Using the Edge Impulse Python SDK to run EON Tuner

The EON Tuner is Edge Impulse's automated machine learning (AutoML) tool to help you find the best combination of blocks and hyperparameters for your model and within your hardware constraints. This example will walk you through uploading data, running the EON Tuner, and interpreting the results.

WARNING: This notebook will add and delete data in your Edge Impulse project, so be careful! We recommend creating a throwaway project when testing this notebook.

To start, create a new project in Edge Impulse. Do not add any data to it.

# If you have not done so already, install the following dependencies
# !python -m pip install matplotlib pandas edgeimpulse

import edgeimpulse as ei
from edgeimpulse.experimental.data import (
    upload_directory
)
from edgeimpulse.experimental.tuner import (
    check_tuner,
    set_impulse_from_trial,
    start_tuner,
    start_custom_tuner,
    tuner_report_as_df,
)
from edgeimpulse.experimental.impulse import (
    build,
)

You will need to obtain an API key from an Edge Impulse project. Log into edgeimpulse.com and create a new project. Open the project, navigate to Dashboard and click on the Keys tab to view your API keys. Double-click on the API key to highlight it, right-click, and select Copy.

Copy API key from Edge Impulse project

Note that you do not actually need to use the project in the Edge Impulse Studio. We just need the API Key.

Paste that API key string in the ei.API_KEY value in the following cell:

# Settings
ei.API_KEY = "ei_dae2..." # Change this to your Edge Impulse API key
deploy_filename = "my_model_cpp.zip"

# Get the project ID
api = ei.experimental.api.EdgeImpulseApi()
project_id = api.default_project_id()

Upload dataset

We start by downloading the continuous motion dataset and uploading it to our project.

# Download and unzip gesture dataset
!mkdir -p dataset/
!wget -P dataset -q https://cdn.edgeimpulse.com/datasets/gestures.zip
!unzip -q dataset/gestures.zip -d dataset/gestures/

# Upload training dataset
resp = upload_directory(
    directory="dataset/gestures/training",
    category="training",
)
print(f"Uploaded {len(resp.successes)} training samples")

# Upload test dataset
resp = upload_directory(
    directory="dataset/gestures/testing",
    category="testing",
)
print(f"Uploaded {len(resp.successes)} testing samples")

# Uncomment the following if you want to delete the temporary dataset folder
#!rm -rf dataset/

Run the Tuner

To start, we need to list the possible target devices we can use for profiling. We need to pick from this list.

# List the available profile targets
ei.model.list_profile_devices()

You should see a list printed such as:

['alif-he',
 'alif-hp',
 'arduino-nano-33-ble',
 'arduino-nicla-vision',
 'portenta-h7',
 'brainchip-akd1000',
 'cortex-m4f-80mhz',
 'cortex-m7-216mhz',
 ...
 'ti-tda4vm']

From there, we start the tuner with start_tuner() and wait for completion via check_tuner(). In this example, we configure the tuner to target for the cortex-m4f-80mhz device. Since we want to classify the motion, we choose classification for our classifcation_type and our dataset as motion continuous. We constrain our model to a latency of 100ms for running the impulse.

NOTE: We set the max trials to 3 here. In a real life situation, you will omit this so the tuner decides the best number of trials.

Once the tuner is done, you can print out the results to determine the best combination of blocks and hyperparameters.

# Choose a device from the list
target_device = "cortex-m4f-80mhz"

# Start tuner. This will take 15+ minutes.
start_tuner(
    target_device=target_device,
    classification_type="classification",
    dataset_category="motion_continuous",
    target_latency=100,
    tuning_max_trials=3,
)

# Wait while checking the tuner's progress.
state = check_tuner(
    wait_for_completion=True
)

Print EON Tuner results

To visualize the results of the tuner trials, you can head to the project page on Edge Impulse Studio.

Alternatively, you can access the results programmatically: the configuration settings and output of the EON Tuner is stored in the variable state. You can access the results of the various trials with state.trials. Note that some trials can fail, so it's a good idea to test the status of each trial.

