opts=dict(showlegend=True,
title='Real output',
ytickmin=0,
ytinkmax=2.0)),
inputs=Image3D(vis, 'inputs'),
fake=Image3D(vis, 'fake'),
valid=Image3D(vis, 'valid'),
outputs=Image3D(vis, 'outputs'),
outputs2=Image3D(vis, 'outputs2'))
x, y = Trainset(FG) # x = image, y=target
# x, y, train_idx, test_idx, ratio = fold_split(FG)
if FG.gm:
transform = Compose([ToFloatTensor()])
else:
transform = Compose([ToFloatTensor(), Normalize(0.5, 0.5)])
# trainset = ADNIDataset2D(FG, x[train_idx], y[train_idx], transform=transform)
# testset = ADNIDataset2D(FG, x[test_idx], y[test_idx], transform=transform)
#
# trainloader = DataLoader(trainset, batch_size=FG.batch_size, shuffle=True,
# pin_memory=True, num_workers=4)
# testloader = DataLoader(testset, batch_size=FG.batch_size, shuffle=True,
# num_workers=4, pin_memory=True)
trainset = ADNIDataset2D(FG, x, y, transform=transform)
trainloader = DataLoader(trainset,
batch_size=FG.batch_size,
shuffle=True,
pin_memory=True)
D = Discriminator(FG).to('cuda:{}'.format(
fake0=Image3D(vis, 'fake0'),
fake1=Image3D(vis, 'fake1'),
valid=Image3D(vis, 'valid'),
outputs0=Image3D(vis, 'outputs0'),
outputs1=Image3D(vis, 'outputs1'))
# dataset setting
x, y = Trainset(FG)
# x, y, train_idx, test_idx, ratio = fold_split(FG)
# transform = Compose([ToFloatTensor(), Normalize(0.5,0.5)])
# trainset = ADNIDataset2D(FG, x, y, transform=transform)
transform = Compose([
ToWoldCoordinateSystem(),
ToTensor(),
Pad(1, 0, 1, 0, 1, 0),
Normalize(0.5, 0.5)
])
trainset = ADNIDataset(FG, x, y, transform=transform)
trainloader = DataLoader(trainset,
batch_size=FG.batch_size,
shuffle=True,
pin_memory=True)
# trainset = ADNIDataset2D(FG, x[train_idx], y[train_idx], transform=transform)
# testset = ADNIDataset2D(FG, x[test_idx], y[test_idx], transform=transform)
# trainloader = DataLoader(trainset, batch_size=FG.batch_size, shuffle=True,
# pin_memory=True, num_workers=4)
# testloader = DataLoader(testset, batch_size=FG.batch_size, shuffle=True,
# num_workers=4, pin_memory=True)
# models
D = infoDiscriminator2D(FG.c_code).to('cuda:{}'.format(FG.devices[0]))
def valid_augment(img,mask):
return DualCompose([Scale(size=pad),ImageOnly(Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)))])(img, mask)
def main():
parser = argparse.ArgumentParser()
arg = parser.add_argument
arg('--jaccard-weight', default=1, type=float)
arg('--device-ids',
type=str,
default='0',
help='For example 0,1 to run on two GPUs')
arg('--fold', type=int, help='fold', default=0)
arg('--root', default='runs/debug', help='checkpoint root')
arg('--batch-size', type=int, default=1)
arg('--n-epochs', type=int, default=100)
arg('--lr', type=float, default=0.0001)
arg('--workers', type=int, default=8)
arg('--type',
type=str,
default='binary',
choices=['binary', 'parts', 'instruments'])
arg('--model',
type=str,
default='UNet',
choices=['UNet', 'UNet11', 'LinkNet34'])
args = parser.parse_args()
root = Path(args.root)
root.mkdir(exist_ok=True, parents=True)
if args.type == 'parts':
num_classes = 4
elif args.type == 'instruments':
num_classes = 8
else:
num_classes = 1
if args.model == 'UNet':
model = UNet(num_classes=num_classes)
elif args.model == 'UNet11':
model = UNet11(num_classes=num_classes, pretrained='vgg')
elif args.model == 'UNet16':
model = UNet16(num_classes=num_classes, pretrained='vgg')
elif args.model == 'LinkNet34':
model = LinkNet34(num_classes=num_classes, pretrained=True)
else:
model = UNet(num_classes=num_classes, input_channels=3)
if torch.cuda.is_available():
if args.device_ids:
device_ids = list(map(int, args.device_ids.split(',')))
else:
device_ids = None
model = nn.DataParallel(model, device_ids=device_ids).cuda()
if args.type == 'binary':
loss = LossBinary(jaccard_weight=args.jaccard_weight)
else:
loss = LossMulti(num_classes=num_classes,
jaccard_weight=args.jaccard_weight)
cudnn.benchmark = True
def make_loader(file_names,
shuffle=False,
transform=None,
problem_type='binary'):
return DataLoader(dataset=RoboticsDataset(file_names,
transform=transform,
problem_type=problem_type),
shuffle=shuffle,
num_workers=args.workers,
batch_size=args.batch_size,
pin_memory=torch.cuda.is_available())
train_file_names, val_file_names = get_split(args.fold)
print('num train = {}, num_val = {}'.format(len(train_file_names),
len(val_file_names)))
train_transform = DualCompose(
[HorizontalFlip(),
VerticalFlip(),
ImageOnly(Normalize())])
val_transform = DualCompose([ImageOnly(Normalize())])
train_loader = make_loader(train_file_names,
shuffle=True,
transform=train_transform,
problem_type=args.type)
valid_loader = make_loader(val_file_names,
transform=val_transform,
problem_type=args.type)
root.joinpath('params.json').write_text(
json.dumps(vars(args), indent=True, sort_keys=True))
if args.type == 'binary':
valid = validation_binary
else:
valid = validation_multi
utils.train(init_optimizer=lambda lr: Adam(model.parameters(), lr=lr),
args=args,
model=model,
criterion=loss,
train_loader=train_loader,
valid_loader=valid_loader,
validation=valid,
fold=args.fold,
num_classes=num_classes)
batch_size=batch_size,
shuffle=False,
drop_last=True)
models_list.append(model_rgb)
data_loaders_list.append(test_loader_rgb)
rgb_included = True
if len(sys.argv) > 3:
model_skeleton = BiGRU(batch_size, input_size, hidden_size, num_layers,
num_classes).to(device)
model_skeleton.load_state_dict(torch.load(sys.argv[3]))
test_dataset_skeleton = UtdMhadDataset(modality='skeleton',
train=False,
transform=Compose([
Normalize((0, 2)),
Resize(num_frames),
FilterJoints(selected_joints),
ToSequence(
sequence_length, input_size)
]))
test_loader_skeleton = DataLoader(dataset=test_dataset_skeleton,
batch_size=batch_size,
shuffle=False,
drop_last=True)
models_list.append(model_skeleton)
data_loaders_list.append(test_loader_skeleton)
skeleton_included = True
# Get three different accuracies for each different fusion rule
max_acc, prod_acc, sum_acc = get_score_fusion_accuracy(data_loaders_list,
def main(rank, args):
# Distributed setup
if args.distributed:
setup_distributed(rank, args.world_size)
not_main_rank = args.distributed and rank != 0
logging.info("Start time: %s", datetime.now())
# Explicitly set seed to make sure models created in separate processes
# start from same random weights and biases
torch.manual_seed(args.seed)
# Empty CUDA cache
torch.cuda.empty_cache()
# Change backend for flac files
torchaudio.set_audio_backend("soundfile")
# Transforms
melkwargs = {
"n_fft": args.win_length,
"n_mels": args.n_bins,
"hop_length": args.hop_length,
}
sample_rate_original = 16000
if args.type == "mfcc":
transforms = torch.nn.Sequential(
torchaudio.transforms.MFCC(
sample_rate=sample_rate_original,
n_mfcc=args.n_bins,
melkwargs=melkwargs,
), )
num_features = args.n_bins
elif args.type == "waveform":
transforms = torch.nn.Sequential(UnsqueezeFirst())
num_features = 1
else:
raise ValueError("Model type not supported")
if args.normalize:
transforms = torch.nn.Sequential(transforms, Normalize())
augmentations = torch.nn.Sequential()
if args.freq_mask:
augmentations = torch.nn.Sequential(
augmentations,
torchaudio.transforms.FrequencyMasking(
freq_mask_param=args.freq_mask),
)
if args.time_mask:
augmentations = torch.nn.Sequential(
augmentations,
torchaudio.transforms.TimeMasking(time_mask_param=args.time_mask),
)
# Text preprocessing
char_blank = "*"
char_space = " "
char_apostrophe = "'"
labels = char_blank + char_space + char_apostrophe + string.ascii_lowercase
language_model = LanguageModel(labels, char_blank, char_space)
# Dataset
training, validation = split_process_librispeech(
[args.dataset_train, args.dataset_valid],
[transforms, transforms],
language_model,
root=args.dataset_root,
folder_in_archive=args.dataset_folder_in_archive,
)
# Decoder
if args.decoder == "greedy":
