Create README.md

PULSAR Radar

PULSAR Radar is a lightweight neural network developed for radar/sensor inference with an emphasis on efficient deployment.

The repository provides models for standard Keras execution, TensorFlow Lite inference, INT8-quantized edge deployment, and embedded C/C++ integration.

Model Variants
File Description
model.keras Best Keras radar model
model.tflite TensorFlow Lite model
model_int8.tflite INT8-quantized TensorFlow Lite model
model.h Model representation for embedded/C integration
Optimization

Model hyperparameters were explored with Keras Tuner Hyperband.

The search space included:

First-stage filters: 16, 32, 64
Second-stage filters: 32, 64, 128
Dense units: 64, 128, 256
Dropout: 0.1 to 0.4
Learning rate: 1e-2, 1e-3, 1e-4

The tuning configuration used:

26 recorded trials
Maximum 15 epochs
Hyperband factor: 3
1 Hyperband iteration

The original tuning artifacts are available under pulsar_radar_hyperband in the source repository.

Usage
Install dependencies
pip install keras tensorflow huggingface_hub
Load the Keras model
from huggingface_hub import hf_hub_download
import keras

model_path = hf_hub_download(
repo_id="enesor0/pulsar-radar",
filename="model.keras"
)

model = keras.models.load_model(model_path)

model.summary()
TensorFlow Lite inference
from huggingface_hub import hf_hub_download
import tensorflow as tf

model_path = hf_hub_download(
repo_id="enesor0/pulsar-radar",
filename="model.tflite"
)

interpreter = tf.lite.Interpreter(model_path=model_path)
interpreter.allocate_tensors()

input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

print("Input:", input_details)
print("Output:", output_details)
INT8 deployment
from huggingface_hub import hf_hub_download
import tensorflow as tf

model_path = hf_hub_download(
repo_id="enesor0/pulsar-radar",
filename="model_int8.tflite"
)

interpreter = tf.lite.Interpreter(model_path=model_path)
interpreter.allocate_tensors()

The quantized variant is provided for environments where inference efficiency, memory footprint, and deployment constraints are important.

Embedded Deployment

The included model.h representation enables integration into embedded C/C++ workflows.

Potential deployment scenarios include:

Edge AI devices
Embedded radar processing
Sensor-processing systems
TinyML experimentation
Low-power inference environments

Compatibility depends on the target hardware and inference runtime.

Model Inputs and Outputs

The current public model artifacts do not include enough documentation to reliably specify:

Exact input tensor shape
Input normalization/preprocessing
Radar feature representation
Output class names
Prediction thresholds

These parameters should be taken from the original training/preprocessing pipeline before inference.

Evaluation

Formal independent test metrics are not currently published in this model card.

Hyperparameter-search validation results should not be presented as final production performance without evaluation on an independent test dataset.

Development Pipeline

The project demonstrates a model-development workflow covering:

Neural network development with Keras
Hyperparameter optimization using Hyperband
TensorFlow Lite conversion
INT8 post-training quantization
Embedded model export

Source repository: enesor0/pulsarmodels on GitHub

Intended Use

PULSAR Radar is intended for:

Machine-learning research
Radar/sensor ML experimentation
Edge inference
Embedded-AI prototyping
Model optimization experiments

The model should be independently tested and validated for the specific sensor configuration before production deployment.

Author

Enes Or

Machine Learning / Software Development

README.md

@@ -0,0 +1,16 @@
+---
+license: apache-2.0
+library_name: keras
+tags:
+- keras
+- tensorflow
+- tensorflow-lite
+- tflite
+- int8
+- quantization
+- edge-ai
+- embedded-ml
+- radar
+- sensor-data
+- hyperparameter-tuning
+---