=25000, epochs=100, steps_per_epoch=50, val_metric="loss", class_weights="balanced", # Evaluation arguments threshold=0.5, val_metric_threshold=0.98, # Export parameters tflm_var_name="ecg_denoise_flatbuffer", tflm_file="ecg_denoise_flatbuffer.h", # Demo params backend="pc", demo_size=800, display_report=True, # Extra arguments verbose=1, seed=42 )

Load denoise task¶

HeartKit provides a TaskFactory that includes a number ready-to-use tasks. Each task provides methods for training, evaluating, exporting, and demoing. We will grab the denoise task and configure it for our use case.

In [8]:

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

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

Download the datasets¶

We will download the synthetic and PTB-XL datasets using heartkit. If already downloaded, this step will be skipped.

In [9]:

task.download(params=params)

task.download(params=params)

Visualize the data¶

Let's visualize a sample ECG signal from the synthetic dataset. Note this contains no noise or artifacts. Augmentations will be applied later to generate noisy samples for training.

In [10]:

ds = hk.DatasetFactory.get(params.datasets[0].name)(
    cacheable=False,
    **params.datasets[0].params
)

ds_gen = ds.signal_generator(
    patient_generator=nse.utils.uniform_id_generator(ds.get_test_patient_ids()),
    frame_size=params.frame_size,
    samples_per_patient=params.samples_per_patient,
    target_rate=params.sampling_rate,
)
ecg = next(ds_gen)

ts = np.arange(0, len(ecg)) / params.sampling_rate
fig, ax = plt.subplots(1, 1, figsize=(9, 4))
ax.plot(ts, ecg, color=plot_theme.primary_color, lw=3)
fig.suptitle("Raw ECG Signal")
ax.set_xlabel("Time (s)")
ax.set_ylabel("Amplitude")
fig.tight_layout()
fig.show()

ds = hk.DatasetFactory.get(params.datasets[0].name)( cacheable=False, **params.datasets[0].params ) ds_gen = ds.signal_generator( patient_generator=nse.utils.uniform_id_generator(ds.get_test_patient_ids()), frame_size=params.frame_size, samples_per_patient=params.samples_per_patient, target_rate=params.sampling_rate, ) ecg = next(ds_gen) ts = np.arange(0, len(ecg)) / params.sampling_rate fig, ax = plt.subplots(1, 1, figsize=(9, 4)) ax.plot(ts, ecg, color=plot_theme.primary_color, lw=3) fig.suptitle("Raw ECG Signal") ax.set_xlabel("Time (s)") ax.set_ylabel("Amplitude") fig.tight_layout() fig.show()

Visualize augmented data¶

Let's visualize the augmented data to understand how the augmentations affect the ECG signals.

In [11]:

preprocessor = hk.datasets.create_augmentation_pipeline(
    augmentations=params.preprocesses,
    sampling_rate=params.sampling_rate,
)

augmenter = hk.datasets.create_augmentation_pipeline(
    augmentations=params.augmentations,
    sampling_rate=params.sampling_rate,
)

preprocessor = hk.datasets.create_augmentation_pipeline( augmentations=params.preprocesses, sampling_rate=params.sampling_rate, ) augmenter = hk.datasets.create_augmentation_pipeline( augmentations=params.augmentations, sampling_rate=params.sampling_rate, )

WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1723838156.202266  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723838156.222145  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723838156.222246  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723838156.223422  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723838156.223495  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723838156.223541  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723838156.268697  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723838156.268787  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723838156.268844  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355