decoder = GreedyDecoder()
else:
raise ValueError("Selected decoder not supported")
# Model
model = Wav2Letter(
num_classes=language_model.length,
input_type=args.type,
num_features=num_features,
)
if args.jit:
model = torch.jit.script(model)
if args.distributed:
n = torch.cuda.device_count() // args.world_size
devices = list(range(rank * n, (rank + 1) * n))
model = model.to(devices[0])
model = torch.nn.parallel.DistributedDataParallel(model,
device_ids=devices)
else:
devices = ["cuda" if torch.cuda.is_available() else "cpu"]
model = model.to(devices[0], non_blocking=True)
model = torch.nn.DataParallel(model)
n = count_parameters(model)
logging.info("Number of parameters: %s", n)
# Optimizer
if args.optimizer == "adadelta":
optimizer = Adadelta(
model.parameters(),
lr=args.learning_rate,
weight_decay=args.weight_decay,
eps=args.eps,
rho=args.rho,
)
elif args.optimizer == "sgd":
optimizer = SGD(
model.parameters(),
lr=args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay,
)
elif args.optimizer == "adam":
optimizer = Adam(
model.parameters(),
lr=args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay,
)
elif args.optimizer == "adamw":
optimizer = AdamW(
model.parameters(),
lr=args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay,
)
else:
raise ValueError("Selected optimizer not supported")
if args.scheduler == "exponential":
scheduler = ExponentialLR(optimizer, gamma=args.gamma)
elif args.scheduler == "reduceonplateau":
scheduler = ReduceLROnPlateau(optimizer, patience=10, threshold=1e-3)
else:
raise ValueError("Selected scheduler not supported")
criterion = torch.nn.CTCLoss(blank=language_model.mapping[char_blank],
zero_infinity=False)
# Data Loader
collate_fn_train = collate_factory(model_length_function, augmentations)
collate_fn_valid = collate_factory(model_length_function)
loader_training_params = {
"num_workers": args.workers,
"pin_memory": True,
"shuffle": True,
"drop_last": True,
}
loader_validation_params = loader_training_params.copy()
loader_validation_params["shuffle"] = False
loader_training = DataLoader(
training,
batch_size=args.batch_size,
collate_fn=collate_fn_train,
**loader_training_params,
)
loader_validation = DataLoader(
validation,
batch_size=args.batch_size,
collate_fn=collate_fn_valid,
**loader_validation_params,
)
# Setup checkpoint
best_loss = 1.0
load_checkpoint = args.checkpoint and os.path.isfile(args.checkpoint)
if args.distributed:
torch.distributed.barrier()
if load_checkpoint:
logging.info("Checkpoint: loading %s", args.checkpoint)
checkpoint = torch.load(args.checkpoint)
args.start_epoch = checkpoint["epoch"]
best_loss = checkpoint["best_loss"]
model.load_state_dict(checkpoint["state_dict"])
optimizer.load_state_dict(checkpoint["optimizer"])
scheduler.load_state_dict(checkpoint["scheduler"])
logging.info("Checkpoint: loaded '%s' at epoch %s", args.checkpoint,
checkpoint["epoch"])
else:
logging.info("Checkpoint: not found")
save_checkpoint(
{
"epoch": args.start_epoch,
"state_dict": model.state_dict(),
"best_loss": best_loss,
"optimizer": optimizer.state_dict(),
"scheduler": scheduler.state_dict(),
},
False,
args.checkpoint,
not_main_rank,
)
if args.distributed:
torch.distributed.barrier()
torch.autograd.set_detect_anomaly(False)
for epoch in range(args.start_epoch, args.epochs):
logging.info("Epoch: %s", epoch)
train_one_epoch(
model,
criterion,
optimizer,
scheduler,
loader_training,
decoder,
language_model,
devices[0],
epoch,
args.clip_grad,
not_main_rank,
not args.reduce_lr_valid,
)
loss = evaluate(
model,
criterion,
loader_validation,
decoder,
language_model,
devices[0],
epoch,
not_main_rank,
)
if args.reduce_lr_valid and isinstance(scheduler, ReduceLROnPlateau):
scheduler.step(loss)
is_best = loss < best_loss
best_loss = min(loss, best_loss)
save_checkpoint(
{
"epoch": epoch + 1,
"state_dict": model.state_dict(),
"best_loss": best_loss,
"optimizer": optimizer.state_dict(),
"scheduler": scheduler.state_dict(),
},
is_best,
args.checkpoint,
not_main_rank,
)
logging.info("End time: %s", datetime.now())
if args.distributed:
torch.distributed.destroy_process_group()
showlegend=True, title='D loss', ytickmin=0, ytinkmax=2.0)),
G_loss = Scalar(vis, 'G_loss', opts=dict(
showlegend=True, title='G loss', ytickmin=0, ytinkmax=10)),
info_loss = Scalar(vis, 'info_loss', opts=dict(
showlegend=True, title='info loss', ytickmin=0, ytinkmax=10)),
input = Image3D(vis, 'input'),
output = Image3D(vis, 'output'),
cont_output = Image3D(vis, 'cont_output'),
disc_output = Image3D(vis, 'disc_output'))
# create train set
x, y = Trainset(FG) # x = image, y=target
if FG.gm == 'true':
transform=Compose([ToWoldCoordinateSystem(), ToTensor()])
else :
transform=Compose([ToWoldCoordinateSystem(), Normalize(0.2,0.9), ToTensor()])
trainset = ADNIDataset(FG, x, y, transform=transform)
trainloader = DataLoader(trainset, batch_size=FG.batch_size,
shuffle=True, pin_memory=True,
num_workers=4)
for _, data in enumerate(trainloader):
data = data['image']
break
if '3d' in FG.model:
G = Generator(FG).to('cuda:{}'.format(FG.devices[0]))
D = Discriminator(FG).to('cuda:{}'.format(FG.devices[0]))
elif '2d' in FG.model:
G = sliceGenerator(FG, output_dim=data.shape[1]).to('cuda:{}'.format(FG.devices[0]))
D = sliceDiscriminator(FG, input_dim=data.shape[1]).to('cuda:{}'.format(FG.devices[0]))
elif 'resnet' in FG.model:
if __name__ == '__main__':
tbwiter = SummaryWriter(log_dir= config.LOG_DIR_PATH)
print("Initializing AlexNet model")
alexnet = alexnetmodel.AlexNet(num_classes = config.PARAMETERS['NUM_CLASSES'])
alexnet = torch.nn.parallel.DataParallel(alexnet, device_ids=config.DEVICE_IDS)
print("Initilaizing transfrmations to apply on the image")
transformations = tf.Compose([
CenterCrop(227),
# HorizontalFlip(),
ToTensor(),
Normalize(mean = (0.485, 0.456, 0.406), std = (0.229, 0.224, 0.225))
]
)
print("Generating train images and labels")
train_X, train_Y, enc = separate_dataset_and_classes()
print("Generating Validation images and labels")
val_X, val_Y = separate_classes_val_test(enc = enc)
print("Initilazing data loader for Alexnet")
train_dataset = alexnetdataloader.AlexNetDataLoader(train_X, train_Y, transform = transformations)
val_dataset = alexnetdataloader.AlexNetDataLoader(val_X, val_Y, None)
train_dataloader = data.DataLoader(
train_dataset,
def main(args):
# keep shuffling be constant every time
seed = 1
torch.manual_seed(seed)
# torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed) # if you are using multi-GPU.
# norm_method = Normalize(args.mean, args.std)
norm_method = Normalize([0, 0, 0], [1, 1, 1])
# scales = [0.5, 0.6, 0.7, 0.8, 0.9]
trans_train = Compose([
# Scale(args.img_size),
# MultiScaleRandomCrop(scales, args.img_size),
# CenterCrop(args.img_size),
RandomHorizontalFlip(),
ToTensor(1),
norm_method
])
trans_test = Compose([
# Scale(args.img_size),
# CenterCrop(args.img_size),
ToTensor(1),
norm_method
])
if args.is_train:
# dataset_train = dataset_R3D(
# root_dir, 'train', args.data_type, args.n_frames_per_clip,
# img_size=args.img_size, stride=1, overlap=True, reverse=True, transform=trans_train)
# dataloader_train = DataLoader(dataset_train, batch_size=args.batch_size,
# shuffle=True,num_workers=12,pin_memory=True)
# dataset_val = dataset_R3D(
# root_dir, 'val', args.data_type, args.n_frames_per_clip,
# img_size=args.img_size, stride=1, overlap=False, transform=trans_test)
# sampler = RandomSampler(dataset_val, replacement=True, num_samples=args.batch_size*5)
# dataloader_val = DataLoader(dataset_val, batch_size=16,
# num_workers=8,sampler=sampler,
# pin_memory=True)
# dataloader for phase 2
mask_trans = transforms.Compose([
# transforms.Resize((126,224)),
# transforms.CenterCrop((126,224)),
transforms.ToTensor()
])
print('Loading phase2 training data.....')
