return_targets=False)
model.load_weights('models/yolov3-tiny.conv.11', only_imagenet=True)
model = model.to(device)
if train:
set_random_seed(12345)
model.unfreeze()
optimizer = optim.SGD(model.get_trainable_parameters(),
lr=model.lr,
momentum=model.momentum,
weight_decay=model.weight_decay,
nesterov=True)
scheduler = step_decay_scheduler(optimizer,
steps=model.steps,
scales=model.scales)
losses = model.fit(train_data=train_data,
val_data=val_data,
optimizer=optimizer,
scheduler=scheduler,
epochs=120,
checkpoint_frequency=120,
num_workers=8)
pickle.dump(losses, open('{}_losses.pkl'.format(model.name), 'wb'))
if predict:
set_random_seed(12345)
def main():
set_random_seed(12345)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
predict = True
train = True
joint = True
ResNet = ResNetBackbone(model=models.resnet101,
model_path='models/resnet101-5d3b4d8f.pth',
pretrained=True,
input_dims=2,
device=device)
faster_rcnn = FasterRCNN(name='FasterRCNN',
anchors=[[121, 108], [147, 163], [176, 93],
[185, 184], [254, 148], [261, 101],
[335, 185], [342, 123], [466, 106]],
use_global_ctx=False,
backbone=ResNet,
num_classes=3,
device=device).to(device)
train_data = SSDataset(root_dir='../../../Data/SS/',
classes='../../../Data/SS/ss.names',
dataset='train',
mu=[0., 0.],
sigma=[0.348, 0.348],
mode='stft_iq',
return_targets=True)
val_data = SSDataset(root_dir='../../../Data/SS/',
classes='../../../Data/SS/ss.names',
dataset='val',
mu=[0., 0.],
sigma=[0.348, 0.348],
mode='stft_iq',
return_targets=True)
test_data = SSDataset(root_dir='../../../Data/SS/',
classes='../../../Data/SS/ss.names',
dataset='test',
mu=[0.141],
sigma=[0.469],
mode='stft',
return_targets=False)
if train:
set_random_seed(12345)
faster_rcnn.fast_rcnn.mini_freeze()
plist = get_trainable_parameters(faster_rcnn)
optimizer = optim.SGD(plist,
lr=1e-3,
momentum=0.9,
weight_decay=5e-4,
nesterov=True)
scheduler = step_decay_scheduler(optimizer,
steps=[-1, 200, 10000, 15000],
scales=[.1, 10., 0.1, 0.1])
losses = faster_rcnn.joint_training(train_data=train_data,
val_data=val_data,
optimizer=optimizer,
scheduler=scheduler,
epochs=18,
shuffle=True,
checkpoint_frequency=18)
pickle.dump(losses, open('losses.pkl', 'wb'))
if predict:
set_random_seed(12345)
faster_rcnn = pickle.load(open('FasterRCNN18.pkl', 'rb'))
faster_rcnn.predict(dataset=test_data,
confidence_threshold=0.001,
overlap_threshold=0.5,
show=True,
export=False)