Welcome to tf2mplabh3!
This project is proudly developed and maintained by Microchip Technology Inc.
It enables you to convert TensorFlow models to C code, ready for seamless integration into your MPLAB® Harmony v3 embedded projects.
tf2mplabh3 is a command-line tool that automates the conversion of TensorFlow models to C code, making it easy to deploy machine learning models on embedded systems using MPLAB® Harmony v3.
- Features
- Installation
- Quick Start
- Usage
- Arguments
- Examples
- How to Use the Hardware Capabilities to Accelerate Inference
- Benchmarking
- License
- Acknowledgments
- Convert TensorFlow SavedModel to C code
- Easy CLI interface
- Verbosity control for logging
- Ready for integration with MPLAB Harmony v3
Use a Linux® based host system. Clone the repository and run the installation script:
git clone --recursive https://github.com/MicrochipTech/tf2mplabh3.git
cd tf2mplabh3
sudo ./install.shgit submodule update --init --recursiveActivate the virtual environment and run the script, by passing the TensorFlow model path as shown in the example below:
source .venv/bin/activate
python3 -m tf2mplabh3 -m examples/mobilenet-v2-tensorflow2-035-128-classification-v2python3 -m tf2mplabh3 [options]| Argument | Description | Default |
|---|---|---|
-m, --model |
Path to TensorFlow SavedModel directory | examples/mobilenet-v2-tensorflow2-035-128-classification-v2 |
-onnx, --onnx_model |
Path to output ONNX intermediate model file | examples/model.onnx |
-c_file, --c_model_file |
Path to output C model file | examples/model.c |
--tag |
SavedModel tag (e.g., serve) |
None |
--signature_def |
Signature def key (e.g., serving_default) |
None |
--onnx2c |
Path to the onnx2c executable | c_deps/onnx2c/build/onnx2c |
-v, --verbosity |
Verbosity level (0=quiet, 1=all logs) |
0 |
--overwrite |
Overwrite existing ONNX or C model files | Not used |
python3 -m tf2mplabh3python3 -m tf2mplabh3 -m path/to/model -v 1In order to ensure an optimized inference time, leverage the features of the MPLAB® XC-32 Compilers by activating the third level of compilation in your MPLAB® X project. Doing this ensures an extended use of the hardware capabilities of the device.
As shown in the example image below:
All benchmarks were performed on:
- Hardware: Microchip SAMA5D29 Curiosity Development Board
- CPU: 1x ARM® Cortex®-A5
- Clock Frequency: 498 MHz
- Compiler: XC32 v4.30
The model used for the benchmarking operations is a mobilenet-v2-tensorflow2-035-128-classification-v2 model.
The following table shows the inference time for the example model converted and run with different optimization levels.
| Optimization Level | Inference Time (ms) | Notes/Flags Used |
|---|---|---|
| None | 7536.600 | No optimization |
| -O1 | 1730.550 | Basic optimization |
| -O2 | 1368.060 | More optimization |
| -O3 | 1188.790 | Optimize for speed |
| -Os | 1382.300 | Optimize for size |
Note:
Inference time was measured as the average over 100 runs.
Results may vary depending on compiler version, memory configuration, and other system activity.
The following table summarizes the results of comparing the logits (raw model outputs) produced by the TensorFlow example model and the compiled model.c file, running on the target.
This comparison was performed to validate the integrity and robustness of the model conversion and deployment process.
All results below were obtained with the MPLAB Harmony v3 compiled at the -O3 optimization level.
| Metric | Value | Description |
|---|---|---|
| Number of images compared | 1000 | Total images used for validation |
| Mean Absolute Error (MAE) | 4.56 × 10⁻⁶ | Average absolute difference per logit between host and target |
| Mean Squared Error (MSE) | 3.58 × 10⁻¹¹ | Average squared difference per logit |
| Maximum Absolute Error | 5.53 × 10⁻⁵ | Largest absolute difference observed for any logit |
| Mean Cosine Similarity | 1.000000 | Average cosine similarity between host and target logit vectors (1.0 = perfect match) |
| Top-1 Agreement | 100.00% | Percentage of images where the predicted class (highest logit) matches |
| Top-5 Agreement | 100.00% | Percentage of images where the top 5 predicted classes match |
| 0-1 Loss (Argmax) | 0.000 | Fraction of images where the predicted class differ between the host and target |
Interpretation:
These results demonstrate that the converted model’s outputs are virtually identical to the initial example, with only negligible differences attributable to floating-point precision. Both Top-1 and Top-5 classification results are in perfect agreement, and the 0-1 loss confirms that there were no mismatches in predicted classes between the host and target. This validates the correctness and robustness of the deployment at the -O3 optimization level.