dataset_train = dataset_unequal.dataset_all(
root_dir,
'train',
n_frames_per_clip=args.n_frames_per_clip,
UnequalSequence=True,
img_size=(args.w, args.h),
stride=2,
reverse=False,
transform=trans_train,
mask_trans=mask_trans)
dataloader_train = DataLoader(dataset_train,
batch_size=128,
shuffle=True,
num_workers=args.num_workers,
pin_memory=True)
print('\n')
print('Loading phase2 validating data.....')
dataset_val = dataset_unequal.dataset_all(
root_dir,
'val',
n_frames_per_clip=args.n_frames_per_clip,
img_size=(args.w, args.h),
stride=2,
UnequalSequence=True,
reverse=False,
transform=trans_test,
mask_trans=mask_trans)
sampler = RandomSampler(dataset_val,
replacement=True,
num_samples=1024)
dataloader_val = DataLoader(dataset_val,
batch_size=64,
sampler=sampler,
num_workers=args.num_workers,
pin_memory=True)
else:
dataset_test = dataset_R3D(root_dir,
'test',
args.data_type,
args.n_frames_per_clip,
img_size=args.img_size,
stride=args.n_frames_per_clip,
overlap=False,
transform=trans_test)
dataloader_test = DataLoader(dataset_test,
batch_size=128,
shuffle=True,
num_workers=8,
pin_memory=True)
if args.is_train:
fine_tune(model_dir,
save_dir,
'resnet-152-kinetics.pth',
dataloader_train,
dataloader_val,
ContinuousTrain=False)
pdb.set_trace()
else:
model_test('output', 'checkpoint-3.pth', dataloader_test)
def main():
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size',
type=int,
default=64,
metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size',
type=int,
default=1000,
metavar='N',
help='input batch size for testing (default: 1000)')
parser.add_argument('--world_size',
type=int,
default=1,
help='number of GPUs to use')
parser.add_argument('--epochs',
type=int,
default=10,
metavar='N',
help='number of epochs to train (default: 10)')
parser.add_argument('--lr',
type=float,
default=0.01,
metavar='LR',
help='learning rate (default: 0.01)')
parser.add_argument('--wd',
type=float,
default=1e-4,
help='weight decay (default: 5e-4)')
parser.add_argument('--lr-decay-every',
type=int,
default=100,
help='learning rate decay by 10 every X epochs')
parser.add_argument('--lr-decay-scalar',
type=float,
default=0.1,
help='--')
parser.add_argument('--momentum',
type=float,
default=0.5,
metavar='M',
help='SGD momentum (default: 0.5)')
parser.add_argument('--no-cuda',
action='store_true',
default=False,
help='disables CUDA training')
parser.add_argument('--seed',
type=int,
default=1,
metavar='S',
help='random seed (default: 1)')
parser.add_argument(
'--log-interval',
type=int,
default=10,
metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--run_test',
default=False,
type=str2bool,
nargs='?',
help='run test only')
parser.add_argument(
'--limit_training_batches',
type=int,
default=-1,
help='how many batches to do per training, -1 means as many as possible'
)
parser.add_argument('--no_grad_clip',
default=False,
type=str2bool,
nargs='?',
help='turn off gradient clipping')
parser.add_argument('--get_flops',
default=False,
type=str2bool,
nargs='?',
help='add hooks to compute flops')
parser.add_argument(
'--get_inference_time',
default=False,
type=str2bool,
nargs='?',
help='runs valid multiple times and reports the result')
parser.add_argument('--mgpu',
default=False,
type=str2bool,
nargs='?',
help='use data paralization via multiple GPUs')
parser.add_argument('--dataset',
default="MNIST",
type=str,
help='dataset for experiment, choice: MNIST, CIFAR10')
parser.add_argument('--data',
metavar='DIR',
default='/imagenet',
help='path to imagenet dataset')
parser.add_argument(
'--model',
default="lenet3",
type=str,
help='model selection, choices: lenet3, vgg, mobilenetv2, resnet18',
choices=[
"lenet3", "vgg", "mobilenetv2", "resnet18", "resnet152",
"resnet50", "resnet50_noskip", "resnet20", "resnet34", "resnet101",
"resnet101_noskip", "densenet201_imagenet", 'densenet121_imagenet',
"multprun_gate5_gpu_0316_1", "mult_prun8_gpu", "multnas5_gpu"
])
parser.add_argument('--tensorboard',
type=str2bool,
nargs='?',
help='Log progress to TensorBoard')
parser.add_argument(
'--save_models',
default=True,
type=str2bool,
nargs='?',
help='if True, models will be saved to the local folder')
parser.add_argument('--fineturn_model',
type=str2bool,
nargs='?',
help='Log progress to TensorBoard')
# ============================PRUNING added
parser.add_argument(
'--pruning_config',
default=None,
type=str,
help=
'path to pruning configuration file, will overwrite all pruning parameters in arguments'
)
parser.add_argument('--group_wd_coeff',
type=float,
default=0.0,
help='group weight decay')
parser.add_argument('--name',
default='test',
type=str,
help='experiment name(folder) to store logs')
parser.add_argument(
'--augment',
default=False,
type=str2bool,
nargs='?',
help=
'enable or not augmentation of training dataset, only for CIFAR, def False'
)
parser.add_argument('--load_model',
default='',
type=str,
help='path to model weights')
parser.add_argument('--pruning',
default=False,
type=str2bool,
nargs='?',
help='enable or not pruning, def False')
parser.add_argument(
'--pruning-threshold',
'--pt',
default=100.0,
type=float,
help=
'Max error perc on validation set while pruning (default: 100.0 means always prune)'
)
parser.add_argument(
'--pruning-momentum',
default=0.0,
type=float,
help=
'Use momentum on criteria between pruning iterations, def 0.0 means no momentum'
)
parser.add_argument('--pruning-step',
default=15,
type=int,
help='How often to check loss and do pruning step')
parser.add_argument('--prune_per_iteration',
default=10,
type=int,
help='How many neurons to remove at each iteration')
parser.add_argument(
'--fixed_layer',
default=-1,
type=int,
help='Prune only a given layer with index, use -1 to prune all')
parser.add_argument('--start_pruning_after_n_iterations',
default=0,
type=int,
help='from which iteration to start pruning')
parser.add_argument('--maximum_pruning_iterations',
default=1e8,
type=int,
help='maximum pruning iterations')
parser.add_argument('--starting_neuron',
default=0,
type=int,
help='starting position for oracle pruning')
parser.add_argument('--prune_neurons_max',
default=-1,
type=int,
help='prune_neurons_max')
parser.add_argument('--pruning-method',
default=0,
type=int,
help='pruning method to be used, see readme.md')
parser.add_argument('--pruning_fixed_criteria',
default=False,
type=str2bool,
nargs='?',
help='enable or not criteria reevaluation, def False')
parser.add_argument('--fixed_network',
default=False,
type=str2bool,
nargs='?',
help='fix network for oracle or criteria computation')
parser.add_argument(
'--zero_lr_for_epochs',
default=-1,
type=int,
help='Learning rate will be set to 0 for given number of updates')
parser.add_argument(
'--dynamic_network',
default=False,
type=str2bool,
nargs='?',
help=
'Creates a new network graph from pruned model, works with ResNet-101 only'
)
parser.add_argument('--use_test_as_train',
default=False,
type=str2bool,
nargs='?',
help='use testing dataset instead of training')
parser.add_argument('--pruning_mask_from',
default='',
type=str,
help='path to mask file precomputed')
parser.add_argument(
'--compute_flops',
default=True,
type=str2bool,
nargs='?',
help=
'if True, will run dummy inference of batch 1 before training to get conv sizes'
)
# ============================END pruning added
best_prec1 = 0
global global_iteration
global group_wd_optimizer
global_iteration = 0
args = parser.parse_args()
use_cuda = not args.no_cuda and torch.cuda.is_available()
torch.manual_seed(args.seed)
args.distributed = args.world_size > 1
if args.distributed:
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size,
rank=0)
device = torch.device("cuda" if use_cuda else "cpu")
# dataset loading section
if args.dataset == "MNIST":
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
train_loader = torch.utils.data.DataLoader(datasets.MNIST(
'../data',
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])),
batch_size=args.batch_size,
shuffle=True,