In [12]:

aug_ecg = augmenter(preprocessor(keras.ops.convert_to_tensor(np.reshape(ecg, (1, -1, 1)))), training=True)
aug_ecg = aug_ecg.numpy().squeeze()

ts = np.arange(0, len(aug_ecg)) / params.sampling_rate
fig, ax = plt.subplots(1, 1, figsize=(9, 4))
plt.plot(ts, aug_ecg, color=plot_theme.primary_color, lw=3)
fig.suptitle("Augmented ECG Signal")
ax.set_xlabel("Time (s)")
ax.set_ylabel("Amplitude")
plt.tight_layout()
plt.show()

aug_ecg = augmenter(preprocessor(keras.ops.convert_to_tensor(np.reshape(ecg, (1, -1, 1)))), training=True) aug_ecg = aug_ecg.numpy().squeeze() ts = np.arange(0, len(aug_ecg)) / params.sampling_rate fig, ax = plt.subplots(1, 1, figsize=(9, 4)) plt.plot(ts, aug_ecg, color=plot_theme.primary_color, lw=3) fig.suptitle("Augmented ECG Signal") ax.set_xlabel("Time (s)") ax.set_ylabel("Amplitude") plt.tight_layout() plt.show()

Visualize the model¶

Let's view the first several layers of the model to understand the architecture better.

In [13]:

model = nse.models.tcn.tcn_from_object(
    x=keras.Input(shape=(params.frame_size, 1), name='inputs'),
    params=architecture["params"],
    num_classes=1
)
model.summary(layer_range=('inputs', model.layers[10].name))

model = nse.models.tcn.tcn_from_object( x=keras.Input(shape=(params.frame_size, 1), name='inputs'), params=architecture["params"], num_classes=1 ) model.summary(layer_range=('inputs', model.layers[10].name))

Model: "TCN"

┏━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┓
┃ Layer (type)        ┃ Output Shape      ┃    Param # ┃ Connected to      ┃
┡━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━┩
│ inputs (InputLayer) │ (None, 256, 1)    │          0 │ -                 │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ reshape (Reshape)   │ (None, 1, 256, 1) │          0 │ inputs[0][0]      │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ ENC.CN              │ (None, 1, 256, 1) │          7 │ reshape[0][0]     │
│ (DepthwiseConv2D)   │                   │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ ENC.BN              │ (None, 1, 256, 1) │          4 │ ENC.CN[0][0]      │
│ (BatchNormalizatio… │                   │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ B1.D1.DW.B1.CN      │ (None, 1, 256, 1) │          7 │ ENC.BN[0][0]      │
│ (DepthwiseConv2D)   │                   │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ B1.D1.DW.B1.BN      │ (None, 1, 256, 1) │          4 │ B1.D1.DW.B1.CN[0… │
│ (BatchNormalizatio… │                   │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ B1.D1.DW.ACT        │ (None, 1, 256, 1) │          0 │ B1.D1.DW.B1.BN[0… │
│ (Activation)        │                   │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ B1.D1.PW.B1.CN      │ (None, 1, 256,    │         16 │ B1.D1.DW.ACT[0][… │
│ (Conv2D)            │ 16)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ B1.D1.PW.B1.BN      │ (None, 1, 256,    │         64 │ B1.D1.PW.B1.CN[0… │
│ (BatchNormalizatio… │ 16)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ B1.D1.PW.ACT        │ (None, 1, 256,    │          0 │ B1.D1.PW.B1.BN[0… │
│ (Activation)        │ 16)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ B2.D1.DW.B1.CN      │ (None, 1, 256,    │        112 │ B1.D1.PW.ACT[0][… │
│ (DepthwiseConv2D)   │ 16)               │            │                   │
└─────────────────────┴───────────────────┴────────────┴───────────────────┘

 Total params: 10,223 (39.93 KB)

 Trainable params: 9,675 (37.79 KB)

 Non-trainable params: 548 (2.14 KB)

Train the model¶

In [14]:

task.train(params)

task.train(params)

INFO     Creating synthetic dataset cache with 5000 patients                                   ecg_synthetic.py:159

Building ecg-synthetic cache: 100%|██████████| 5000/5000 [00:57<00:00, 86.91it/s] 

INFO     Validation steps per epoch: 39                                                              datasets.py:85

Training:   0%|           0/100 ETA: ?s,  ?epochs/sWARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1723838225.604155  751478 service.cc:146] XLA service 0x7a52b8001f20 initialized for platform CUDA (this does not guarantee that XLA will be used). Devices:
I0000 00:00:1723838225.604174  751478 service.cc:154]   StreamExecutor device (0): NVIDIA GeForce RTX 4090, Compute Capability 8.9
I0000 00:00:1723838232.858832  751478 device_compiler.h:188] Compiled cluster using XLA!  This line is logged at most once for the lifetime of the process.
Training: 100%|██████████ 100/100 ETA: 00:00s,   1.59s/epochs

