train.py

import os
import cv2
import dlib
import numpy as np
from imutils import face_utils
import tensorflow as tf
import pickle
import onnx
import onnxruntime as ort
from onnx_tf.backend import prepare


def area_of(left_top, right_bottom):
    """
    Compute the areas of rectangles given two corners.
    Args:
        left_top (N, 2): left top corner.
        right_bottom (N, 2): right bottom corner.
    Returns:
        area (N): return the area.
    """
    hw = np.clip(right_bottom - left_top, 0.0, None)
    return hw[..., 0] * hw[..., 1]


def iou_of(boxes0, boxes1, eps=1e-5):
    """
    Return intersection-over-union (Jaccard index) of boxes.
    Args:
        boxes0 (N, 4): ground truth boxes.
        boxes1 (N or 1, 4): predicted boxes.
        eps: a small number to avoid 0 as denominator.
    Returns:
        iou (N): IoU values.
    """
    overlap_left_top = np.maximum(boxes0[..., :2], boxes1[..., :2])
    overlap_right_bottom = np.minimum(boxes0[..., 2:], boxes1[..., 2:])

    overlap_area = area_of(overlap_left_top, overlap_right_bottom)
    area0 = area_of(boxes0[..., :2], boxes0[..., 2:])
    area1 = area_of(boxes1[..., :2], boxes1[..., 2:])
    return overlap_area / (area0 + area1 - overlap_area + eps)


def hard_nms(box_scores, iou_threshold, top_k=-1, candidate_size=200):
    """
    Perform hard non-maximum-supression to filter out boxes with iou greater
    than threshold
    Args:
        box_scores (N, 5): boxes in corner-form and probabilities.
        iou_threshold: intersection over union threshold.
        top_k: keep top_k results. If k <= 0, keep all the results.
        candidate_size: only consider the candidates with the highest scores.
    Returns:
        picked: a list of indexes of the kept boxes
    """
    scores = box_scores[:, -1]
    boxes = box_scores[:, :-1]
    picked = []
    indexes = np.argsort(scores)
    indexes = indexes[-candidate_size:]
    while len(indexes) > 0:
        current = indexes[-1]
        picked.append(current)
        if 0 < top_k == len(picked) or len(indexes) == 1:
            break
        current_box = boxes[current, :]
        indexes = indexes[:-1]
        rest_boxes = boxes[indexes, :]
        iou = iou_of(
            rest_boxes,
            np.expand_dims(current_box, axis=0),
        )
        indexes = indexes[iou <= iou_threshold]

    return box_scores[picked, :]


def predict(width, height, confidences, boxes, prob_threshold, iou_threshold=0.5, top_k=-1):
    """
    Select boxes that contain human faces
    Args:
        width: original image width
        height: original image height
        confidences (N, 2): confidence array
        boxes (N, 4): boxes array in corner-form
        iou_threshold: intersection over union threshold.
        top_k: keep top_k results. If k <= 0, keep all the results.
    Returns:
        boxes (k, 4): an array of boxes kept
        labels (k): an array of labels for each boxes kept
        probs (k): an array of probabilities for each boxes being in corresponding labels
    """
    boxes = boxes[0]
    confidences = confidences[0]
    picked_box_probs = []
    picked_labels = []
    for class_index in range(1, confidences.shape[1]):
        probs = confidences[:, class_index]
        mask = probs > prob_threshold
        probs = probs[mask]
        if probs.shape[0] == 0:
            continue
        subset_boxes = boxes[mask, :]
        box_probs = np.concatenate([subset_boxes, probs.reshape(-1, 1)], axis=1)
        box_probs = hard_nms(box_probs,
                             iou_threshold=iou_threshold,
                             top_k=top_k,
                             )
        picked_box_probs.append(box_probs)
        picked_labels.extend([class_index] * box_probs.shape[0])
    if not picked_box_probs:
        return np.array([]), np.array([]), np.array([])
    picked_box_probs = np.concatenate(picked_box_probs)
    picked_box_probs[:, 0] *= width
    picked_box_probs[:, 1] *= height
    picked_box_probs[:, 2] *= width
    picked_box_probs[:, 3] *= height
    return picked_box_probs[:, :4].astype(np.int32), np.array(picked_labels), picked_box_probs[:, 4]


onnx_path = 'models/ultra_light/ultra_light_models/ultra_light_640.onnx'
onnx_model = onnx.load(onnx_path)
predictor = prepare(onnx_model)
ort_session = ort.InferenceSession(onnx_path)
input_name = ort_session.get_inputs()[0].name

shape_predictor = dlib.shape_predictor('models/facial_landmarks/shape_predictor_5_face_landmarks.dat')
fa = face_utils.facealigner.FaceAligner(shape_predictor, desiredFaceWidth=112, desiredLeftEye=(0.3, 0.3))

TRAINING_BASE = 'faces/training/'

dirs = os.listdir(TRAINING_BASE)
images = []
names = []

for label in dirs:
    for i, fn in enumerate(os.listdir(os.path.join(TRAINING_BASE, label))):
        print(f"start collecting faces from {label}'s data")
        cap = cv2.VideoCapture(os.path.join(TRAINING_BASE, label, fn))
        frame_count = 0
        while True:
            # read video frame
            ret, raw_img = cap.read()
            # process every 5 frames
            if frame_count % 5 == 0 and raw_img is not None:
                h, w, _ = raw_img.shape
                img = cv2.cvtColor(raw_img, cv2.COLOR_BGR2RGB)
                img = cv2.resize(img, (640, 480))
                img_mean = np.array([127, 127, 127])
                img = (img - img_mean) / 128
                img = np.transpose(img, [2, 0, 1])
                img = np.expand_dims(img, axis=0)
                img = img.astype(np.float32)

                confidences, boxes = ort_session.run(None, {input_name: img})
                boxes, labels, probs = predict(w, h, confidences, boxes, 0.7)

                # if face detected
                if boxes.shape[0] > 0:
                    x1, y1, x2, y2 = boxes[0, :]
                    gray = cv2.cvtColor(raw_img, cv2.COLOR_BGR2GRAY)
                    aligned_face = fa.align(raw_img, gray, dlib.rectangle(left=x1, top=y1, right=x2, bottom=y2))
                    aligned_face = cv2.resize(aligned_face, (112, 112))

                    cv2.imwrite(f'faces/tmp/{label}_{frame_count}.jpg', aligned_face)

                    aligned_face = aligned_face - 127.5
                    aligned_face = aligned_face * 0.0078125
                    images.append(aligned_face)
                    names.append(label)

            frame_count += 1
            if frame_count == cap.get(cv2.CAP_PROP_FRAME_COUNT):
                break

with tf.Graph().as_default():
    with tf.Session() as sess:
        print("loading checkpoint ...")
        saver = tf.train.import_meta_graph('models/mfn/m1/mfn.ckpt.meta')
        saver.restore(sess, 'models/mfn/m1/mfn.ckpt')

        images_placeholder = tf.get_default_graph().get_tensor_by_name("input:0")
        embeddings = tf.get_default_graph().get_tensor_by_name("embeddings:0")
        phase_train_placeholder = tf.get_default_graph().get_tensor_by_name("phase_train:0")

        feed_dict = {images_placeholder: images, phase_train_placeholder: False}
        embeds = sess.run(embeddings, feed_dict=feed_dict)
        with open("embeddings/embeddings.pkl", "wb") as f:
            pickle.dump((embeds, names), f)
        print("Done!")