class Tester(object):
def __init__(self, log_dir):
init_seeds(0)
self.device = gpu.select_device()
self.log_dir = log_dir
self.yolov4 = Build_Model(weight_path=None, resume=False)
self.yolov4 = self.yolov4.to(self.device)
self.__load_best_weights()
def __load_best_weights(self):
best_weight = os.path.join(log_dir, "checkpoints", "best.pt")
chkpt = torch.load(best_weight, map_location=self.device)
self.yolov4.load_state_dict(chkpt)
del chkpt
def test(self):
logger.info(self.yolov4)
evaluator = COCOAPIEvaluator(cfg=cfg,
img_size=cfg.VAL.TEST_IMG_SIZE,
confthre=cfg.VAL.CONF_THRESH,
nmsthre=cfg.VAL.NMS_THRESH)
logger.info("\n=============== start testing ===============")
start = time.time()
coco_stat = evaluator.evaluate(self.yolov4)
end = time.time()
logger.info(
"Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = {:.04f}"
.format(coco_stat[0]))
logger.info(
"Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = {:.04f}"
.format(coco_stat[1]))
logger.info(
"Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = {:.04f}"
.format(coco_stat[2]))
logger.info(
"Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = {:.04f}"
.format(coco_stat[3]))
logger.info(
"Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = {:.04f}"
.format(coco_stat[4]))
logger.info(
"Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = {:.04f}"
.format(coco_stat[5]))
logger.info(
"Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = {:.04f}"
.format(coco_stat[6]))
logger.info(
"Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = {:.04f}"
.format(coco_stat[7]))
logger.info(
"Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = {:.04f}"
.format(coco_stat[8]))
logger.info(
"Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = {:.04f}"
.format(coco_stat[9]))
logger.info(
"Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = {:.04f}"
.format(coco_stat[10]))
logger.info(
"Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = {:.04f}"
.format(coco_stat[11]))
logger.info("cost time:{:.4f}s".format(end - start))
logger.info("FPS:{:.4f}".format(
len(evaluator.dataloader) / (end - start)))
class Trainer(object):
_resume = False
_fine_tune = False
def __init__(self, log_dir, resume=False, fine_tune=False):
init_seeds(0)
self._fine_tune = fine_tune
self._resume = resume
if self._fine_tune:
self.__prepare_fine_tune()
self.device = gpu.select_device()
self.start_epoch = 0
self.best_mAP = 0.
self.accumulate = cfg.TRAIN.ACCUMULATE
self.log_dir = log_dir
self.weight_path = "yolov4.weights"
self.multi_scale_train = cfg.TRAIN.MULTI_SCALE_TRAIN
if self.multi_scale_train:
print('Using multi scales training')
else:
print('train img size is {}'.format(cfg.TRAIN.TRAIN_IMG_SIZE))
self.train_dataset = data.Build_Train_Dataset(anno_file=cfg.TRAIN.ANNO_FILE, anno_file_type="train", img_size=cfg.TRAIN.TRAIN_IMG_SIZE)
self.epochs = cfg.TRAIN.YOLO_EPOCHS if cfg.MODEL.MODEL_TYPE == 'YOLOv4' else cfg.TRAIN.Mobilenet_YOLO_EPOCHS
self.train_dataloader = DataLoader(self.train_dataset,
batch_size=cfg.TRAIN.BATCH_SIZE//cfg.TRAIN.ACCUMULATE,
num_workers=cfg.TRAIN.NUMBER_WORKERS,
shuffle=True, pin_memory=True)
self.yolov4 = Build_Model(weight_path="yolov4.weights", resume=resume)