**kwargs)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('../data',
train=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])),
batch_size=args.test_batch_size,
shuffle=True,
**kwargs)
elif args.dataset == "CIFAR10":
# Data loading code
normalize = transforms.Normalize(
mean=[x / 255.0 for x in [125.3, 123.0, 113.9]],
std=[x / 255.0 for x in [63.0, 62.1, 66.7]])
if args.augment:
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
else:
transform_train = transforms.Compose([
transforms.ToTensor(),
normalize,
])
transform_test = transforms.Compose([transforms.ToTensor(), normalize])
kwargs = {'num_workers': 8, 'pin_memory': True}
train_loader = torch.utils.data.DataLoader(datasets.CIFAR10(
'../data', train=True, download=True, transform=transform_train),
batch_size=args.batch_size,
shuffle=True,
drop_last=True,
**kwargs)
test_loader = torch.utils.data.DataLoader(
datasets.CIFAR10('../data', train=False, transform=transform_test),
batch_size=args.test_batch_size,
shuffle=True,
**kwargs)
elif args.dataset == "Imagenet":
traindir = os.path.join(args.data, 'train')
valdir = os.path.join(args.data, 'val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
else:
train_sampler = None
kwargs = {'num_workers': 16}
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
sampler=train_sampler,
pin_memory=True,
**kwargs)
if args.use_test_as_train:
train_loader = torch.utils.data.DataLoader(
datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size,
shuffle=(train_sampler is None),
**kwargs)
test_loader = torch.utils.data.DataLoader(datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size,
shuffle=False,
pin_memory=True,
**kwargs)
#wm
elif args.dataset == "mult_5T":
args.data_root = ['/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/CX_20200709',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TK_20200709',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/ZR_20200709',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TX_20200616',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/WM_20200709']
args.data_root_val = ['/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/CX_20200709',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TK_20200709',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/ZR_20200709',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TX_20200616',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/WM_20200709']
args.train_data_list = ['/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/CX_20200709/txt/cx_train.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TK_20200709/txt/tk_train.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/ZR_20200709/txt/zr_train.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TX_20200616/txt/tx_train.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/WM_20200709/txt/wm_train.txt']
args.val_data_list = ['/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/CX_20200709/txt/cx_val.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TK_20200709/txt/tk_val.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/ZR_20200709/txt/zr_val.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TX_20200616/txt/tx_val.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/WM_20200709/txt/wm_val.txt']
num_tasks = len(args.data_root)
args.ngpu = 8
args.workers = 8
args.train_batch_size = [40, 40, 40, 40, 40] #36
args.val_batch_size = [100, 100, 100, 100, 100]
args.loss_weight = [1.0, 1.0, 1.0, 1.0, 1.0]
args.val_num_classes = [[0, 1, 2, 3, 4], [0, 1, 2], [0, 1], [0, 1],
[0, 1, 2, 3, 4, 5, 6]]
args.mixup_alpha = None #None
for i in range(num_tasks):
args.train_batch_size[i] *= args.ngpu
args.val_batch_size[i] *= args.ngpu
pixel_mean = [0.406, 0.456, 0.485]
pixel_std = [0.225, 0.224, 0.229]
#私人定制：
train_dataset = []
for i in range(num_tasks):
if i == 1:
train_dataset.append(
FileListLabeledDataset(
args.train_data_list[i],
args.data_root[i],
Compose([
RandomResizedCrop(
112,
scale=(0.94, 1.),
ratio=(1. / 4., 4. / 1.)
), #scale=(0.7, 1.2), ratio=(1. / 1., 4. / 1.)
RandomHorizontalFlip(),
ColorJitter(brightness=[0.5, 1.5],
contrast=[0.5, 1.5],
saturation=[0.5, 1.5],
hue=0),
ToTensor(),
Lighting(1, [0.2175, 0.0188, 0.0045],
[[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]]),
Normalize(pixel_mean, pixel_std),
])))
else:
train_dataset.append(
FileListLabeledDataset(
args.train_data_list[i], args.data_root[i],
Compose([
RandomResizedCrop(112,
scale=(0.7, 1.2),
ratio=(1. / 1., 4. / 1.)),
RandomHorizontalFlip(),
ColorJitter(brightness=[0.5, 1.5],
contrast=[0.5, 1.5],
saturation=[0.5, 1.5],
hue=0),
ToTensor(),
Lighting(1, [0.2175, 0.0188, 0.0045],
[[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]]),
Normalize(pixel_mean, pixel_std),
])))
#原来的
# train_dataset = [FileListLabeledDataset(
# args.train_data_list[i], args.data_root[i],
# Compose([
# RandomResizedCrop(112,scale=(0.7, 1.2), ratio=(1. / 1., 4. / 1.)),
# RandomHorizontalFlip(),
# ColorJitter(brightness=[0.5,1.5], contrast=[0.5,1.5], saturation=[0.5,1.5], hue= 0),
# ToTensor(),
# Lighting(1, [0.2175, 0.0188, 0.0045], [[-0.5675, 0.7192, 0.4009], [-0.5808, -0.0045, -0.8140], [-0.5836, -0.6948, 0.4203]]),
# Normalize(pixel_mean, pixel_std),]),
# memcached=False,
# memcached_client="") for i in range(num_tasks)]
args.num_classes = [td.num_class for td in train_dataset]
train_longest_size = max([
int(np.ceil(len(td) / float(bs)))
for td, bs in zip(train_dataset, args.train_batch_size)
])
train_sampler = [
GivenSizeSampler(td,
total_size=train_longest_size * bs,
rand_seed=0)
for td, bs in zip(train_dataset, args.train_batch_size)
]
train_loader = [
DataLoader(train_dataset[k],
batch_size=args.train_batch_size[k],
shuffle=False,
num_workers=args.workers,
pin_memory=False,
sampler=train_sampler[k]) for k in range(num_tasks)
]
val_dataset = [
FileListLabeledDataset(
args.val_data_list[i],
args.data_root_val[i],
Compose([
Resize((112, 112)),
# CenterCrop(112),
ToTensor(),
Normalize(pixel_mean, pixel_std),
]),
memcached=False,
memcached_client="") for i in range(num_tasks)
]
val_longest_size = max([
int(np.ceil(len(vd) / float(bs)))
for vd, bs in zip(val_dataset, args.val_batch_size)
])
test_loader = [
DataLoader(val_dataset[k],
batch_size=args.val_batch_size[k],
shuffle=False,
num_workers=args.workers,
pin_memory=False) for k in range(num_tasks)
]
if args.model == "lenet3":
model = LeNet(dataset=args.dataset)
elif args.model == "vgg":
model = vgg11_bn(pretrained=True)
elif args.model == "resnet18":
model = PreActResNet18()
elif (args.model == "resnet50") or (args.model == "resnet50_noskip"):
if args.dataset == "CIFAR10":
model = PreActResNet50(dataset=args.dataset)
else:
from models.resnet import resnet50
skip_gate = True
if "noskip" in args.model:
skip_gate = False
if args.pruning_method not in [22, 40]:
skip_gate = False
model = resnet50(skip_gate=skip_gate)
elif args.model == "resnet34":
if not (args.dataset == "CIFAR10"):
from models.resnet import resnet34
model = resnet34()
elif args.model == "multprun_gate5_gpu_0316_1":
from models.multitask import MultiTaskWithLoss
model = MultiTaskWithLoss(backbone=args.model,
num_classes=args.num_classes,
feature_dim=2560,
spatial_size=112,
arc_fc=False,
feat_bn=False)
print(model)
elif args.model == "mult_prun8_gpu":
from models.multitask import MultiTaskWithLoss
model = MultiTaskWithLoss(backbone=args.model,
num_classes=args.num_classes,
feature_dim=18,
spatial_size=112,
arc_fc=False,
feat_bn=False)
print(model)
elif args.model == "multnas5_gpu": #作为修改项
from models.multitask import MultiTaskWithLoss
model = MultiTaskWithLoss(backbone=args.model,
num_classes=args.num_classes,
feature_dim=512,
spatial_size=112,
arc_fc=False,
feat_bn=False)
print(model)
elif "resnet101" in args.model:
if not (args.dataset == "CIFAR10"):
from models.resnet import resnet101
if args.dataset == "Imagenet":
classes = 1000
if "noskip" in args.model:
model = resnet101(num_classes=classes, skip_gate=False)
else:
model = resnet101(num_classes=classes)
elif args.model == "resnet20":
if args.dataset == "CIFAR10":
NotImplementedError(
"resnet20 is not implemented in the current project")
# from models.resnet_cifar import resnet20
# model = resnet20()
elif args.model == "resnet152":
model = PreActResNet152()
elif args.model == "densenet201_imagenet":
from models.densenet_imagenet import DenseNet201
model = DenseNet201(gate_types=['output_bn'], pretrained=True)
elif args.model == "densenet121_imagenet":
from models.densenet_imagenet import DenseNet121
model = DenseNet121(gate_types=['output_bn'], pretrained=True)
else:
print(args.model, "model is not supported")
####end dataset preparation
if args.dynamic_network:
# attempts to load pruned model and modify it be removing pruned channels
# works for resnet101 only
if (len(args.load_model) > 0) and (args.dynamic_network):
if os.path.isfile(args.load_model):
load_model_pytorch(model, args.load_model, args.model)
else:
print("=> no checkpoint found at '{}'".format(args.load_model))
exit()
dynamic_network_change_local(model)
# save the model
log_save_folder = "%s" % args.name
if not os.path.exists(log_save_folder):
os.makedirs(log_save_folder)
if not os.path.exists("%s/models" % (log_save_folder)):
os.makedirs("%s/models" % (log_save_folder))
model_save_path = "%s/models/pruned.weights" % (log_save_folder)
model_state_dict = model.state_dict()
if args.save_models:
save_checkpoint({'state_dict': model_state_dict},
False,
filename=model_save_path)
print("model is defined")
# aux function to get size of feature maps
# First it adds hooks for each conv layer
# Then runs inference with 1 image
output_sizes = get_conv_sizes(args, model)
if use_cuda and not args.mgpu:
model = model.to(device)
elif args.distributed:
model.cuda()
print(
"\n\n WARNING: distributed pruning was not verified and might not work correctly"
)
model = torch.nn.parallel.DistributedDataParallel(model)
elif args.mgpu:
model = torch.nn.DataParallel(model).cuda()
else:
model = model.to(device)
print(
"model is set to device: use_cuda {}, args.mgpu {}, agrs.distributed {}"
.format(use_cuda, args.mgpu, args.distributed))
weight_decay = args.wd
if args.fixed_network:
weight_decay = 0.0
# remove updates from gate layers, because we want them to be 0 or 1 constantly
if 1:
parameters_for_update = []
parameters_for_update_named = []
for name, m in model.named_parameters():
if "gate" not in name:
parameters_for_update.append(m)
parameters_for_update_named.append((name, m))
else:
print("skipping parameter", name, "shape:", m.shape)
total_size_params = sum(
[np.prod(par.shape) for par in parameters_for_update])
print("Total number of parameters, w/o usage of bn consts: ",
total_size_params)
optimizer = optim.SGD(parameters_for_update,
lr=args.lr,
momentum=args.momentum,
weight_decay=weight_decay)
if 1:
# helping optimizer to implement group lasso (with very small weight that doesn't affect training)
# will be used to calculate number of remaining flops and parameters in the network
group_wd_optimizer = group_lasso_decay(
parameters_for_update,
group_lasso_weight=args.group_wd_coeff,
named_parameters=parameters_for_update_named,
output_sizes=output_sizes)
cudnn.benchmark = True
# define objective
criterion = nn.CrossEntropyLoss()
###=======================added for pruning
# logging part
log_save_folder = "%s" % args.name
if not os.path.exists(log_save_folder):
os.makedirs(log_save_folder)
if not os.path.exists("%s/models" % (log_save_folder)):
os.makedirs("%s/models" % (log_save_folder))
train_writer = None
if args.tensorboard:
try:
# tensorboardX v1.6
train_writer = SummaryWriter(log_dir="%s" % (log_save_folder))
except:
# tensorboardX v1.7
train_writer = SummaryWriter(logdir="%s" % (log_save_folder))
time_point = time.strftime("%Y_%m_%d_%H_%M_%S", time.localtime())
textfile = "%s/log_%s.txt" % (log_save_folder, time_point)
stdout = Logger(textfile)
sys.stdout = stdout
print(" ".join(sys.argv))
# initializing parameters for pruning
# we can add weights of different layers or we can add gates (multiplies output with 1, useful only for gradient computation)
pruning_engine = None
if args.pruning:
pruning_settings = dict()
if not (args.pruning_config is None):
pruning_settings_reader = PruningConfigReader()
pruning_settings_reader.read_config(args.pruning_config)
pruning_settings = pruning_settings_reader.get_parameters()
# overwrite parameters from config file with those from command line
# needs manual entry here
# user_specified = [key for key in vars(default_args).keys() if not (vars(default_args)[key]==vars(args)[key])]
# argv_of_interest = ['pruning_threshold', 'pruning-momentum', 'pruning_step', 'prune_per_iteration',
# 'fixed_layer', 'start_pruning_after_n_iterations', 'maximum_pruning_iterations',
# 'starting_neuron', 'prune_neurons_max', 'pruning_method']
has_attribute = lambda x: any([x in a for a in sys.argv])
if has_attribute('pruning-momentum'):
pruning_settings['pruning_momentum'] = vars(
args)['pruning_momentum']
if has_attribute('pruning-method'):
pruning_settings['method'] = vars(args)['pruning_method']
pruning_parameters_list = prepare_pruning_list(
pruning_settings,
model,
model_name=args.model,
pruning_mask_from=args.pruning_mask_from,
name=args.name)
print("Total pruning layers:", len(pruning_parameters_list))
folder_to_write = "%s" % log_save_folder + "/"
log_folder = folder_to_write
pruning_engine = pytorch_pruning(pruning_parameters_list,
pruning_settings=pruning_settings,
log_folder=log_folder)
pruning_engine.connect_tensorboard(train_writer)
pruning_engine.dataset = args.dataset
pruning_engine.model = args.model
pruning_engine.pruning_mask_from = args.pruning_mask_from
pruning_engine.load_mask()
gates_to_params = connect_gates_with_parameters_for_flops(
args.model, parameters_for_update_named)
pruning_engine.gates_to_params = gates_to_params
###=======================end for pruning
# loading model file
if (len(args.load_model) > 0) and (not args.dynamic_network):
if os.path.isfile(args.load_model):
if args.fineturn_model:
checkpoint = torch.load(args.load_model)
state_dict = checkpoint['state_dict']
model = load_module_state_dict_checkpoint(model, state_dict)
else:
load_model_pytorch(model, args.load_model, args.model)
else:
print("=> no checkpoint found at '{}'".format(args.load_model))
exit()
if args.tensorboard and 0:
if args.dataset == "CIFAR10":
dummy_input = torch.rand(1, 3, 32, 32).to(device)
elif args.dataset == "Imagenet":
dummy_input = torch.rand(1, 3, 224, 224).to(device)
train_writer.add_graph(model, dummy_input)
for epoch in range(1, args.epochs + 1):
if args.distributed:
train_sampler.set_epoch(epoch)
adjust_learning_rate(args, optimizer, epoch, args.zero_lr_for_epochs,
train_writer)
if not args.run_test and not args.get_inference_time:
train(args,
model,
device,
train_loader,
optimizer,
epoch,
criterion,
train_writer=train_writer,
pruning_engine=pruning_engine)
if args.pruning:
# skip validation error calculation and model saving
if pruning_engine.method == 50: continue
# evaluate on validation set
prec1 = validate(args,
test_loader,
model,
device,
criterion,
epoch,
train_writer=train_writer)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
model_save_path = "%s/models/checkpoint.weights" % (log_save_folder)
paths = "%s/models" % (log_save_folder)
model_state_dict = model.state_dict()
if args.save_models:
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model_state_dict,
'best_prec1': best_prec1,
},
is_best,
filename=model_save_path)
states = {
'epoch': epoch + 1,
'state_dict': model_state_dict,
}
torch.save(states, '{}/{}.pth.tar'.format(paths, epoch + 1))
# scale = 4
img_width = 256 #1280 // scale
img_height = 256 #1024 // scale
offset = 30
loss_ratio = 0.5
train_transform = DualCompose([
Resize(w=img_width + offset, h=img_height + offset),
RandomCrop(size=(img_height, img_width)),
HorizontalFlip(),
VerticalFlip(),
RandomColorDual(limit=0.3, prob=1.0),
# RandomBrightnessDual(limit=0.3),
# RandomContrastDual(limit=0.3),
# RandomSaturationDual(limit=0.3),
Normalize(normalize_mask=True)
])
valid_transform = DualCompose(
[Resize(w=img_width, h=img_height),
Normalize(normalize_mask=True)])
train_dataset = Challenge2018ColorizationDataset(
image_file_names=train_file_names,
to_augment=True,
transform=train_transform)
val_dataset = Challenge2018ColorizationDataset(
image_file_names=val_file_names[::20],
to_augment=True,
transform=valid_transform)
def init(batch_size, state, input_sizes, dataset, mean, std, base, workers=10, method='baseline'):
# Return data_loaders
# depending on whether the state is
# 0: training
# 1: fast validation by mean IoU (validation set)
# 2: just testing (test set)
# 3: just testing (validation set)
# Transformations
# ! Can't use torchvision.Transforms.Compose
transforms_test = Compose(
[Resize(size_image=input_sizes[0], size_label=input_sizes[0]),
ToTensor(),
Normalize(mean=mean, std=std)])
transforms_train = Compose(
[Resize(size_image=input_sizes[0], size_label=input_sizes[0]),
RandomRotation(degrees=3),
ToTensor(),
Normalize(mean=mean, std=std, normalize_target=True if method == 'lstr' else False)])
# Batch builder
if method == 'lstr':
collate_fn = dict_collate_fn
else:
collate_fn = None
if state == 0:
if method == 'lstr':
if dataset == 'tusimple':
data_set = TuSimple(root=base, image_set='train', transforms=transforms_train,
padding_mask=True, process_points=True)
elif dataset == 'culane':
data_set = CULane(root=base, image_set='train', transforms=transforms_train,
padding_mask=True, process_points=True)
elif dataset == 'llamas':
data_set = LLAMAS(root=base, image_set='train', transforms=transforms_train,
padding_mask=True, process_points=True)
else:
raise ValueError
else:
data_set = StandardLaneDetectionDataset(root=base, image_set='train', transforms=transforms_train,
data_set=dataset)
data_loader = torch.utils.data.DataLoader(dataset=data_set, batch_size=batch_size, collate_fn=collate_fn,
num_workers=workers, shuffle=True)
validation_set = StandardLaneDetectionDataset(root=base, image_set='val',
transforms=transforms_test, data_set=dataset)
validation_loader = torch.utils.data.DataLoader(dataset=validation_set, batch_size=batch_size * 4,
num_workers=workers, shuffle=False, collate_fn=collate_fn)
return data_loader, validation_loader
elif state == 1 or state == 2 or state == 3:
image_sets = ['valfast', 'test', 'val']
if method == 'lstr':
if dataset == 'tusimple':
data_set = TuSimple(root=base, image_set=image_sets[state - 1], transforms=transforms_test,
padding_mask=False, process_points=False)
elif dataset == 'culane':
data_set = CULane(root=base, image_set=image_sets[state - 1], transforms=transforms_test,
padding_mask=False, process_points=False)
elif dataset == 'llamas':
data_set = LLAMAS(root=base, image_set=image_sets[state - 1], transforms=transforms_test,
padding_mask=False, process_points=False)
else:
raise ValueError
else:
data_set = StandardLaneDetectionDataset(root=base, image_set=image_sets[state - 1],
transforms=transforms_test, data_set=dataset)
data_loader = torch.utils.data.DataLoader(dataset=data_set, batch_size=batch_size, collate_fn=collate_fn,
num_workers=workers, shuffle=False)
return data_loader
else:
raise ValueError
adding_noise = True
gaussian_std = 0.05
n_epochs = 500
report_each = 1200
## Configuring data loader
# fold = 0
fold = 4
train_file_names, val_file_names = get_split(fold=fold)
# color image in range [0, 255] will become [-1, 1]
train_transform = DualCompose([
Resize(size=img_size),
HorizontalFlip(),
VerticalFlip(),
Normalize(normalize_mask=False),
MaskOnly([MaskErodeDilation(kernel_size_lower=0, kernel_size_upper=6)])
])
valid_transform = DualCompose(
[Resize(size=img_size),
Normalize(normalize_mask=False)])
batch_size = 6
num_workers = 4
train_loader = DataLoader(dataset=RoboticsDataset(mode='train',
file_names=train_file_names,
transform=train_transform,
problem_type='binary'),
shuffle=True,
num_workers=num_workers,
from dataset import RoboticsDataset
import cv2
from models import UNet16, LinkNet34, UNet11, UNet
import torch
from pathlib import Path
from tqdm import tqdm
import numpy as np
import utils
import prepare_data
from torch.utils.data import DataLoader
from torch.nn import functional as F
from prepare_data import (original_height, original_width, h_start, w_start)
from transforms import (ImageOnly, Normalize, DualCompose)
img_transform = DualCompose([ImageOnly(Normalize())])
def get_model(model_path, model_type='unet11', problem_type='binary'):
"""
:param model_path:
:param model_type: 'UNet', 'UNet16', 'UNet11', 'LinkNet34'
:param problem_type: 'binary', 'parts', 'instruments'
:return:
"""
if problem_type == 'binary':
num_classes = 1
elif problem_type == 'parts':
num_classes = 4
elif problem_type == 'instruments':
num_classes = 8
def main():
parser = argparse.ArgumentParser()
arg = parser.add_argument
arg('--jaccard-weight', type=float, default=1)
arg('--root', type=str, default='runs/debug', help='checkpoint root')
arg('--image-path', type=str, default='data', help='image path')
arg('--batch-size', type=int, default=2)
arg('--n-epochs', type=int, default=100)
arg('--optimizer', type=str, default='Adam', help='Adam or SGD')
arg('--lr', type=float, default=0.001)
arg('--workers', type=int, default=10)
arg('--model',
type=str,
default='UNet16',
choices=[
'UNet', 'UNet11', 'UNet16', 'LinkNet34', 'FCDenseNet57',
'FCDenseNet67', 'FCDenseNet103'
])
arg('--model-weight', type=str, default=None)
arg('--resume-path', type=str, default=None)
arg('--attribute',
type=str,
default='all',
choices=[
'pigment_network', 'negative_network', 'streaks',
'milia_like_cyst', 'globules', 'all'
])
args = parser.parse_args()
## folder for checkpoint
root = Path(args.root)
root.mkdir(exist_ok=True, parents=True)
image_path = args.image_path
#print(args)
if args.attribute == 'all':
num_classes = 5
else:
num_classes = 1
args.num_classes = num_classes
### save initial parameters
print('--' * 10)
print(args)
print('--' * 10)
root.joinpath('params.json').write_text(
json.dumps(vars(args), indent=True, sort_keys=True))
## load pretrained model
if args.model == 'UNet':
model = UNet(num_classes=num_classes)
elif args.model == 'UNet11':
model = UNet11(num_classes=num_classes, pretrained='vgg')
elif args.model == 'UNet16':
model = UNet16(num_classes=num_classes, pretrained='vgg')
elif args.model == 'LinkNet34':
model = LinkNet34(num_classes=num_classes, pretrained=True)
elif args.model == 'FCDenseNet103':
model = FCDenseNet103(num_classes=num_classes)
else:
model = UNet(num_classes=num_classes, input_channels=3)
## multiple GPUs
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
model.to(device)
## load pretrained model
if args.model_weight is not None:
state = torch.load(args.model_weight)
#epoch = state['epoch']
#step = state['step']
model.load_state_dict(state['model'])
print('--' * 10)
print('Load pretrained model', args.model_weight)
#print('Restored model, epoch {}, step {:,}'.format(epoch, step))
print('--' * 10)
## replace the last layer
## although the model and pre-trained weight have differernt size (the last layer is different)
## pytorch can still load the weight
## I found that the weight for one layer just duplicated for all layers
## therefore, the following code is not necessary
# if args.attribute == 'all':
# model = list(model.children())[0]
# num_filters = 32
# model.final = nn.Conv2d(num_filters, num_classes, kernel_size=1)
# print('--' * 10)
# print('Load pretrained model and replace the last layer', args.model_weight, num_classes)
# print('--' * 10)
# if torch.cuda.device_count() > 1:
# model = nn.DataParallel(model)
# model.to(device)
## model summary
print_model_summay(model)
## define loss
loss_fn = LossBinary(jaccard_weight=args.jaccard_weight)
## It enables benchmark mode in cudnn.
## benchmark mode is good whenever your input sizes for your network do not vary. This way, cudnn will look for the
## optimal set of algorithms for that particular configuration (which takes some time). This usually leads to faster runtime.
## But if your input sizes changes at each iteration, then cudnn will benchmark every time a new size appears,
## possibly leading to worse runtime performances.
cudnn.benchmark = True
## get train_test_id
train_test_id = get_split()
## train vs. val
print('--' * 10)
print('num train = {}, num_val = {}'.format(
(train_test_id['Split'] == 'train').sum(),
(train_test_id['Split'] != 'train').sum()))
print('--' * 10)
train_transform = DualCompose(
[HorizontalFlip(),
VerticalFlip(),
ImageOnly(Normalize())])
val_transform = DualCompose([ImageOnly(Normalize())])
## define data loader
train_loader = make_loader(train_test_id,
image_path,
args,
train=True,
shuffle=True,
transform=train_transform)
valid_loader = make_loader(train_test_id,
image_path,
args,
train=False,
shuffle=True,
transform=val_transform)
if True:
print('--' * 10)
print('check data')
train_image, train_mask, train_mask_ind = next(iter(train_loader))
print('train_image.shape', train_image.shape)
print('train_mask.shape', train_mask.shape)
print('train_mask_ind.shape', train_mask_ind.shape)
print('train_image.min', train_image.min().item())
print('train_image.max', train_image.max().item())
print('train_mask.min', train_mask.min().item())
print('train_mask.max', train_mask.max().item())
print('train_mask_ind.min', train_mask_ind.min().item())
print('train_mask_ind.max', train_mask_ind.max().item())
print('--' * 10)
valid_fn = validation_binary
###########
## optimizer
if args.optimizer == 'Adam':
optimizer = Adam(model.parameters(), lr=args.lr)
elif args.optimizer == 'SGD':
optimizer = SGD(model.parameters(), lr=args.lr, momentum=0.9)
## loss
criterion = loss_fn
## change LR
scheduler = ReduceLROnPlateau(optimizer,
'min',
factor=0.8,
patience=5,
verbose=True)
##########
## load previous model status
previous_valid_loss = 10
model_path = root / 'model.pt'
if args.resume_path is not None and model_path.exists():
state = torch.load(str(model_path))
epoch = state['epoch']
step = state['step']
model.load_state_dict(state['model'])
epoch = 1
step = 0
try:
previous_valid_loss = state['valid_loss']
except:
previous_valid_loss = 10
print('--' * 10)
print('Restored previous model, epoch {}, step {:,}'.format(
epoch, step))
print('--' * 10)
else:
epoch = 1
step = 0
#########
## start training
log = root.joinpath('train.log').open('at', encoding='utf8')
writer = SummaryWriter()
meter = AllInOneMeter()
#if previous_valid_loss = 10000
print('Start training')
print_model_summay(model)
previous_valid_jaccard = 0
for epoch in range(epoch, args.n_epochs + 1):
model.train()
random.seed()
#jaccard = []
start_time = time.time()
meter.reset()
w1 = 1.0
w2 = 0.5
w3 = 0.5
try:
train_loss = 0
valid_loss = 0
# if epoch == 1:
# freeze_layer_names = get_freeze_layer_names(part='encoder')
# set_freeze_layers(model, freeze_layer_names=freeze_layer_names)
# #set_train_layers(model, train_layer_names=['module.final.weight','module.final.bias'])
# print_model_summay(model)
# elif epoch == 5:
# w1 = 1.0
# w2 = 0.0
# w3 = 0.5
# freeze_layer_names = get_freeze_layer_names(part='encoder')
# set_freeze_layers(model, freeze_layer_names=freeze_layer_names)
# # set_train_layers(model, train_layer_names=['module.final.weight','module.final.bias'])
# print_model_summay(model)
#elif epoch == 3:
# set_train_layers(model, train_layer_names=['module.dec5.block.0.conv.weight','module.dec5.block.0.conv.bias',
# 'module.dec5.block.1.weight','module.dec5.block.1.bias',
# 'module.dec4.block.0.conv.weight','module.dec4.block.0.conv.bias',
# 'module.dec4.block.1.weight','module.dec4.block.1.bias',
# 'module.dec3.block.0.conv.weight','module.dec3.block.0.conv.bias',
# 'module.dec3.block.1.weight','module.dec3.block.1.bias',
# 'module.dec2.block.0.conv.weight','module.dec2.block.0.conv.bias',
# 'module.dec2.block.1.weight','module.dec2.block.1.bias',
# 'module.dec1.conv.weight','module.dec1.conv.bias',
# 'module.final.weight','module.final.bias'])
# print_model_summa zvgf t5y(model)
# elif epoch == 50:
# set_freeze_layers(model, freeze_layer_names=None)
# print_model_summay(model)
for i, (train_image, train_mask,
train_mask_ind) in enumerate(train_loader):
# inputs, targets = variable(inputs), variable(targets)
train_image = train_image.permute(0, 3, 1, 2)
train_mask = train_mask.permute(0, 3, 1, 2)
train_image = train_image.to(device)
train_mask = train_mask.to(device).type(torch.cuda.FloatTensor)
train_mask_ind = train_mask_ind.to(device).type(
torch.cuda.FloatTensor)
# if args.problem_type == 'binary':
# train_mask = train_mask.to(device).type(torch.cuda.FloatTensor)
# else:
# #train_mask = train_mask.to(device).type(torch.cuda.LongTensor)
# train_mask = train_mask.to(device).type(torch.cuda.FloatTensor)
outputs, outputs_mask_ind1, outputs_mask_ind2 = model(
train_image)
#print(outputs.size())
#print(outputs_mask_ind1.size())
#print(outputs_mask_ind2.size())
### note that the last layer in the model is defined differently
# if args.problem_type == 'binary':
# train_prob = F.sigmoid(outputs)
# loss = criterion(outputs, train_mask)
# else:
# #train_prob = outputs
# train_prob = F.sigmoid(outputs)
# loss = torch.tensor(0).type(train_mask.type())
# for feat_inx in range(train_mask.shape[1]):
# loss += criterion(outputs, train_mask)
train_prob = F.sigmoid(outputs)
train_mask_ind_prob1 = F.sigmoid(outputs_mask_ind1)
train_mask_ind_prob2 = F.sigmoid(outputs_mask_ind2)
loss1 = criterion(outputs, train_mask)
#loss1 = F.binary_cross_entropy_with_logits(outputs, train_mask)
#loss2 = nn.BCEWithLogitsLoss()(outputs_mask_ind1, train_mask_ind)
#print(train_mask_ind.size())
#weight = torch.ones_like(train_mask_ind)
#weight[:, 0] = weight[:, 0] * 1
#weight[:, 1] = weight[:, 1] * 14
#weight[:, 2] = weight[:, 2] * 14
#weight[:, 3] = weight[:, 3] * 4
#weight[:, 4] = weight[:, 4] * 4
#weight = weight * train_mask_ind + 1
#weight = weight.to(device).type(torch.cuda.FloatTensor)
loss2 = F.binary_cross_entropy_with_logits(
outputs_mask_ind1, train_mask_ind)
loss3 = F.binary_cross_entropy_with_logits(
outputs_mask_ind2, train_mask_ind)
#loss3 = criterion(outputs_mask_ind2, train_mask_ind)
loss = loss1 * w1 + loss2 * w2 + loss3 * w3
#print(loss1.item(), loss2.item(), loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
#jaccard += [get_jaccard(train_mask, (train_prob > 0).float()).item()]
meter.add(train_prob, train_mask, train_mask_ind_prob1,
train_mask_ind_prob2, train_mask_ind, loss1.item(),
loss2.item(), loss3.item(), loss.item())
# print(train_mask.data.shape)
# print(train_mask.data.sum(dim=-2).shape)
# print(train_mask.data.sum(dim=-2).sum(dim=-1).shape)
# print(train_mask.data.sum(dim=-2).sum(dim=-1).sum(dim=0).shape)
# intersection = train_mask.data.sum(dim=-2).sum(dim=-1)
# print(intersection.shape)
# print(intersection.dtype)
# print(train_mask.data.shape[0])
#torch.zeros([2, 4], dtype=torch.float32)
#########################
## at the end of each epoch, evualte the metrics
epoch_time = time.time() - start_time
train_metrics = meter.value()
train_metrics['epoch_time'] = epoch_time
train_metrics['image'] = train_image.data
train_metrics['mask'] = train_mask.data
train_metrics['prob'] = train_prob.data
#train_jaccard = np.mean(jaccard)
#train_auc = str(round(mtr1.value()[0],2))+' '+str(round(mtr2.value()[0],2))+' '+str(round(mtr3.value()[0],2))+' '+str(round(mtr4.value()[0],2))+' '+str(round(mtr5.value()[0],2))
valid_metrics = valid_fn(model, criterion, valid_loader, device,
num_classes)
##############
## write events
write_event(log,
step,
epoch=epoch,
train_metrics=train_metrics,
valid_metrics=valid_metrics)
#save_weights(model, model_path, epoch + 1, step)
#########################
## tensorboard
write_tensorboard(writer,
model,
epoch,
train_metrics=train_metrics,
valid_metrics=valid_metrics)
#########################
## save the best model
valid_loss = valid_metrics['loss1']
valid_jaccard = valid_metrics['jaccard']
if valid_loss < previous_valid_loss:
save_weights(model, model_path, epoch + 1, step, train_metrics,
valid_metrics)
previous_valid_loss = valid_loss
print('Save best model by loss')
if valid_jaccard > previous_valid_jaccard:
save_weights(model, model_path, epoch + 1, step, train_metrics,
valid_metrics)
previous_valid_jaccard = valid_jaccard
print('Save best model by jaccard')
#########################
## change learning rate
scheduler.step(valid_metrics['loss1'])
except KeyboardInterrupt:
