Bring your own model (BYOM)

Upload your own model directly into your Edge Impulse project (TensorFlow SavedModel, ONNX, or TensorFlow Lite)

Bring your own model or BYOM allows you to optimize and deploy your own pretrained model (TensorFlow SavedModel, ONNX, or TensorFlow Lite) to any edge device, directly from your Edge Impulse project.

Getting Started

First, create a new project in Edge Impulse.

Also make sure you have your own pretrained model available locally on your computer, in one of the following formats: TensorFlow SavedModel (saved_model.zip), ONNX model (.onnx) or TensorFlow Lite model (.tflite)

For this guide, we will be uploading a pretrained image classification TFLite model for plant disease classification.

Then, from the Dashboard, of your Edge Impulse project under "Getting started", select Upload your model:

Step 1: Upload your model

Upload your trained model: Upload a TensorFlow SavedModel (saved_model.zip), ONNX model (.onnx) or TensorFlow Lite model (.tflite) to get started.
Model performance: Do you want performance characteristics (latency, RAM and ROM) for a specific device? Select "No" to show the performance for a range of device types, or "Yes" to run performance profiling for any of our available officially supported Edge Impulse development platforms.
If your model is not already quantized, you can also upload a .npy file to Upload representative features (Optional) - for example, your validation set - as an .npy file. This way, we can automatically quantize this model for better on-device performance. See quantization below.

After configuring the settings for uploading your model, select Upload your model and wait for your model to upload, you can check the upload status via the "Upload progress" section.

Quantization

You can optionally quantize a model during deployment. A quantized model will use an internal int8 numeric representation rather than float32, which can result in reduced memory usage and faster computation on many targets.

Quantization requires a sample of data that is representative of the range (maximum and minimum) of values in your training data. In the Studio, we require a numpy file (.npy). Each element of the array must have the same shape as your model's input.

Note that quantization is a form of lossy compression and may result in a reduction in model performance. It's important to evaluate your model after quantization to ensure it still performs well enough for your use case.

Step 2: Process your model

Depending on the model you have uploaded in Step 1, the configuration settings available for Step 2 will change.

For this guide, we have selected the following configuration model settings for optimal processing for an image classification model with input shape (300, 300, 3) in RGB format, Classification model output and 16 output labels: Tomato Healthy, Tomato Septoria Leaf Spot, Tomato Bacterial Spot, Tomato Blight, Cabbage Healthy, Tomato Spider Mite, Tomato Leaf Mold, Tomato_Yellow Leaf Curl Virus, Soy_Frogeye_Leaf_Spot, Soy_Downy_Mildew, Maize_Ravi_Corn_Rust, Maize_Healthy, Maize_Grey_Leaf_Spot, Maize_Lethal_Necrosis, Soy_Healthy, Cabbage Black Rot

After configuring your model settings, select Save model to view your model's on-device performance information for both MCUs and microprocessors (if applicable, depending on your model's arena size).

Check model behavior

Optionally upload test data to ensure correct model settings and proper model processing:

Limitations

There are a couple of restrictions to converting models with our tooling:

The model must have 1 input tensor.
You need to have a batch dimension (e.g. if your input is an image of 320x320x3 then your input shape should be (1,320,320,3)), and the batch size must be equal to 1. For ONNX models you can use a variable batch size (we'll set it to 1).
For classification and regression models: The model must have 1 output tensor, and the output tensor should have a batch dimension as well. E.g. a classification model with 3 classes should have output shape (1,3).
Regression models must have an output shape of (1,1).

Troubleshooting

--saved-model /tmp/saved_model does not exist:

If you encountered the following error:

Job started
Converting SavedModel...
Scheduling job in cluster...
Job started
Application exited with code 1
INFO: No representative features passed in, won't quantize this model

Extracting saved model...
Extracting saved model OK

--saved-model /tmp/saved_model does not exist

Converting SavedModel failed, see above

Job failed (see above)

Make sure to upload a .zip archive containing at minimum a saved_model directory that contains your saved_model.pb.

Could not profile: No uploaded model yet

If you encounter the following error:

Could not profile: No uploaded model yet

This often means that the model you are attempting to upload is unsupported. Only the following model formats are supported at this time:

TensorFlow SavedModel (in .zip archive)
ONNX (.onnx)
TensorFlow Lite (.tflite or .lite)

This model won’t run on MCUs.

If you see a message along these lines:

This model won’t run on MCUs. Unsupported ops: Pow.

Then that means that part of your network is not supported by EON Compiler, and cannot run as-is on MCUs. A list of all supported ops can be found in our SDK; but there might be certain input/output types that are not supported even if the op is in this list (you'll get a proper error code if this is the case). If you're an enterprise customer please contact your solutions engineer, most of the times we can add the missing op within a few days.

Clear model and revert back to impulse mode

When you upload a model via BYOM, your project is converted to BYOM mode where the Impulse design page is replaced by your model. To remove your model and return your project to the default impulse mode (so you can upload data and train a model from scratch), go to your project's Dashboard and click either Add existing data or Collect new data. You will be asked if you want to remove the existing model. Click Yes, switch.

Preprocessing your data using Edge Impulse DSP blocks

If you want to use our built-in processing blocks to preprocess your data, for example to turn your audio files into spectrograms before training your network, then you should:

Use the code in the edgeimpulse/processing-blocks repo to preprocess your data before training your model. This contains the Python version of the processing blocks.
Then, on the embedded side, call extract_XXXX_features to preprocess your sensor data, and pass the resulting features into the ei_run_classifier function.

Here's an end-to-end example of the embedded code to preprocess using an MFCC block:

#include <stdio.h>
#include "edge-impulse-sdk/classifier/ei_run_classifier.h"

// Audio (int16)
static const int16_t features[] = {
    0, 0, 1, 0, 0, 0, ...rest
};

static const float frequency = 16000.0f;

// We need the audio as float, but don't want to allocate the whole features buffer as float in one go
// as that'd double the required RAM. So dynamically convert i16->f32 when the MFCC algorithm requires.
static int input_signal_get_data(size_t offset, size_t length, float *out_ptr) {
    return numpy::int16_to_float(features + offset, out_ptr, length);
}

int main(int argc, char **argv) {
    // construct input signal (input is i16)
    signal_t input_signal;
    input_signal.total_length = sizeof(features) / sizeof(features[0]);
    input_signal.get_data = &input_signal_get_data;
    
    // construct an MFCC config (should match the params used to preprocess before training)
    ei_dsp_config_mfcc_t config = {
        1, // uint32_t block_id;
        3, // uint16_t implementation_version;
        1, // int axes;
        13, // int num_cepstral;
        0.032, // float frame_length;
        0.024, // float frame_stride;
        32, // int num_filters;
        256, // int fft_length;
        101, // int win_size;
        300, // int low_frequency;
        0, // int high_frequency;
        0.98f, // float pre_cof;
        1, // int pre_shift;
    };
    
    // calculate the size of the output matrix
    matrix_size_t out_matrix_size =
        speechpy::feature::calculate_mfcc_buffer_size(
            input_signal.total_length, frequency, config.frame_length, config.frame_stride, config.num_cepstral, config.implementation_version);
    printf("out_matrix_size = %u x %u\n", out_matrix_size.rows, out_matrix_size.cols);
    
    // create a new output matrix
    matrix_t out_matrix(out_matrix_size.rows, out_matrix_size.cols);
    int mfcc_ret = extract_mfcc_features(&input_signal, &out_matrix, (void *)&config, frequency);
    if (mfcc_ret != EIDSP_OK) {
        printf("ERR: MFCC failed w/ %d", mfcc_ret);
        return 1;
    }
    
    // print the features so you can check adherence to your Python results
    printf("Features:\n");
    for (size_t ix = 0; ix < out_matrix.rows * out_matrix.cols; ix++) {
        printf("%f ", out_matrix.buffer[ix]);
    }
    printf("\n");
    
    // this you'll pass into the NN
    signal_t nn_signal;
    ei::numpy::signal_from_buffer(out_matrix.buffer, out_matrix_size.rows * out_matrix_size.cols, &nn_signal);
    
    // call ei_run_classifier as usual

PreviousImpulse design NextProcessing blocks

Last updated 6 months ago

Job started Converting SavedModel... Scheduling job in cluster... Job started Application exited with code 1 INFO: No representative features passed in, won't quantize this model Extracting saved model... Extracting saved model OK --saved-model /tmp/saved_model does not exist Converting SavedModel failed, see above Job failed (see above)

#include <stdio.h> #include "edge-impulse-sdk/classifier/ei_run_classifier.h" // Audio (int16) static const int16_t features[] = { 0, 0, 1, 0, 0, 0, ...rest }; static const float frequency = 16000.0f; // We need the audio as float, but don't want to allocate the whole features buffer as float in one go // as that'd double the required RAM. So dynamically convert i16->f32 when the MFCC algorithm requires. static int input_signal_get_data(size_t offset, size_t length, float *out_ptr) { return numpy::int16_to_float(features + offset, out_ptr, length); } int main(int argc, char **argv) { // construct input signal (input is i16) signal_t input_signal; input_signal.total_length = sizeof(features) / sizeof(features[0]); input_signal.get_data = &input_signal_get_data; // construct an MFCC config (should match the params used to preprocess before training) ei_dsp_config_mfcc_t config = { 1, // uint32_t block_id; 3, // uint16_t implementation_version; 1, // int axes; 13, // int num_cepstral; 0.032, // float frame_length; 0.024, // float frame_stride; 32, // int num_filters; 256, // int fft_length; 101, // int win_size; 300, // int low_frequency; 0, // int high_frequency; 0.98f, // float pre_cof; 1, // int pre_shift; }; // calculate the size of the output matrix matrix_size_t out_matrix_size = speechpy::feature::calculate_mfcc_buffer_size( input_signal.total_length, frequency, config.frame_length, config.frame_stride, config.num_cepstral, config.implementation_version); printf("out_matrix_size = %u x %u\n", out_matrix_size.rows, out_matrix_size.cols); // create a new output matrix matrix_t out_matrix(out_matrix_size.rows, out_matrix_size.cols); int mfcc_ret = extract_mfcc_features(&input_signal, &out_matrix, (void *)&config, frequency); if (mfcc_ret != EIDSP_OK) { printf("ERR: MFCC failed w/ %d", mfcc_ret); return 1; } // print the features so you can check adherence to your Python results printf("Features:\n"); for (size_t ix = 0; ix < out_matrix.rows * out_matrix.cols; ix++) { printf("%f ", out_matrix.buffer[ix]); } printf("\n"); // this you'll pass into the NN signal_t nn_signal; ei::numpy::signal_from_buffer(out_matrix.buffer, out_matrix_size.rows * out_matrix_size.cols, &nn_signal); // call ei_run_classifier as usual