self.yolov4 = self.yolov4.to(self.device)
self.optimizer = optim.SGD(self.yolov4.parameters(), lr=cfg.TRAIN.LR_INIT,
momentum=cfg.TRAIN.MOMENTUM, weight_decay=cfg.TRAIN.WEIGHT_DECAY)
self.criterion = YoloV4Loss(anchors=cfg.MODEL.ANCHORS, strides=cfg.MODEL.STRIDES,
iou_threshold_loss=cfg.TRAIN.IOU_THRESHOLD_LOSS)
self.scheduler = cosine_lr_scheduler.CosineDecayLR(self.optimizer,
T_max=self.epochs*len(self.train_dataloader),
lr_init=cfg.TRAIN.LR_INIT,
lr_min=cfg.TRAIN.LR_END,
warmup=cfg.TRAIN.WARMUP_EPOCHS*len(self.train_dataloader))
if resume: self.__load_resume_weights()
if self._fine_tune: self.__load_best_weights()
def __prepare_fine_tune(self):
cfg.defrost()
cfg.TRAIN.ANNO_FILE = cfg.FINE_TUNE.ANNO_FILE
cfg.TRAIN.YOLO_EPOCHS = cfg.FINE_TUNE.YOLO_EPOCHS
cfg.TRAIN.LR_INIT = cfg.FINE_TUNE.LR_INIT
cfg.TRAIN.LR_END = cfg.FINE_TUNE.LR_END
cfg.TRAIN.WARMUP_EPOCHS = cfg.FINE_TUNE.WARMUP_EPOCHS
cfg.freeze()
def __load_best_weights(self):
best_weight = os.path.join(log_dir,"checkpoints", "best.pt")
last_weight = os.path.join(log_dir,"checkpoints", "last.pt")
shutil.copy2(best_weight,
best_weight.replace("best.pt","best_before_fine_tune.pt"))
shutil.copy2(last_weight,
last_weight.replace("last.pt","last_before_fine_tune.pt"))
last_chkpt = torch.load(last_weight, map_location=self.device)
best_chkpt = torch.load(best_weight, map_location=self.device)
self.yolov4.load_state_dict(best_chkpt)
self.best_mAP = 0
del last_chkpt, best_chkpt
def __load_resume_weights(self):
last_weight = os.path.join(log_dir,"checkpoints", "last.pt")
chkpt = torch.load(last_weight, map_location=self.device)
self.yolov4.load_state_dict(chkpt['model'])
self.start_epoch = chkpt['epoch'] + 1
if chkpt['optimizer'] is not None:
self.optimizer.load_state_dict(chkpt['optimizer'])
self.best_mAP = chkpt['best_mAP']
del chkpt
def __save_model_weights(self, epoch, mAP):
if mAP > self.best_mAP:
self.best_mAP = mAP
best_weight = os.path.join(log_dir,"checkpoints", "best.pt")
last_weight = os.path.join(log_dir,"checkpoints", "last.pt")
chkpt = {'epoch': epoch,
'best_mAP': self.best_mAP,
'model': self.yolov4.module.state_dict() if torch.cuda.device_count()>1 else self.yolov4.state_dict(),
'optimizer': self.optimizer.state_dict()}
torch.save(chkpt, last_weight)
if self.best_mAP == mAP:
torch.save(chkpt['model'], best_weight)
if self._fine_tune and epoch % 5 == 0:
torch.save(chkpt['model'], os.path.join(log_dir,"checkpoints", 'backup_fine_tune_epoch_{:02d}.pt'.format(epoch)))
del chkpt
def train(self):
global writer
logger.info("Training start,img size is: {:d},batchsize is: {:d}, subdivision: {:d}, worker number is {:d}".format(cfg.TRAIN.TRAIN_IMG_SIZE, cfg.TRAIN.BATCH_SIZE, cfg.TRAIN.ACCUMULATE, cfg.TRAIN.NUMBER_WORKERS))
logger.info(self.yolov4)
n_train = len(self.train_dataset)
n_step = n_train // (cfg.TRAIN.BATCH_SIZE//cfg.TRAIN.ACCUMULATE) + 1
logger.info("Train datasets number is : {}".format(n_train))
evaluator = COCOAPIEvaluator(cfg=cfg,
img_size=cfg.VAL.TEST_IMG_SIZE,
confthre=cfg.VAL.CONF_THRESH,
nmsthre=cfg.VAL.NMS_THRESH)
if torch.cuda.device_count() > 1: self.yolov4 = torch.nn.DataParallel(self.yolov4)
logger.info("\n=============== start training ===============")
