def test_Resnet50Encoder():
net = Resnet50Encoder().eval()
print(count_parameters(net))
x = torch.rand((4, 3, 512, 512))
y = net(x)
for yi in y:
print(yi.size())
def test_retinanet34_shared():
model = RRetinaNetShared(encoder=Resnet34Encoder(),
num_classes=1,
image_size=(768, 768))
model.set_encoder_training_enabled(False)
model.eval()
model.cuda()
print(count_parameters(model))
boxes = np.array([[40, 50, 100, 20, 15], [200, 200, 16, 4, -33]],
dtype=np.float32)
labels = np.array([0, 0], dtype=int)
true_loc, true_cls = model.box_coder.encode(boxes, labels)
image = torch.randn(1, 3, model.box_coder.image_height,
model.box_coder.image_width)
pred_bboxes, pred_labels = model(image.cuda())
print(pred_bboxes.size())
print(pred_labels.size())
cls_loss = SSDBinaryLabelLoss()
loc_loss = RSSDLocationLoss()
# Make tensor of proper shape & batch_size
true_loc = torch.unsqueeze(torch.from_numpy(true_loc), 0)
true_cls = torch.unsqueeze(torch.from_numpy(true_cls).long(), 0)
floss = cls_loss(pred_bboxes, pred_labels, true_loc.cuda(),
true_cls.cuda())
lloss = loc_loss(pred_bboxes, pred_labels, true_loc.cuda(),
true_cls.cuda())
print('Random input', floss, lloss)
def test_se_resnext_unet():
net = UNetModel(num_classes=1,
encoder=SEResNeXt50Encoder,
decoder_block=UnetDecoderBlockSE,
central_block=UnetCentralBlock)
net.eval()
net.set_encoder_training_enabled(False)
print(count_parameters(net))
x = torch.rand((4, 3, 512, 512))
mask = net(x)
print(mask.size())
print(net)
def test():
model = FPNSSD512(1)
image = torch.randn(4, 3, 512, 512)
output = model(image)
print(count_parameters(model))
model.cuda()
image = image.cuda()
with torch.cuda.profiler.profile() as prof:
model(image) # Warmup CUDA memory allocator and profiler
with torch.autograd.profiler.emit_nvtx():
model(image)
print(prof)
features = []
x = image
for mod in self.encoders:
x = mod(x)
features.append(x)
return features
def forward(self, image):
features = self.get_encoder_features(image)
x = self.center(F.max_pool2d(features[-1], kernel_size=2, stride=2))
for encoder_features, decoder_block in zip(reversed(features),
self.decoder):
x = decoder_block(x, encoder_features)
x = F.interpolate(x,
scale_factor=4,
mode='bilinear',
align_corners=True)
x = self.final(x)
return x
if __name__ == '__main__':
net = SenetUnet(num_classes=2).eval()
print(count_parameters(net))
mask = net(torch.rand((4, 3, 256, 256)))
print(mask.size())
def resnext50(**kwargs):
return ResNeXt([3, 4, 6, 3], **kwargs)
def resnext101(pretrained=True, input_3x3=True, abn_block=ABN, **kwargs):
model = ResNeXt([3, 4, 23, 3],
input_3x3=input_3x3,
abn_block=abn_block,
classes=1000,
**kwargs)
if pretrained and input_3x3:
checkpoint = torch.load(
os.path.join('pretrain', 'resnext101_ipabn_lr_512.pth.tar'))
state_dict = checkpoint['state_dict']
new_state_dict = OrderedDict()
for k, v in state_dict.items():
name = k[7:] # remove `module.`
new_state_dict[name] = v
model.load_state_dict(new_state_dict)
return model
def resnext152(**kwargs):
return ResNeXt([3, 8, 36, 3], **kwargs)
if __name__ == '__main__':
print(count_parameters(resnext50()))
print(count_parameters(resnext101()))
print(count_parameters(resnext152()))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--seed', type=int, default=42, help='Random seed')
parser.add_argument('-dd',
'--data-dir',
type=str,
default='d:\\datasets\\airbus',
help='Data dir')
parser.add_argument('-m',
'--model',
type=str,
default='rretina_net',
help='')
parser.add_argument('-b',
'--batch-size',
type=int,
default=4,
help='Batch Size during training, e.g. -b 64')
parser.add_argument('-e',
'--epochs',
type=int,
default=150,
help='Epoch to run')
parser.add_argument('-es',
'--early-stopping',
type=int,
default=None,
help='Maximum number of epochs without improvement')
parser.add_argument('-f',
'--fold',
default=0,
type=int,
help='Fold to train')
parser.add_argument('-fe',
'--freeze-encoder',
type=int,
default=0,
help='Freeze encoder parameters for N epochs')
parser.add_argument('-ft', '--fine-tune', action='store_true')
parser.add_argument('--fast', action='store_true')
parser.add_argument('-lr',
'--learning-rate',
type=float,
default=1e-3,
help='Initial learning rate')
parser.add_argument('-lrs',
'--lr-scheduler',
default=None,
help='LR scheduler')
parser.add_argument('-o',
'--optimizer',
default='Adam',
help='Name of the optimizer')
parser.add_argument('-r',
'--resume',
type=str,
default=None,
help='Checkpoint filename to resume')
parser.add_argument('-w',
'--workers',
default=4,
type=int,
help='Num workers')
parser.add_argument('-wd',
'--weight-decay',
type=float,
default=0,
help='L2 weight decay')
parser.add_argument('-p', '--patch-size', type=int, default=768, help='')
parser.add_argument('-ew', '--encoder-weights', default=None, type=str)
args = parser.parse_args()
set_manual_seed(args.seed)
train_session_args = vars(args)
train_session = get_random_name()
current_time = datetime.now().strftime('%b%d_%H_%M')
prefix = f'{current_time}_{args.model}_f{args.fold}_{train_session}_{args.patch_size}'
if args.fast:
prefix += '_fast'
print(prefix)
print(args)
log_dir = os.path.join('runs', prefix)
exp_dir = os.path.join('experiments', args.model, prefix)
os.makedirs(exp_dir, exist_ok=True)
model = get_model(args.model,
num_classes=1,
image_size=(args.patch_size, args.patch_size))
print(count_parameters(model))
train_loader, valid_loader = get_dataloaders(
args.data_dir,
box_coder=model.box_coder,
fold=args.fold,
patch_size=args.patch_size,
train_batch_size=args.batch_size,
valid_batch_size=args.batch_size,
fast=args.fast)
# Declare variables we will use during training
start_epoch = 0
train_history = pd.DataFrame()
best_metric_val = 0
best_lb_checkpoint = os.path.join(exp_dir, f'{prefix}.pth')
if args.encoder_weights:
classifier = get_model('seresnext_cls', num_classes=1)
restore_checkpoint(auto_file(args.encoder_weights), classifier)
encoder_state = classifier.encoder.state_dict()
model.encoder.load_state_dict(encoder_state)
del classifier
if args.resume:
fname = auto_file(args.resume)
start_epoch, train_history, best_score = restore_checkpoint(
fname, model)
print(train_history)
print('Resuming training from epoch', start_epoch, ' and score',
best_score, args.resume)
writer = SummaryWriter(log_dir)
writer.add_text('train/params',
'```' + json.dumps(train_session_args, indent=2) + '```',
0)
# log_model_graph(writer, model)
config_fname = os.path.join(exp_dir, f'{train_session}.json')
with open(config_fname, 'w') as f:
f.write(json.dumps(train_session_args, indent=2))
# Main training phase
model.cuda()
trainable_parameters = filter(lambda p: p.requires_grad,
model.parameters())
optimizer = get_optimizer(args.optimizer,
trainable_parameters,
args.learning_rate,
weight_decay=args.weight_decay)
# scheduler = ReduceLROnPlateau(optimizer, mode='max', patience=50, factor=0.5, min_lr=1e-5)
scheduler = None
train_history, best_metric_val, start_epoch = train(
model,
optimizer,
scheduler,
train_loader,
valid_loader,
writer,
start_epoch,
epochs=args.epochs,
early_stopping=args.early_stopping,
train_history=train_history,
experiment_dir=exp_dir,
best_metric_val=best_metric_val,
checkpoint_filename=best_lb_checkpoint)
train_history.to_csv(os.path.join(exp_dir, 'train_history.csv'),
index=False)
print('Training finished')
del train_loader, valid_loader, optimizer
# Restore to best model
restore_checkpoint(best_lb_checkpoint, model)
# Make OOF predictions
_, valid_ids = get_train_test_split_for_fold(args.fold)
validset_full = D.RSSDDataset(sample_ids=valid_ids,
data_dir=args.data_dir,
transform=get_transform(
training=False,
width=args.patch_size,
height=args.patch_size),
box_coder=model.box_coder)
oof_predictions = model.predict_as_csv(validset_full,
batch_size=args.batch_size,
workers=args.workers)
oof_predictions.to_csv(os.path.join(exp_dir,
f'{prefix}_oof_predictions.csv'),
index=None)
del validset_full
testset_full = D.RSSDDataset(sample_ids=all_test_ids(args.data_dir),
test=True,
data_dir=args.data_dir,
transform=get_transform(
training=False,
width=args.patch_size,
height=args.patch_size),
box_coder=model.box_coder)
test_predictions = model.predict_as_csv(testset_full,
batch_size=args.batch_size,
workers=args.workers)
test_predictions.to_csv(os.path.join(exp_dir,
f'{prefix}_test_predictions.csv'),
index=None)
print('Predictions saved')