From there, you will want to sort the results based on some metric. In this example, we will sort based on int8 test set accuracy from highest to lowest.

Note: Edge Impulse supports only one learning block per project at this time (excluding anomaly detection blocks). As a result, we will use the first learning block (e.g. learning_blocks[0]) in the list to extract metrics.

import json

# The easiest way to view the results is to look at the EON Tuner page on your project
print(f"Navigate to https://studio.edgeimpulse.com/studio/{project_id}/tuner to see the results")

# Set quantization ("float32" or "int8")
qtzn = "int8"

# Filter out all failed trials
results = [r for r in state.trials if r.status == "completed"]

# Extract int8 accuracies from the trial results
accuracies = []
for result in results:
    accuracy = result.impulse.learn_blocks[0]["metrics"]["test"][qtzn]["accuracy"]
    accuracies.append(accuracy)

# Sort the results based on int8 accuracies
acc_results = zip(accuracies, results)
sorted_results = sorted(acc_results, reverse=True, key=lambda x: list(x)[0])
sorted_results = [result for _, result in sorted_results]

Now that we have the sorted results, we can extract the values we care about. We will print out the following metrics along with the impulse configuration (processing/learning block configuration and hyperparameters) of the top-performing trial.

This will help you determine if the impulse can fit on your target hardware and run fast enough for your needs. The impulse configuration can be used to recreate the processing and learning blocks on Edge Impulse. Later, we will set the project impulse based on the trial ID to simply deploy (rather than re-train).

Note: we assume the first learning block has the metrics we care about.

def get_metrics(results, qtzn, idx):
    """Calculate metrics for a given trial index"""

    metrics = {}

    # Get model accuracy results
    result_metrics = results[idx].impulse.learn_blocks[0]["metrics"]
    metrics["val-acc"] = result_metrics['validation'][qtzn]['accuracy']
    metrics["test-acc"] = result_metrics['test'][qtzn]['accuracy']
    
    # Calculate processing block RAM
    metrics["processing-block-ram"] = 0
    for i, dsp_block in enumerate(results[idx].impulse.dsp_blocks):
        metrics["processing-block-ram"] += dsp_block["performance"]["ram"]

    # Get latency, RAM, and ROM usage
    device_performance = results[idx].device_performance[qtzn]
    metrics["learning-block-latency-ms"] = device_performance['latency']
    metrics["learning-block-tflite-ram"] = device_performance['tflite']['ramRequired']
    metrics["learning-block-tflite-rom"] = device_performance['tflite']['romRequired']
    metrics["learning-block-eon-ram"] = device_performance['eon']['ramRequired']
    metrics["learning-block-eon-rom"] = device_performance['eon']['romRequired']

    return metrics

# The top performing impulse is the first element (sorted by highest int8 accuracy on test set)
trial_idx = 0

# Print info about the processing (DSP) blocks and store RAM usage
print("Processing blocks")
print("===")
for i, dsp_block in enumerate(sorted_results[trial_idx].impulse.dsp_blocks):
    print(f"Processing block {i}")
    print("---")
    print("Block:")
    print(json.dumps(dsp_block["block"], indent=2))
    print("Config:")
    print(json.dumps(dsp_block["config"], indent=2))
print()

# Print info about the learning blocks
print("Learning blocks")
print("===")
for i, learn_block in enumerate(sorted_results[trial_idx].impulse.learn_blocks):
    print(f"Learn block {i}")
    print("---")
    print("Block:")
    print(json.dumps(learn_block["block"], indent=2))
    print("Config:")
    print(json.dumps(learn_block["config"], indent=2))
    metadata = learn_block["metadata"]
    qtzn_metadata = [m for m in metadata["modelValidationMetrics"] if m.get("type") == qtzn]
print()