for epoch in range(self.start_epoch, self.epochs):
start = time.time()
self.yolov4.train()
with tqdm(total=n_train, unit="imgs", desc=f'Epoch {epoch}/{self.epochs}', ncols=30) as pbar:
for i, (imgs, label_sbbox, label_mbbox, label_lbbox, sbboxes, mbboxes, lbboxes) in enumerate(self.train_dataloader):
imgs = imgs.to(self.device)
label_sbbox = label_sbbox.to(self.device)
label_mbbox = label_mbbox.to(self.device)
label_lbbox = label_lbbox.to(self.device)
sbboxes = sbboxes.to(self.device)
mbboxes = mbboxes.to(self.device)
lbboxes = lbboxes.to(self.device)
p, p_d = self.yolov4(imgs)
loss, loss_ciou, loss_conf, loss_cls = self.criterion(p, p_d, label_sbbox, label_mbbox,
label_lbbox, sbboxes, mbboxes, lbboxes)
loss.backward()
# Accumulate gradient for x batches before optimizing
if i % self.accumulate == 0:
self.scheduler.step(n_step*epoch + i)
self.optimizer.step()
self.optimizer.zero_grad()
# Print batch results
if i % (5*self.accumulate) == 0:
logger.info("{:3}: total_loss:{:.4f} | loss_ciou:{:.4f} | loss_conf:{:.4f} | loss_cls:{:.4f} | lr:{:.6f}".format(
self.train_dataset.img_size, loss, loss_ciou, loss_conf, loss_cls, self.optimizer.param_groups[0]['lr']
))
writer.add_scalar('train/loss_ciou', loss_ciou, n_step * epoch + i)
writer.add_scalar('train/loss_conf', loss_conf, n_step * epoch + i)
writer.add_scalar('train/loss_cls', loss_cls, n_step * epoch + i)
writer.add_scalar('train/train_loss', loss, n_step * epoch + i)
writer.add_scalar('train/lr', self.optimizer.param_groups[0]['lr'], n_step * epoch + i)
# multi-sclae training (320-608 pixels) every 10 batches
if self.multi_scale_train and (i+1) % (5*self.accumulate) == 0:
self.train_dataset.img_size = random.choice(range(10, 20)) * 32
pbar.update(imgs.shape[0])
mAP = 0.
# evaluator = COCOAPIEvaluator(cfg=cfg,
# img_size=cfg.VAL.TEST_IMG_SIZE,
# confthre=cfg.VAL.CONF_THRESH,
# nmsthre=cfg.VAL.NMS_THRESH)
coco_stat = evaluator.evaluate(self.yolov4)
logger.info("Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = {:.04f}".format(coco_stat[0]))
logger.info("Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = {:.04f}".format(coco_stat[1]))
logger.info("Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = {:.04f}".format(coco_stat[2]))
logger.info("Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = {:.04f}".format(coco_stat[3]))
logger.info("Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = {:.04f}".format(coco_stat[4]))
logger.info("Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = {:.04f}".format(coco_stat[5]))
logger.info("Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = {:.04f}".format(coco_stat[6]))
logger.info("Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = {:.04f}".format(coco_stat[7]))
logger.info("Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = {:.04f}".format(coco_stat[8]))
logger.info("Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = {:.04f}".format(coco_stat[9]))
logger.info("Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = {:.04f}".format(coco_stat[10]))
logger.info("Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = {:.04f}".format(coco_stat[11]))
writer.add_scalar('val/mAP_50_95', coco_stat[0], epoch)
writer.add_scalar('val/mAP_50', coco_stat[1], epoch)
writer.add_scalar('val/mAP_75', coco_stat[2], epoch)
writer.add_scalar('val/mAP_small', coco_stat[3], epoch)
writer.add_scalar('val/mAP_medium', coco_stat[4], epoch)
writer.add_scalar('val/mAP_large', coco_stat[5], epoch)
writer.add_scalar('val/mAR_max_1', coco_stat[6], epoch)
writer.add_scalar('val/mAR_max_10', coco_stat[7], epoch)
writer.add_scalar('val/mAR_max_100',coco_stat[8], epoch)
writer.add_scalar('val/mAR_small', coco_stat[9], epoch)
writer.add_scalar('val/mAR_medium', coco_stat[10], epoch)
writer.add_scalar('val/mAR_large', coco_stat[11], epoch)
self.__save_model_weights(epoch, coco_stat[0])
logger.info('save weights done')
end = time.time()
logger.info("cost time:{:.4f}s".format(end - start))
logger.info("=====Training Finished. best_test_mAP:{:.3f}%====".format(self.best_mAP))
class LAMR_Tester(object):
id_map = ["pedestrian"]
def __init__(self, log_dir, test_images):
init_seeds(0)
self.device = gpu.select_device()
self.log_dir = log_dir
self.yolov4 = Build_Model(weight_path=None, resume=False)
self.yolov4 = self.yolov4.to(self.device)
self.dataset = Naive_Test_Dataset(test_images)
self.dataloader = torch.utils.data.DataLoader(
self.dataset,
batch_size=cfg.VAL.BATCH_SIZE,
shuffle=False,
pin_memory=True,
num_workers=cfg.VAL.NUMBER_WORKERS)
self.__load_best_weights()
def __load_best_weights(self):
best_weight = os.path.join(log_dir, "checkpoints", "best.pt")
chkpt = torch.load(best_weight, map_location=self.device)
self.yolov4.load_state_dict(chkpt)
del chkpt
def test(self):
logger.info(self.yolov4)
self.yolov4.eval()
Tensor = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.FloatTensor
results_path = os.path.join("/data", "mock_detections",
os.path.basename(self.log_dir), "day",
"val")
if not os.path.exists(results_path):
os.makedirs(results_path)
for i, (img_path, img, info_img) in tqdm(enumerate(self.dataloader),
desc="Test to ECP... ",
unit="imgs",
total=len(self.dataloader)):
info_img = [float(info) for info in info_img]
data_dict = {
"tags": [],
"imageheight": int(info_img[0]),
"imagewidth": int(info_img[1]),
"children": [],
"identity": "frame"
}
city_name = os.path.basename(os.path.dirname(img_path[0]))
os.makedirs(os.path.join(results_path, city_name), exist_ok=True)
result_json_path = os.path.join(
results_path,
os.path.basename(img_path[0]).replace("png", "json"))
with torch.no_grad():
img = Variable(img.type(Tensor))
_, outputs = self.yolov4(img)
outputs = outputs.unsqueeze(0)
outputs = postprocess(outputs, len(cfg.DATASET.CLASSES), 0.01,
cfg.VAL.NMS_THRESH)
if outputs[0] is None:
with open(result_json_path, "w") as json_fh:
json.dump(data_dict, json_fh, indent=4)
continue
outputs = outputs[0].cpu().data
for output in outputs:
x1 = float(output[0])
y1 = float(output[1])
x2 = float(output[2])
y2 = float(output[3])
box = yolobox2label((y1, x1, y2, x2), info_img)
data_dict["children"].append({
"tags": [],
"children": [],
"identity":
self.id_map[int(output[6])],
"score":
float(output[4]) * float(output[5]),
"y0":
float(box[0]),
"x0":
float(box[1]),
"y1":
float(box[2]),
"x1":
float(box[3])
}) # ECP Formats
with open(result_json_path, "w") as json_fh:
json.dump(data_dict, json_fh, indent=4)