train_ds,
epochs=cfg.train.epochs,
batch_size=cfg.train.batch_size,
verbose=cfg.train.verbose,
validation_data=None if not cfg.train.val else val_ds,
validation_steps=val_steps[cfg.data.dataset] //
strategy.num_replicas_in_sync,
steps_per_epoch=spe)
yolo.save_weights(weights_path='./models/{}/weights'.format(
cfg.model.name),
weights_type=cfg.train.save_weights_type)
pickle.dump(hist.history,
open('./models/{}/history.pkl'.format(cfg.model.name), 'wb'))
return yolo
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-c', '--cfg', default='default')
args = parser.parse_args()
cfg = CN(new_allowed=True)
cfg.merge_from_file(osp.join('configs', args.cfg + '.yaml'))
cfg.model.name = args.cfg
tpu, strategy = detect_hardware(tpu_name=None)
yolo = train(cfg, strategy)
predict(yolo, cfg, dataset=cfg.data.dataset, split='val')
import numpy as np
from lr_schedules import WarmupCosineDecay, WarmupPiecewise
import os.path as osp
from utils import detect_hardware
from dataset.dataloader import load_representative_tfds, visualize, prediction_tf_lite, prediction_examples
from dataset.coco import cn as cfg
import argparse
from lr_schedules import WarmupCosineDecay
import math
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-c', '--cfg', required=True)
args = parser.parse_args()
tpu, strategy = detect_hardware(None)
if tpu:
cfg.TRAIN.ACCELERATOR = args.tpu
else:
cfg.TRAIN.ACCELERATOR = 'GPU/CPU'
cfg.merge_from_file('configs/' + args.cfg)
cfg.MODEL.NAME = args.cfg.split('.yaml')[0]
if cfg.DATASET.OUTPUT_SHAPE[-1] == 17:
cfg.DATASET.KP_FLIP = cfg.DATASET.KP_FLIP[:17]
cfg.MODEL.TFRECORDS = 'data/tfrecords_foot/val'
train_ds = load_representative_tfds(cfg)
def representative_dataset():
for img in train_ds:
AP_large = cocoEval.stats[5]
return mAP, AP_50, AP_75, AP_small, AP_medium, AP_large # AP
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--tpu', default=None)
parser.add_argument('-c', '--cfg', required=True) # yaml
parser.add_argument('--det', type=int, default=-1)
parser.add_argument('--ckpt', default=None)
parser.add_argument('--split', default='val')
args = parser.parse_args()
from dataset.coco import cn as cfg
cfg.merge_from_file('configs/' + args.cfg)
cfg.MODEL.NAME = args.cfg.split('.')[0]
if cfg.DATASET.OUTPUT_SHAPE[-1] == 17:
cfg.DATASET.KP_FLIP = cfg.DATASET.KP_FLIP[:17]
if args.ckpt:
cfg.MODEL.NAME += '_{}'.format(args.ckpt)
if args.det >= 0:
cfg.VAL.DET = bool(args.det)
tpu, strategy = detect_hardware(args.tpu)
if args.split == 'val':
AP, _ = validate(strategy, cfg, split='val')
print('AP: {:.5f}'.format(AP))
else:
validate(strategy, cfg, split='test')