# Print metrics
metrics = get_metrics(sorted_results, qtzn, trial_idx)
print(f"Metrics ({qtzn}) for best trial")
print("===")
print(f"Validation accuracy: {metrics['val-acc']}")
print(f"Test accuracy: {metrics['test-acc']}")
print(f"Estimated processing blocks RAM (bytes): {metrics['processing-block-ram']}")
print(f"Estimated learning blocks latency (ms): {metrics['learning-block-latency-ms']}")
print(f"Estimated learning blocks RAM (bytes): {metrics['learning-block-tflite-ram']}")
print(f"Estimated learning blocks ROM (bytes): {metrics['learning-block-tflite-rom']}")
print(f"Estimated learning blocks RAM with EON Compiler (bytes): {metrics['learning-block-eon-ram']}")
print(f"Estimated learning blocks ROM with EON Compiler (bytes): {metrics['learning-block-eon-rom']}")

Graph results

You can optionally use a plotting package like matplotlib to graph the results from the top results to compare the metrics.

import matplotlib.pyplot as plt

# Get metrics for the top 3 trials (sorted by int8 test set accuracy)
num_trials = 3
top_metrics = [get_metrics(sorted_results, qtzn, idx) for idx in range(num_trials)]

# Construct metrics for plotting
test_accs = [top_metrics[x]['test-acc'] for x in range(num_trials)]
proc_rams = [top_metrics[x]['processing-block-ram'] for x in range(num_trials)]
learn_latencies = [top_metrics[x]['learning-block-latency-ms'] for x in range(num_trials)]
learn_tflite_rams = [top_metrics[x]['learning-block-tflite-ram'] for x in range(num_trials)]
learn_tflite_roms = [top_metrics[x]['learning-block-tflite-rom'] for x in range(num_trials)]
learn_eon_rams = [top_metrics[x]['learning-block-eon-ram'] for x in range(num_trials)]
learn_eon_roms = [top_metrics[x]['learning-block-eon-rom'] for x in range(num_trials)]

# Create plots
fig, axs = plt.subplots(7, 1, figsize=(8, 15))
indices = range(num_trials)

# Plot test accuracies
axs[0].barh(indices, test_accs)
axs[0].set_title("Test set accuracy")
axs[0].set_xlabel("Accuracy")
axs[0].set_ylabel("Trial")

# Plot processing block RAM
axs[1].barh(indices, proc_rams)
axs[1].set_title("Processing block RAM")
axs[1].set_xlabel("RAM (bytes)")
axs[1].set_ylabel("Trial")

# Plot learning block latency
axs[2].barh(indices, learn_latencies)
axs[2].set_title("Learning block latency")
axs[2].set_xlabel("Latency (ms)")
axs[2].set_ylabel("Trial")

# Plot learning block RAM (TFLite)
axs[3].barh(indices, learn_tflite_rams)
axs[3].set_title("Learning block RAM (TFLite)")
axs[3].set_xlabel("RAM (bytes)")
axs[3].set_ylabel("Trial")

# Plot learning block ROM (TFLite)
axs[4].barh(indices, learn_tflite_roms)
axs[4].set_title("Learning block ROM (TFLite)")
axs[4].set_xlabel("ROM (bytes)")
axs[4].set_ylabel("Trial")

# Plot learning block RAM (EON)
axs[5].barh(indices, learn_eon_rams)
axs[5].set_title("Learning block RAM (EON)")
axs[5].set_xlabel("RAM (bytes)")
axs[5].set_ylabel("Trial")

# Plot learning block ROM (EON)
axs[6].barh(indices, learn_eon_roms)
axs[6].set_title("Learning block ROM (EON)")
axs[6].set_xlabel("ROM (bytes)")
axs[6].set_ylabel("Trial")

# Prevent overlap
plt.tight_layout()

Results as a DataFrame

If you have pandas installed, you can make the previous section much easier by reporting metrics as a DataFrame.

