| |
| |
| |
|
|
| import argparse |
| import os |
| import numpy as np |
| from glob import glob |
| import tqdm |
|
|
| try: |
| import tflite_runtime.interpreter as tflite |
| except ImportError: |
| import tensorflow as tf |
| tflite = tf.lite |
|
|
| import cv2 |
|
|
| THRESHOLD = 0.6 |
|
|
| |
| LFW_DIR = "lfw-deepfunneled" |
|
|
| |
| LFW_PAIRS_FILE = "pairsDevTest.txt" |
|
|
|
|
| def cosine_similarity(a, b): |
| return 1 - np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b)) |
|
|
|
|
| def load_img(filename): |
| img = cv2.imread(filename, 1) |
| img = img[45:-45, 45:-45] |
| img = np.array(img, dtype=np.uint8) |
| return img[None, ...] |
|
|
|
|
| def main(): |
| parser = argparse.ArgumentParser(description="Evaluate FaceNet512 on LFW pairsDevTest") |
| parser.add_argument('-m', '--model', |
| default='original_model/facenet512_uint8_float32.tflite', |
| help='Path to the TFLite model file') |
| args = parser.parse_args() |
|
|
| interpreter = tflite.Interpreter(model_path=args.model) |
| interpreter.allocate_tensors() |
| input_details = interpreter.get_input_details() |
| output_details = interpreter.get_output_details() |
|
|
| with open(LFW_PAIRS_FILE, 'r') as f: |
| pairs = f.readlines()[1:] |
| pairs = [p.strip().split("\t") for p in pairs] |
|
|
| image_filenames = set() |
| for line in pairs[:500]: |
| image_filenames.add(os.sep.join([LFW_DIR, line[0], |
| f"{line[0]}_{int(line[1]):04d}.jpg"])) |
| image_filenames.add(os.sep.join([LFW_DIR, line[0], |
| f"{line[0]}_{int(line[2]):04d}.jpg"])) |
| for line in pairs[500:]: |
| image_filenames.add(os.sep.join([LFW_DIR, line[0], |
| f"{line[0]}_{int(line[1]):04d}.jpg"])) |
| image_filenames.add(os.sep.join([LFW_DIR, line[2], |
| f"{line[2]}_{int(line[3]):04d}.jpg"])) |
|
|
| feature_vectors = dict() |
| for f in tqdm.tqdm(image_filenames, desc="Running inferences"): |
| img = load_img(f) |
| interpreter.set_tensor(input_details[0]['index'], img) |
| interpreter.invoke() |
| out = interpreter.get_tensor(output_details[0]['index']) |
| |
| if output_details[0]['dtype'] == np.uint8: |
| scale, zero_point = output_details[0]['quantization'] |
| out = (out.astype(np.float32) - zero_point) * scale |
| feature_vectors[f] = out[0] |
|
|
| n_correct = 0 |
|
|
| |
| for line in pairs[:500]: |
| f1 = os.sep.join([LFW_DIR, line[0], f"{line[0]}_{int(line[1]):04d}.jpg"]) |
| f2 = os.sep.join([LFW_DIR, line[0], f"{line[0]}_{int(line[2]):04d}.jpg"]) |
| result = cosine_similarity(feature_vectors[f1], feature_vectors[f2]) |
| n_correct += 1 if result < THRESHOLD else 0 |
|
|
| |
| for line in pairs[500:]: |
| f1 = os.sep.join([LFW_DIR, line[0], f"{line[0]}_{int(line[1]):04d}.jpg"]) |
| f2 = os.sep.join([LFW_DIR, line[2], f"{line[2]}_{int(line[3]):04d}.jpg"]) |
| result = cosine_similarity(feature_vectors[f1], feature_vectors[f2]) |
| n_correct += 1 if result > THRESHOLD else 0 |
|
|
| print(f"Quantized model accuracy: {n_correct / 1000.0:.1%}") |
|
|
|
|
| if __name__ == '__main__': |
| main() |
|
|
