dataset_train, dataset_test = get_dataset(args)
time_limits = 5 * 60 * 60 # 5 days
epochs = 12
# use coco pre-trained model for custom datasets
transfer = None if ('voc' in args.dataset_name) or (
'coco' in args.dataset_name) else 'coco'
detector = task.fit(dataset_train,
num_trials=30,
epochs=epochs,
net=ag.Categorical('darknet53', 'mobilenet1.0'),
meta_arch=args.meta_arch,
lr=ag.Categorical(1e-2, 5e-3, 1e-3, 5e-4, 1e-4, 5e-5),
transfer=transfer,
data_shape=ag.Categorical(320, 416),
nthreads_per_trial=16,
ngpus_per_trial=1,
batch_size=8,
lr_decay_epoch=ag.Categorical('80,90', '85,95'),
warmup_epochs=ag.Int(1, 10),
warmup_iters=ag.Int(250, 1000),
wd=ag.Categorical(1e-4, 5e-4, 2.5e-4),
syncbn=ag.Bool(),
label_smooth=ag.Bool(),
time_limits=time_limits,
dist_ip_addrs=[])
test_map = detector.evaluate(dataset_test)
print("mAP on test dataset: {}".format(test_map[1][1]))
4,
'data_shape': [640, 800],
'lr_decay_epoch':
ag.Categorical([24, 28], [35], [50, 55], [40], [45], [55],
[30, 35], [20]),
# 'warmup_epochs': 2,
'warmup_iters':
ag.Int(5, 500),
'wd':
ag.Categorical(5e-4, 2.5e-4, 1e-4),
'transfer':
transfer,
'label_smooth':
False,
'syncbn':
ag.Bool(),
'num_trials':
30,
'time_limits':
time_limits,
'dist_ip_addrs':
None
}
elif args.meta_arch == 'yolo3':
kwargs = {
'meta_arch': args.meta_arch,
'backbone': ag.Categorical('darknet53', 'mobilenet1.0'),
'lr': ag.Categorical(1e-2, 5e-3, 1e-3, 5e-4, 1e-4, 5e-5),
'batch_size': 32,
'epochs': epochs,
'nthreads_per_trial': 16,
'no_mixup_epochs':
20,
'reuse_pred_weights':
True,
'horovod':
False,
'grace_period':
None,
'auto_search':
True,
'seed':
223,
'wd':
ag.Categorical(1e-4, 5e-4, 2.5e-4),
'syncbn':
ag.Bool(),
'label_smooth':
False,
'time_limits':
time_limits,
'dist_ip_addrs': []
}
# Construct a object detection task based on the config.
task = ObjectDetection(search_args, FasterRCNNEstimator)
# Automatically fit a model.
estimator = task.fit()
# Evaluate the final model on test set.
test_map = estimator.evaluate()
dataset_train, dataset_test = get_dataset(args)
time_limits = 5 * 24 * 60 * 60 # 5 days
epochs = 20
# use coco pre-trained model for custom datasets
transfer = None if ('voc' in args.dataset_name) or ('coco' in args.dataset_name) else 'coco'
if args.meta_arch == 'yolo3':
kwargs = {'num_trials': 30, 'epochs': epochs,
'net': ag.Categorical('darknet53', 'mobilenet1.0'), 'meta_arch': args.meta_arch,
'lr': ag.Categorical(1e-2, 5e-3, 1e-3, 5e-4, 1e-4, 5e-5), 'transfer': transfer,
'data_shape': ag.Categorical(320, 416), 'nthreads_per_trial': 16,
'ngpus_per_trial': 8, 'batch_size': 64,
'lr_decay_epoch': ag.Categorical('80,90', '85,95'),
'warmup_epochs': ag.Int(1, 10), 'warmup_iters': ag.Int(250, 1000),
'wd': ag.Categorical(1e-4, 5e-4, 2.5e-4), 'syncbn': ag.Bool(),
'label_smooth': ag.Bool(), 'time_limits': time_limits, 'dist_ip_addrs': []}
elif args.meta_arch == 'faster_rcnn':
kwargs = {'num_trials': 30, 'epochs': ag.Categorical(30, 40, 50, 60),
'net': ag.Categorical('resnest101', 'resnest50'),
'meta_arch': args.meta_arch,
'lr': ag.Categorical(0.02, 0.01, 0.005, 0.002, 2e-4, 5e-4), 'transfer': transfer,
'data_shape': (640, 800), 'nthreads_per_trial': 16,
'ngpus_per_trial': 8, 'batch_size': 16,
'lr_decay_epoch': ag.Categorical('24,28', '35', '50,55', '40', '45', '55',
'30, 35', '20'),
'warmup_iters': ag.Int(5, 500),
'wd': ag.Categorical(1e-4, 5e-4, 2.5e-4), 'syncbn': True,
'label_smooth': False, 'time_limits': time_limits, 'dist_ip_addrs': []}
else:
raise NotImplementedError('%s is not implemented.', args.meta_arch)