39/39 ━━━━━━━━━━━━━━━━━━━━ 0s 975us/step - cos: 0.7118 - loss: 0.0511 - mae: 0.1445 - mse: 0.0452 - snr: 11.9220

INFO     [VAL SET]COS=0.7079, LOSS=0.0528, MAE=0.1466, MSE=0.0469, SNR=11.9038                         train.py:149

Model evaluation¶

Now that we have trained the model, we will evaluate the model on the test dataset. Similar to training, we will provide the high-level configuration to the task process.

In [15]:

task.evaluate(params)

task.evaluate(params)

INFO     Creating synthetic dataset cache with 5000 patients                                   ecg_synthetic.py:159

Building ecg-synthetic cache: 100%|██████████| 5000/5000 [00:57<00:00, 87.16it/s] 

39/39 ━━━━━━━━━━━━━━━━━━━━ 2s 25ms/step - cos: 0.7238 - loss: 0.0443 - mae: 0.1328 - mse: 0.0384 - snr: 12.3671

INFO     [TEST SET] COS=0.7245, LOSS=0.0437, MAE=0.1316, MSE=0.0377, SNR=12.3787                     evaluate.py:37

Export model to TF Lite / TFLM¶

Once we have trained and evaluated the model, we need to export the model into a format that can be used for inference on the edge. Currently, we export the model to TensorFlow Lite flatbuffer format. This will also generate a C header file that can be used with TensorFlow Lite for Microcontrollers (TFLM).

For this model, we will export as a 32-bit floating point model.

NOTE: We utilize CONCRETE mode to lower the model to concrete functions before converting. This is because TF (MLIR) fails to properly lower the dilated convolutional layers.

In [16]:

quantization = hk.QuantizationParams(
    enabled=True,
    format="FP32",
    io_type="float32",
    conversion="CONCRETE",
)
params.quantization = quantization

quantization = hk.QuantizationParams( enabled=True, format="FP32", io_type="float32", conversion="CONCRETE", ) params.quantization = quantization

In [18]:

# TF dumps a lot of info to stdout, so we redirect it to /dev/null
with open(os.devnull, 'w') as devnull:
    with contextlib.redirect_stdout(devnull), contextlib.redirect_stderr(devnull):
        task.export(params)

# TF dumps a lot of info to stdout, so we redirect it to /dev/null with open(os.devnull, 'w') as devnull: with contextlib.redirect_stdout(devnull), contextlib.redirect_stderr(devnull): task.export(params)

INFO     Creating synthetic dataset cache with 5000 patients                                   ecg_synthetic.py:159

[08/16/24 20:02:23] WARNING  WARNING:absl:Please consider providing the trackable_obj argument in the  lite.py:2166
                             from_concrete_functions. Providing without the trackable_obj argument is              
                             deprecated and it will use the deprecated conversion path.                            

INFO     Validating model results                                                                      export.py:83

I0000 00:00:1723838543.688860  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723838543.688944  751181 devices.cc:67] Number of eligible GPUs (core count >= 8, compute capability >= 0.0): 1
I0000 00:00:1723838543.689113  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723838543.689169  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723838543.689214  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723838543.689287  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723838543.689333  751181 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
W0000 00:00:1723838543.815333  751181 tf_tfl_flatbuffer_helpers.cc:392] Ignored output_format.
W0000 00:00:1723838543.815348  751181 tf_tfl_flatbuffer_helpers.cc:395] Ignored drop_control_dependency.
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.