import pandas as pd

# Convert the state metrics into a DataFrame
df = tuner_report_as_df(state)
df.head()

# Print column names
for col in df.columns:
    print(col)

# Sort the DataFrame by validation (int8) accuracy
df = df.sort_values(by="test_int8_accuracy", ascending=False)

# Print out best trial metrics
print(f"Trial ID: {df.iloc[0]['id']}")
print(f"Test accuracy (int8): {df.iloc[0]['test_int8_accuracy']}")
print(f"Estimated learning blocks latency (ms): {df.iloc[0]['device_performance_int8_latency']}")
print(f"Estimated learning blocks RAM (bytes): {df.iloc[0]['device_performance_int8_tflite_ram_required']}")
print(f"Estimated learning blocks ROM (bytes): {df.iloc[0]['device_performance_int8_tflite_rom_required']}")
print(f"Estimated learning blocks RAM with EON Compiler (bytes): {df.iloc[0]['device_performance_int8_eon_ram_required']}")
print(f"Estimated learning blocks ROM with EON Compiler (bytes): {df.iloc[0]['device_performance_int8_eon_rom_required']}")

Set trial as impulse and deploy

We can replace the current impulse with the top performing trial from the EON Tuner. From there, we can deploy it, just like we would any impulse.

# Get the ID for the top-performing trial and set that to our impulse. This will take about a minute.
trial_id = df.iloc[0].trial_id
response = set_impulse_from_trial(trial_id)
job_id = response.id

# Make sure the impulse update was successful
if not hasattr(response, "success") or getattr(response, "success") == False:
    raise RuntimeError("Could not set project impulse to trial impulse")

# List the available profile target devices
ei.model.list_deployment_targets()

You should see a list printed such as:

['zip',
 'arduino',
 'cubemx',
 'wasm',
 ...
 'runner-linux-aarch64-jetson-orin-6-0']

The most generic target is to download a .zip file that holds a C++ library containing the inference runtime and your trained model, so we choose 'zip' from the above list. To do that, we first need to create a Classification object which contains our label strings (and other optional information about the model). These strings will be added to the C++ library metadata so you can access them in your edge application.

Note that instead of writing the raw bytes to a file, you can also specify an output_directory argument in the .deploy() function. Your deployment file(s) will be downloaded to that directory.

Important! The deployment targets list will change depending on the values provided for model, model_output_type, and model_input_type in the next part. For example, you will not see openmv listed once you upload a model (e.g. using .profile() or .deploy()) if model_input_type is not set to ei.model.input_type.ImageInput(). If you attempt to deploy to an unavailable target, you will receive the error Could not deploy: deploy_target: .... If model_input_type is not provided, it will default to OtherInput. See this page for more information about input types.

# Build and download C++ library with the trained model
deploy_bytes = None
try:
    deploy_bytes = build(
        deploy_model_type=qtzn,
        engine="tflite",
        deploy_target="zip"
    )
except Exception as e:
    print(f"Could not deploy: {e}")
    
# Write the downloaded raw bytes to a file
if deploy_bytes:
    with open(deploy_filename, 'wb') as f:
        f.write(deploy_bytes.getvalue())

Your model C++ library should be downloaded as the file my_model_cpp.zip in the same directory as this notebook. You are now ready to use your C++ model in your embedded and edge device application! To use the C++ model for local inference, see our documentation here.

Configure custom search space

By default, the EON Tuner will make a guess at a search space based on the type of data you uploaded (e.g. using spectral-analysis blocks for feature extraction). As a result, you can run the tuner without needing to construct a search space. However, you may want to define your own search space.

The best way to define a search space is to open your project (after uploading data), head to the EON Tuner page, click Run EON Tuner, and select the Space tab.

EON Tuner search space

The search space is defined in JSON format, so we can just copy that to create a dictionary. This is a good place to start for tuning blocks and hyperparameters.

Note: Functions to get available blocks and search space parameters coming soon

from edgeimpulse_api import (
    OptimizeConfig,
    TunerSpaceImpulse,
)

