raj5517/imu-activity-classifier

IMU Activity Classifier — Pruning + INT8 Quantization

Compact 1D-CNN for human activity recognition from 6-axis IMU signals. Trained on UCI HAR dataset, compressed via magnitude pruning (78% sparsity)

• INT8 quantization for edge/microcontroller deployment.

Results

Model	Size	Accuracy	Latency
Baseline FP32	358 KB	93.99%	46.1 ms
Pruned FP16 TFLite	191 KB	92.40%	0.054 ms
Pruned INT8 TFLite	113 KB	92.43%	0.026 ms

• Size reduction: 68.4%
• Latency speedup: 1775x
• Accuracy drop: 1.56%

Classes

WALKING · WALKING_UPSTAIRS · WALKING_DOWNSTAIRS · SITTING · STANDING · LAYING

Compression Pipeline

1. Baseline 1D-CNN trained on UCI HAR (93.99% accuracy)
2. Magnitude pruning with PolynomialDecay → 78% sparsity
3. INT8 post-training quantization → 113KB TFLite model

Confusion Matrix

Compression Summary

Training Curves

Usage

import tensorflow as tf
import numpy as np

interpreter = tf.lite.Interpreter("imu_pruned_int8.tflite")
interpreter.allocate_tensors()
input_det  = interpreter.get_input_details()
output_det = interpreter.get_output_details()

# sample shape: (1, 128, 9) — float32
sample = np.random.randn(1, 128, 9).astype(np.float32)
interpreter.set_tensor(input_det[0]['index'], sample)
interpreter.invoke()
output = interpreter.get_tensor(output_det[0]['index'])

ACTIVITIES = ['WALKING','WALKING_UPSTAIRS','WALKING_DOWNSTAIRS',
              'SITTING','STANDING','LAYING']
print(ACTIVITIES[np.argmax(output)])

Links

• GitHub: https://github.com/RAj5517/imu_activity_classifier
• Dataset: UCI HAR (University of California Irvine)