INFO     [TF METRICS] LOSS=0.0396 MAE=0.1357 MSE=0.0396 RMSE=0.1991 COSINE=0.7178                      export.py:90

INFO     [TFL METRICS] LOSS=0.0396 MAE=0.1357 MSE=0.0396 RMSE=0.1991 COSINE=0.7177                     export.py:91

INFO     Validation passed (0.0000)                                                                    export.py:99

ECG Denoising Demo¶

Finally, we will demonstrate how to use the trained ECG denoiser model to remove noise and artifacts from raw ECG signals. We will load a sample ECG signal, add noise to it, and then denoise it using the trained model. We will visualize the original, noisy, and denoised ECG signals to compare the results.

In [19]:

model = nse.models.load_model(params.model_file)

model = nse.models.load_model(params.model_file)

In [30]:

ecg = next(ds_gen)
aug_ecg = augmenter(preprocessor(keras.ops.convert_to_tensor(np.reshape(ecg, (1, -1, 1)))), training=True).numpy().squeeze()
clean_ecg = model.predict(np.reshape(aug_ecg, (1, -1, 1)))
snr = nse.metrics.Snr()
snr.update_state(ecg.reshape(1, -1, 1), aug_ecg.reshape(1, -1, 1))
aug_snr = snr.result().numpy()
snr.reset_state()
snr.update_state(ecg.reshape(1, -1, 1), clean_ecg.reshape(1, -1, 1))
clean_snr = snr.result().numpy()

ecg = next(ds_gen) aug_ecg = augmenter(preprocessor(keras.ops.convert_to_tensor(np.reshape(ecg, (1, -1, 1)))), training=True).numpy().squeeze() clean_ecg = model.predict(np.reshape(aug_ecg, (1, -1, 1))) snr = nse.metrics.Snr() snr.update_state(ecg.reshape(1, -1, 1), aug_ecg.reshape(1, -1, 1)) aug_snr = snr.result().numpy() snr.reset_state() snr.update_state(ecg.reshape(1, -1, 1), clean_ecg.reshape(1, -1, 1)) clean_snr = snr.result().numpy()

1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 9ms/step

In [31]:

fig, ax = plt.subplots(3, 1, figsize=(9, 5), sharex=True)
ax[0].plot(ts, ecg.squeeze(), color=plot_theme.primary_color, lw=3)
ax[1].plot(ts, aug_ecg.squeeze(), color=plot_theme.secondary_color, lw=3)
ax[2].plot(ts, clean_ecg.squeeze(), color=plot_theme.tertiary_color, lw=3)

ax[0].set_ylabel("Reference")
ax[1].set_ylabel("Noisy")
ax[2].set_ylabel("Denoised")

ax[1].text(0.98, 0.15, f"{aug_snr:4.02f} dB SNR", transform=ax[1].transAxes, ha="right", va="top", weight='bold')
ax[2].text(0.98, 0.15, f"{clean_snr:4.02f} dB SNR", transform=ax[2].transAxes, ha="right", va="top", weight='bold')
# Disable y-axis ticks for all plots
for axes in ax:
    axes.yaxis.set_ticks([])
ax[-1].set_xlabel("Time (s)")
fig.suptitle("ECG Denoising Demo")
fig.tight_layout()
fig.show()

fig, ax = plt.subplots(3, 1, figsize=(9, 5), sharex=True) ax[0].plot(ts, ecg.squeeze(), color=plot_theme.primary_color, lw=3) ax[1].plot(ts, aug_ecg.squeeze(), color=plot_theme.secondary_color, lw=3) ax[2].plot(ts, clean_ecg.squeeze(), color=plot_theme.tertiary_color, lw=3) ax[0].set_ylabel("Reference") ax[1].set_ylabel("Noisy") ax[2].set_ylabel("Denoised") ax[1].text(0.98, 0.15, f"{aug_snr:4.02f} dB SNR", transform=ax[1].transAxes, ha="right", va="top", weight='bold') ax[2].text(0.98, 0.15, f"{clean_snr:4.02f} dB SNR", transform=ax[2].transAxes, ha="right", va="top", weight='bold') # Disable y-axis ticks for all plots for axes in ax: axes.yaxis.set_ticks([]) ax[-1].set_xlabel("Time (s)") fig.suptitle("ECG Denoising Demo") fig.tight_layout() fig.show()

In [ ]: