了解了yolov3模型和損失函數(shù)的計算后,我們來分析yolov3訓(xùn)練和測試數(shù)據(jù)集的制作代碼客税。
這里有必要說明一下况褪,這里的數(shù)據(jù)集就是我們在訓(xùn)練yolov3神經(jīng)網(wǎng)絡(luò)時所需要的圖片資源和相應(yīng)的真實框標(biāo)記。
數(shù)據(jù)集的制作流程如下:
- 從文件中讀取圖片數(shù)據(jù)和標(biāo)記的真實框數(shù)據(jù)(2個左上角數(shù)據(jù)+2個右下角數(shù)據(jù))更耻、真實框的類別(用一個整數(shù)表示)
- 生成3個尺度的特征圖测垛,找出每個特征圖中真實框位置處對應(yīng)的3個先驗框,然后這3個先驗框和真實框計算iou值秧均,然后根據(jù)iou值大于指定閾值的條件食侮,找出符合條件的先驗框,最后把真實框數(shù)據(jù)填充到先驗框在特征圖中對應(yīng)的位置處目胡。還有锯七,把類別通過one-hot編碼,也填充到先驗框在特征圖中對應(yīng)的位置處誉己。
- 生成3個尺度的真實框張量眉尸。在上一步中,我們可以找到每個尺度的特征圖中和真實框匹配的先驗框巨双,這樣噪猾,我們就可以把真實框保存進(jìn)這個尺度的真實框張量中。
- 數(shù)據(jù)基本已經(jīng)制作完成了筑累,返回3個尺度的特征圖和3個尺度的真實框張量袱蜡。
數(shù)據(jù)集的制作流程基本就是這樣的,下面我們看具體的代碼實現(xiàn)慢宗。
class Dataset(object):
"""implement Dataset here"""
def __init__(self, dataset_type):
self.annot_path = cfg.TRAIN.ANNOT_PATH if dataset_type == 'train' else cfg.TEST.ANNOT_PATH
self.input_sizes = cfg.TRAIN.INPUT_SIZE if dataset_type == 'train' else cfg.TEST.INPUT_SIZE
self.batch_size = cfg.TRAIN.BATCH_SIZE if dataset_type == 'train' else cfg.TEST.BATCH_SIZE
self.data_aug = cfg.TRAIN.DATA_AUG if dataset_type == 'train' else cfg.TEST.DATA_AUG
self.train_input_sizes = cfg.TRAIN.INPUT_SIZE#訓(xùn)練集圖片尺寸
self.strides = np.array(cfg.YOLO.STRIDES)#每個特征圖中的一個格子代表原始圖像中的幾個格子
self.classes = utils.read_class_names(cfg.YOLO.CLASSES)#類別的索引
self.num_classes = len(self.classes)#類別的個數(shù)
self.anchors = np.array(utils.get_anchors(cfg.YOLO.ANCHORS))#3個尺度的3個不同大小的先驗框坪蚁,一共有9個
self.anchor_per_scale = cfg.YOLO.ANCHOR_PER_SCALE#一個尺度上有幾個先驗框,這里是3個
self.max_bbox_per_scale = 150#一個尺度上的最大先驗框個數(shù)
self.annotations = self.load_annotations(dataset_type)#加載訓(xùn)練集標(biāo)簽
self.num_samples = len(self.annotations)#樣本數(shù)量
self.num_batchs = int(np.ceil(self.num_samples / self.batch_size))#一共有幾個 batch
self.batch_count = 0 # 計數(shù)
load_annotations
def load_annotations(self, dataset_type):
with open(self.annot_path, 'r') as f:
txt = f.readlines()
annotations = [line.strip() for line in txt if len(line.strip().split()[1:]) != 0]
np.random.shuffle(annotations)
return annotations
def __iter__(self):
return self
def __next__(self):
with tf.device('/cpu:0'):
self.train_input_size = random.choice(self.train_input_sizes)#輸入的圖片尺寸大小
self.train_output_sizes = self.train_input_size // self.strides#計算3個尺度的輸出大小
# 初始化一個批次的樣本
batch_image = np.zeros((self.batch_size, self.train_input_size, self.train_input_size, 3), dtype=np.float32)
# 初始化一個批次的輸出(批次大小為4婆廊,三個尺度上的輸出尺寸分別為52迅细、26巫橄、13淘邻,一個尺度上共有3個先驗框,2個坐標(biāo)中心點+2個高寬信息+1個置信度信息+80個類別信息)
batch_label_sbbox = np.zeros((self.batch_size, self.train_output_sizes[0], self.train_output_sizes[0],
self.anchor_per_scale, 5 + self.num_classes), dtype=np.float32)#初始化小尺度的輸出湘换,shape = [4宾舅,52统阿,52,3筹我,85]
batch_label_mbbox = np.zeros((self.batch_size, self.train_output_sizes[1], self.train_output_sizes[1],
self.anchor_per_scale, 5 + self.num_classes), dtype=np.float32)#初始化中尺度的輸出扶平,shape = [4, 26, 26, 3, 85]
batch_label_lbbox = np.zeros((self.batch_size, self.train_output_sizes[2], self.train_output_sizes[2],
self.anchor_per_scale, 5 + self.num_classes), dtype=np.float32)# 初始化大尺度的輸出,shape = [4蔬蕊,13结澄,13,3岸夯,85]
batch_sbboxes = np.zeros((self.batch_size, self.max_bbox_per_scale, 4), dtype=np.float32)#小尺度的真實框張量
batch_mbboxes = np.zeros((self.batch_size, self.max_bbox_per_scale, 4), dtype=np.float32)#中尺度的真實框張量
batch_lbboxes = np.zeros((self.batch_size, self.max_bbox_per_scale, 4), dtype=np.float32)#大尺度的真實框張量
num = 0#記錄遍歷到一個批次(4張圖片)中的第幾張圖片了
if self.batch_count < self.num_batchs:#如果記錄的 batch 個數(shù)還沒達(dá)到總 batch 個數(shù)
while num < self.batch_size:#如果這個批次(4張圖片)還沒遍歷完
index = self.batch_count * self.batch_size + num#記錄現(xiàn)在是第幾個樣本
if index >= self.num_samples: index -= self.num_samples#如果這個樣本已經(jīng)超過了總個圖片數(shù)量麻献,那么從頭再開始
annotation = self.annotations[index]#根據(jù)索引取出這個圖片的信息(包括圖片,真實框信息猜扮,類別信息)
image, bboxes = self.parse_annotation(annotation)#解析出圖片信息和真實框信息
label_sbbox, label_mbbox, label_lbbox, sbboxes, mbboxes, lbboxes = self.preprocess_true_boxes(bboxes)#返回保存有真實框信息的3個尺度的特征圖和3個真實框張量信息
batch_image[num, :, :, :] = image
batch_label_sbbox[num, :, :, :, :] = label_sbbox
batch_label_mbbox[num, :, :, :, :] = label_mbbox
batch_label_lbbox[num, :, :, :, :] = label_lbbox
batch_sbboxes[num, :, :] = sbboxes
batch_mbboxes[num, :, :] = mbboxes
batch_lbboxes[num, :, :] = lbboxes
num += 1
self.batch_count += 1
batch_smaller_target = batch_label_sbbox, batch_sbboxes
batch_medium_target = batch_label_mbbox, batch_mbboxes
batch_larger_target = batch_label_lbbox, batch_lbboxes
return batch_image, (batch_smaller_target, batch_medium_target, batch_larger_target)
else:
self.batch_count = 0
np.random.shuffle(self.annotations)
raise StopIteration
圖片的一系列增強(qiáng)操作
# 圖片增強(qiáng)操作之-水平翻轉(zhuǎn)
def random_horizontal_flip(self, image, bboxes):
if random.random() < 0.5:
_, w, _ = image.shape
image = image[:, ::-1, :]
bboxes[:, [0,2]] = w - bboxes[:, [2,0]]
return image, bboxes
# 圖片增強(qiáng)操作之-隨機(jī)裁剪
def random_crop(self, image, bboxes):
if random.random() < 0.5:
h, w, _ = image.shape
max_bbox = np.concatenate([np.min(bboxes[:, 0:2], axis=0), np.max(bboxes[:, 2:4], axis=0)], axis=-1)
max_l_trans = max_bbox[0]
max_u_trans = max_bbox[1]
max_r_trans = w - max_bbox[2]
max_d_trans = h - max_bbox[3]
crop_xmin = max(0, int(max_bbox[0] - random.uniform(0, max_l_trans)))
crop_ymin = max(0, int(max_bbox[1] - random.uniform(0, max_u_trans)))
crop_xmax = max(w, int(max_bbox[2] + random.uniform(0, max_r_trans)))
crop_ymax = max(h, int(max_bbox[3] + random.uniform(0, max_d_trans)))
image = image[crop_ymin : crop_ymax, crop_xmin : crop_xmax]
bboxes[:, [0, 2]] = bboxes[:, [0, 2]] - crop_xmin
bboxes[:, [1, 3]] = bboxes[:, [1, 3]] - crop_ymin
return image, bboxes
# 圖片增強(qiáng)操作之-隨機(jī)移動
def random_translate(self, image, bboxes):
if random.random() < 0.5:
h, w, _ = image.shape
max_bbox = np.concatenate([np.min(bboxes[:, 0:2], axis=0), np.max(bboxes[:, 2:4], axis=0)], axis=-1)
max_l_trans = max_bbox[0]
max_u_trans = max_bbox[1]
max_r_trans = w - max_bbox[2]
max_d_trans = h - max_bbox[3]
tx = random.uniform(-(max_l_trans - 1), (max_r_trans - 1))
ty = random.uniform(-(max_u_trans - 1), (max_d_trans - 1))
M = np.array([[1, 0, tx], [0, 1, ty]])
image = cv2.warpAffine(image, M, (w, h))
bboxes[:, [0, 2]] = bboxes[:, [0, 2]] + tx
bboxes[:, [1, 3]] = bboxes[:, [1, 3]] + ty
return image, bboxes
parse_annotation
#解析出圖片信息和真實框信息
def parse_annotation(self, annotation):
line = annotation.split()
image_path = line[0]
if not os.path.exists(image_path):
raise KeyError("%s does not exist ... " %image_path)
image = cv2.imread(image_path)#讀取圖片信息
bboxes = np.array([list(map(int, box.split(','))) for box in line[1:]])#將真實框數(shù)據(jù)由字符串型轉(zhuǎn)換為整型
# 一系列的數(shù)據(jù)增強(qiáng)操作
if self.data_aug:
image, bboxes = self.random_horizontal_flip(np.copy(image), np.copy(bboxes))
image, bboxes = self.random_crop(np.copy(image), np.copy(bboxes))
image, bboxes = self.random_translate(np.copy(image), np.copy(bboxes))
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image, bboxes = utils.image_preporcess(np.copy(image), [self.train_input_size, self.train_input_size], np.copy(bboxes))#將圖片尺寸處理成神經(jīng)網(wǎng)絡(luò)需要的圖片尺寸勉吻,相應(yīng)的真實框的數(shù)據(jù)信息也要進(jìn)行處理(主要是數(shù)據(jù)偏移處理)
return image, bboxes
bbox_iou
# 計算兩個框框的iou值,處理方法與yolov3.py文件中的bbox_iou方法大致相同
def bbox_iou(self, boxes1, boxes2):
boxes1 = np.array(boxes1)
boxes2 = np.array(boxes2)
boxes1_area = boxes1[..., 2] * boxes1[..., 3]
boxes2_area = boxes2[..., 2] * boxes2[..., 3]
boxes1 = np.concatenate([boxes1[..., :2] - boxes1[..., 2:] * 0.5,
boxes1[..., :2] + boxes1[..., 2:] * 0.5], axis=-1)
boxes2 = np.concatenate([boxes2[..., :2] - boxes2[..., 2:] * 0.5,
boxes2[..., :2] + boxes2[..., 2:] * 0.5], axis=-1)
left_up = np.maximum(boxes1[..., :2], boxes2[..., :2])
right_down = np.minimum(boxes1[..., 2:], boxes2[..., 2:])
inter_section = np.maximum(right_down - left_up, 0.0)
inter_area = inter_section[..., 0] * inter_section[..., 1]
union_area = boxes1_area + boxes2_area - inter_area
return inter_area / union_area
preprocess_true_boxes
def preprocess_true_boxes(self, bboxes):
# 3個尺度的特征圖信息
label = [np.zeros((self.train_output_sizes[i], self.train_output_sizes[i], self.anchor_per_scale,
5 + self.num_classes)) for i in range(3)]
# 3個尺度的真實框張量
bboxes_xywh = [np.zeros((self.max_bbox_per_scale, 4)) for _ in range(3)]
bbox_count = np.zeros((3,))
for bbox in bboxes:#遍歷真實框數(shù)組
bbox_coor = bbox[:4]#取出每個真實框的坐標(biāo)信息(2個左上角坐標(biāo)+2個右下角坐標(biāo))
bbox_class_ind = bbox[4]#每個真實框的類別
#開始對類別進(jìn)行one-hot編碼
onehot = np.zeros(self.num_classes, dtype=np.float)
onehot[bbox_class_ind] = 1.0
uniform_distribution = np.full(self.num_classes, 1.0 / self.num_classes)
deta = 0.01
#最終處理完成的類別one-hot編碼
smooth_onehot = onehot * (1 - deta) + deta * uniform_distribution
#將真實框的坐標(biāo)從(2個左上角坐標(biāo)+2個右下角坐標(biāo))轉(zhuǎn)換為(2個中心點坐標(biāo)+2個高寬)
bbox_xywh = np.concatenate([(bbox_coor[2:] + bbox_coor[:2]) * 0.5, bbox_coor[2:] - bbox_coor[:2]], axis=-1)
#將真實框的坐標(biāo)轉(zhuǎn)換為特征圖上的坐標(biāo)旅赢,bbox_xywh的shape=[1,4]齿桃,self.strides的shape=[3,1],于是bbox_xywh_scaled的shape=[1,3,4]煮盼,即計算出了每個真實框在3個尺度下的坐標(biāo)信息
bbox_xywh_scaled = 1.0 * bbox_xywh[np.newaxis, :] / self.strides[:, np.newaxis]
iou = []#保存真實框與9個先驗框計算的iou值
exist_positive = False#真實框是否有相匹配的先驗框的標(biāo)志
for i in range(3):#遍歷3個尺度
#每個尺度下有3個先驗框
anchors_xywh = np.zeros((self.anchor_per_scale, 4))
#3個先驗框的中心坐標(biāo)短纵,讓它在真實框所在的格子的中心坐標(biāo)上。
anchors_xywh[:, 0:2] = np.floor(bbox_xywh_scaled[i, 0:2]).astype(np.int32) + 0.5
#3個先驗框的尺寸大小
anchors_xywh[:, 2:4] = self.anchors[i]
#計算真實框與這3個先驗框的iou值
iou_scale = self.bbox_iou(bbox_xywh_scaled[i][np.newaxis, :], anchors_xywh)
#保存iou值
iou.append(iou_scale)
#找出符合條件的iou值
iou_mask = iou_scale > 0.3
if np.any(iou_mask):
#到這一步孕似,說明有符合條件的先驗框踩娘,于是找出真實框所在的格子坐標(biāo),即第幾行幾列
xind, yind = np.floor(bbox_xywh_scaled[i, 0:2]).astype(np.int32)
#首先清除下該尺度下對應(yīng)格子的信息
label[i][yind, xind, iou_mask, :] = 0
#在該尺度的格子中填充進(jìn)真實框坐標(biāo)信息
label[i][yind, xind, iou_mask, 0:4] = bbox_xywh
#該尺度的格子中的置信度置為1喉祭,表明該格子中有檢測到物體
label[i][yind, xind, iou_mask, 4:5] = 1.0
#填充該尺度的格子中的類別信息养渴,就是前面已經(jīng)做好的類別one-hot編碼
label[i][yind, xind, iou_mask, 5:] = smooth_onehot
#找出該尺度下真實框張量中的索引
bbox_ind = int(bbox_count[i] % self.max_bbox_per_scale)
#在真實框張量的索引位置處填充進(jìn)真實框信息
bboxes_xywh[i][bbox_ind, :4] = bbox_xywh
#該索引下的值+1
bbox_count[i] += 1
#表明有符合真實框要求的先驗框
exist_positive = True
if not exist_positive:
#如果沒有符合真實框要求的先驗框,那就在9個iou值中找出最大的那個值的索引
best_anchor_ind = np.argmax(np.array(iou).reshape(-1), axis=-1)
#計算這個最大的先驗框所在的尺度
best_detect = int(best_anchor_ind / self.anchor_per_scale)
#計算在該尺度下的哪個先驗框
best_anchor = int(best_anchor_ind % self.anchor_per_scale)
#同樣泛烙,找出在該尺度下的格子坐標(biāo)理卑,即第幾行第幾列
xind, yind = np.floor(bbox_xywh_scaled[best_detect, 0:2]).astype(np.int32)
#和上面的處理過程一樣,首先清除該尺度下的對應(yīng)格子信息
label[best_detect][yind, xind, best_anchor, :] = 0
#在該格子中填充真實框信息
label[best_detect][yind, xind, best_anchor, 0:4] = bbox_xywh
#把該格子的置信度置為1
label[best_detect][yind, xind, best_anchor, 4:5] = 1.0
#填充該格子的類別one-hot
label[best_detect][yind, xind, best_anchor, 5:] = smooth_onehot
#這也和上面一樣蔽氨,找出該尺度下的真實框張量的索引
bbox_ind = int(bbox_count[best_detect] % self.max_bbox_per_scale)
#把真實框信息填充到真實框張量的索引位置處
bboxes_xywh[best_detect][bbox_ind, :4] = bbox_xywh
#該尺度下的索引值+1
bbox_count[best_detect] += 1
#3個尺度的特征信息
label_sbbox, label_mbbox, label_lbbox = label
#3個尺度的真實框張量信息
sbboxes, mbboxes, lbboxes = bboxes_xywh
return label_sbbox, label_mbbox, label_lbbox, sbboxes, mbboxes, lbboxes
len
#返回數(shù)據(jù)集的批次大小
def __len__(self):
return self.num_batchs
這就是yolov3數(shù)據(jù)集的制作流程和代碼藐唠,在下一篇中我們來看yolov3是如何實現(xiàn)訓(xùn)練神經(jīng)網(wǎng)絡(luò)。