evaluate.py

#!/usr/bin/env python3
# Copyright 2022-2024,2026 NXP
# SPDX-License-Identifier: MIT
#
# Evaluate a Deepface emotion TFLite model on the FER2013 PrivateTest split.
#
# Usage:
#   python evaluate.py \
#       --model original_model/emotion_uint8_float32.tflite \
#       --fer2013_csv PATH/TO/fer2013/fer2013.csv

import argparse
import sys

import numpy as np
import tensorflow as tf

LABELS = ['angry', 'disgust', 'fear', 'happy', 'sad', 'surprise', 'neutral']

parser = argparse.ArgumentParser(
    description="Evaluate Deepface emotion TFLite model on FER2013 PrivateTest split."
)
parser.add_argument("-m", "--model",
                    default="original_model/emotion_uint8_float32.tflite",
                    type=str,
                    help="Path to the TFLite model file.")
parser.add_argument("--fer2013_csv", default="fer2013/fer2013.csv", type=str,
                    help="Path to the FER2013 CSV file.")
args = parser.parse_args()

interpreter = tf.lite.Interpreter(args.model)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
input_scale, input_zero_point = input_details[0]["quantization"]
print("Loaded model:", args.model)

print(f"Reading FER2013 dataset from {args.fer2013_csv} ...")
with open(args.fer2013_csv, 'r') as f:
    lines = f.readlines()

lines = [li.strip().split(',') for li in lines[1:]]
lines = [li for li in lines if li[2] == "PrivateTest"]
lines = [(int(li[0]), np.array(li[1].split(' ')).reshape((48, 48, 1)).astype(np.float32))
         for li in lines]

print(f"  {len(lines)} PrivateTest images found.")

n_correct = 0
total = len(lines)

for i, (label, pixels) in enumerate(lines):
    # Normalize to [0, 1] and add batch dimension.
    im = pixels[None, ...] / 255.0

    # Quantize to uint8 using model quantization params.
    im_q = (im / input_scale + input_zero_point).astype(np.uint8)

    interpreter.set_tensor(input_details[0]['index'], im_q)
    interpreter.invoke()
    out = interpreter.get_tensor(output_details[0]['index'])

    if int(out.argmax()) == label:
        n_correct += 1

    if (i + 1) % 500 == 0:
        print(f"  [{i + 1}/{total}]  accuracy so far: {n_correct / (i + 1) * 100:.2f}%")
        sys.stdout.flush()

accuracy = n_correct / total * 100 if total > 0 else 0.0
print(f"\nResults for {args.model}")
print(f"  Images evaluated : {total}")
print(f"  Accuracy         : {accuracy:.2f}%")