# Configure the search space
space = {
    "inputBlocks": [
      {
        "type": "time-series",
        "window": [
          {"windowSizeMs": 9000, "windowIncreaseMs": 9000},
          {"windowSizeMs": 10000, "windowIncreaseMs": 10000}
        ],
        "frequencyHz": [62.5],
        "padZeros": [True]
      }
    ],
    "dspBlocks": [
      {
        "type": "spectral-analysis",
        "analysis-type": ["FFT"],
        "fft-length": [16, 64],
        "scale-axes": [1],
        "filter-type": ["none"],
        "filter-cutoff": [3],
        "filter-order": [6],
        "do-log": [True],
        "do-fft-overlap": [True]
      },
      {
        "type": "spectral-analysis",
        "analysis-type": ["Wavelet"],
        "wavelet": ["haar", "bior1.3"],
        "wavelet-level": [1, 2]
      },
      {"type": "raw", "scale-axes": [1]}
    ],
    "learnBlocks": [
      {
        "id": 4,
        "type": "keras",
        "dimension": ["dense"],
        "denseBaseNeurons": [40, 20],
        "denseLayers": [2, 3],
        "dropout": [0.25, 0.5],
        "learningRate": [0.0005],
        "trainingCycles": [30]
      }
    ]
  }

# Wrap the search space
ts = TunerSpaceImpulse.from_dict(space)

# Create a custom configuration
config = OptimizeConfig(
    name=None,
    target_device={"name": "cortex-m4f-80mhz"},
    classification_type="classification",
    dataset_category="motion_continuous",
    target_latency=100,
    tuning_max_trials=2,
    space=[ts]
)

# Start tuner and wait for it to complete
start_custom_tuner(
    config=config
)
state = check_tuner(
    wait_for_completion=True
)

# The easiest way to view the results is to look at the EON Tuner page on your project
print(f"Navigate to https://studio.edgeimpulse.com/studio/{project_id}/tuner to see the results")

# Set quantization ("float32" or "int8")
qtzn = "int8"

# Filter out all failed trials
results = [r for r in state.trials if r.status == "completed"]

# Extract float32 accuracies from the trial results
accuracies = []
for result in results:
    accuracy = result.impulse.learn_blocks[0]["metrics"]["test"][qtzn]["accuracy"]
    accuracies.append(accuracy)

# Sort the results based on int8 accuracies
acc_results = zip(accuracies, results)
sorted_results = sorted(acc_results, reverse=True, key=lambda x: list(x)[0])
sorted_results = [result for _, result in sorted_results]

# The top performing impulse is the first element (sorted by highest int8 accuracy on test set)
trial_idx = 0

# Print info about the processing (DSP) blocks and store RAM usage
print("Processing blocks")
print("===")
for i, dsp_block in enumerate(sorted_results[trial_idx].impulse.dsp_blocks):
    print(f"Processing block {i}")
    print("---")
    print("Block:")
    print(json.dumps(dsp_block["block"], indent=2))
    print("Config:")
    print(json.dumps(dsp_block["config"], indent=2))
print()

# Print info about the learning blocks
print("Learning blocks")
print("===")
for i, learn_block in enumerate(sorted_results[trial_idx].impulse.learn_blocks):
    print(f"Learn block {i}")
    print("---")
    print("Block:")
    print(json.dumps(learn_block["block"], indent=2))
    print("Config:")
    print(json.dumps(learn_block["config"], indent=2))
    metadata = learn_block["metadata"]
    qtzn_metadata = [m for m in metadata["modelValidationMetrics"] if m.get("type") == qtzn]
print()

# Print metrics
metrics = get_metrics(sorted_results, qtzn, trial_idx)
print(f"Metrics ({qtzn}) for best trial")
print("===")
print(f"Validation accuracy: {metrics['val-acc']}")
print(f"Test accuracy: {metrics['test-acc']}")
print(f"Estimated processing blocks RAM (bytes): {metrics['processing-block-ram']}")
print(f"Estimated learning blocks latency (ms): {metrics['learning-block-latency-ms']}")
print(f"Estimated learning blocks RAM (bytes): {metrics['learning-block-tflite-ram']}")
print(f"Estimated learning blocks ROM (bytes): {metrics['learning-block-tflite-rom']}")
print(f"Estimated learning blocks RAM with EON Compiler (bytes): {metrics['learning-block-eon-ram']}")
print(f"Estimated learning blocks ROM with EON Compiler (bytes): {metrics['learning-block-eon-rom']}")