# print('--' * 10)
# print('Ctrl+C, saving snapshot')
# save_weights(model, model_path, epoch, step)
# print('done.')
# print('--' * 10)
writer.close()
#return
writer.close()
from tqdm import tqdm
import numpy as np
import utils
# import prepare_data
from torch.utils.data import DataLoader
from torch.nn import functional as F
import torch
from transforms import (ImageOnly,
Normalize,
CenterCrop,
DualCompose)
img_transform = DualCompose([
CenterCrop(512),
ImageOnly(Normalize())
])
def get_model(model_path, model_type):
"""
:param model_path:
:param model_type: 'UNet', 'UNet11', 'UNet16', 'AlbuNet34'
:return:
"""
num_classes = 1
if model_type == 'UNet11':
model = UNet11(num_classes=num_classes)
from prepare_train_val import get_split
from dataset import Polyp
import cv2
from models import UNet, UNet11, UNet16, AlbuNet34, MDeNet, EncDec, hourglass, MDeNetplus
import torch
from pathlib import Path
from tqdm import tqdm
import numpy as np
import utils
# import prepare_data
from torch.utils.data import DataLoader
from torch.nn import functional as F
from transforms import (ImageOnly, Normalize, CenterCrop, DualCompose)
img_transform = DualCompose([CenterCrop(512), ImageOnly(Normalize())])
def get_model(model_path, model_type):
"""
:param model_path:
:param model_type: 'UNet', 'UNet11', 'UNet16', 'AlbuNet34'
:return:
"""
num_classes = 1
if model_type == 'UNet11':
model = UNet11(num_classes=num_classes)
elif model_type == 'UNet16':
def main():
parser = argparse.ArgumentParser()
arg = parser.add_argument
arg('--jaccard-weight', default=0.3, type=float)
arg('--device-ids',
type=str,
default='0',
help='For example 0,1 to run on two GPUs')
arg('--fold', type=int, help='fold', default=0)
arg('--root', default='runs/debug', help='checkpoint root')
arg('--batch-size', type=int, default=1)
arg('--limit', type=int, default=10000, help='number of images in epoch')
arg('--n-epochs', type=int, default=100)
arg('--lr', type=float, default=0.0001)
arg('--workers', type=int, default=12)
arg('--multiple-output', type=bool, default=False)
arg('--model',
type=str,
default='UNet',
choices=[
'UNet', 'UNet11', 'LinkNet34', 'UNet16', 'AlbuNet34', 'SEAlbuNet',
'MultiSEAlbuNet'
])
args = parser.parse_args()
root = Path(args.root)
root.mkdir(exist_ok=True, parents=True)
num_classes = 1
if args.model == 'UNet':
model = UNet(num_classes=num_classes)
elif args.model == 'UNet11':
model = UNet11(num_classes=num_classes, pretrained=True)
elif args.model == 'UNet16':
model = UNet16(num_classes=num_classes, pretrained=True)
elif args.model == 'LinkNet34':
model = LinkNet34(num_classes=num_classes, pretrained=True)
elif args.model == 'AlbuNet':
model = AlbuNet34(num_classes=num_classes, pretrained=True)
elif args.model == 'SEAlbuNet':
model = SEAlbuNet34(num_classes=num_classes, pretrained=True)
elif args.model == 'MultiSEAlbuNet':
model = MultiSEAlbuNet34(num_classes=num_classes, pretrained=True)
else:
model = UNet(num_classes=num_classes, input_channels=3)
if torch.cuda.is_available():
if args.device_ids:
device_ids = list(map(int, args.device_ids.split(',')))
else:
device_ids = None
model = nn.DataParallel(model, device_ids=device_ids).cuda()
if args.multiple_output == False:
loss = LossBinary(jaccard_weight=args.jaccard_weight)
else:
loss = MultiLossBinary(jaccard_weight=args.jaccard_weight)
cudnn.benchmark = True
def make_loader(file_names, shuffle=False, transform=None, limit=None):
if args.multiple_output == False:
return DataLoader(dataset=AngyodysplasiaDataset(
file_names, transform=transform, limit=limit),
shuffle=shuffle,
num_workers=args.workers,
batch_size=args.batch_size,
pin_memory=torch.cuda.is_available())
else:
return DataLoader(dataset=MultiAngyodysplasiaDataset(
file_names, transform=transform, limit=limit),
shuffle=shuffle,
num_workers=args.workers,
batch_size=args.batch_size,
pin_memory=torch.cuda.is_available())
train_file_names, val_file_names = get_split(args.fold)
print('num train = {}, num_val = {}'.format(len(train_file_names),
len(val_file_names)))
train_transform = DualCompose([
CenterCrop(512),
HorizontalFlip(),
VerticalFlip(),
Rotate(),
ImageOnly(RandomHueSaturationValue()),
ImageOnly(Normalize())
])
val_transform = DualCompose([CenterCrop(512), ImageOnly(Normalize())])
train_loader = make_loader(train_file_names,
shuffle=True,
transform=train_transform,
limit=args.limit)
valid_loader = make_loader(val_file_names, transform=val_transform)
root.joinpath('params.json').write_text(
json.dumps(vars(args), indent=True, sort_keys=True))
if args.multiple_output == False:
utils.train(init_optimizer=lambda lr: Adam(model.parameters(), lr=lr),
args=args,
model=model,
criterion=loss,
train_loader=train_loader,
valid_loader=valid_loader,
validation=validation_binary,
fold=args.fold)
else:
utils.train_multi(
init_optimizer=lambda lr: Adam(model.parameters(), lr=lr),
args=args,
model=model,
criterion=loss,
train_loader=train_loader,
valid_loader=valid_loader,
validation=multi_validation_binary,
fold=args.fold)
def init(batch_size, state, input_sizes, std, mean, dataset, city_aug=0):
# Return data_loaders
# depending on whether the state is
# 1: training
# 2: just testing
# Transformations
# ! Can't use torchvision.Transforms.Compose
if dataset == 'voc':
base = base_voc
workers = 4
transform_train = Compose([
ToTensor(),
RandomResize(min_size=input_sizes[0], max_size=input_sizes[1]),
RandomCrop(size=input_sizes[0]),
RandomHorizontalFlip(flip_prob=0.5),
Normalize(mean=mean, std=std)
])
transform_test = Compose([
ToTensor(),
ZeroPad(size=input_sizes[2]),
Normalize(mean=mean, std=std)
])
elif dataset == 'city' or dataset == 'gtav' or dataset == 'synthia': # All the same size
if dataset == 'city':
base = base_city
elif dataset == 'gtav':
base = base_gtav
else:
base = base_synthia
outlier = False if dataset == 'city' else True # GTAV has f****d up label ID
workers = 8
if city_aug == 3: # SYNTHIA & GTAV
if dataset == 'gtav':
transform_train = Compose([
ToTensor(),
Resize(size_label=input_sizes[1],
size_image=input_sizes[1]),
RandomCrop(size=input_sizes[0]),
RandomHorizontalFlip(flip_prob=0.5),
Normalize(mean=mean, std=std),
LabelMap(label_id_map_city, outlier=outlier)
])
else:
transform_train = Compose([
ToTensor(),
RandomCrop(size=input_sizes[0]),
RandomHorizontalFlip(flip_prob=0.5),
Normalize(mean=mean, std=std),
LabelMap(label_id_map_synthia, outlier=outlier)
])
transform_test = Compose([
ToTensor(),
Resize(size_image=input_sizes[2], size_label=input_sizes[2]),
Normalize(mean=mean, std=std),
LabelMap(label_id_map_city)
])
elif city_aug == 2: # ERFNet
transform_train = Compose([
ToTensor(),
Resize(size_image=input_sizes[0], size_label=input_sizes[0]),
LabelMap(label_id_map_city, outlier=outlier),
RandomTranslation(trans_h=2, trans_w=2),
RandomHorizontalFlip(flip_prob=0.5)
])
transform_test = Compose([
ToTensor(),
Resize(size_image=input_sizes[0], size_label=input_sizes[2]),
LabelMap(label_id_map_city)
])
elif city_aug == 1: # City big
transform_train = Compose([
ToTensor(),
RandomCrop(size=input_sizes[0]),
LabelMap(label_id_map_city, outlier=outlier),
RandomTranslation(trans_h=2, trans_w=2),
RandomHorizontalFlip(flip_prob=0.5),
Normalize(mean=mean, std=std)
])
transform_test = Compose([
ToTensor(),
Resize(size_image=input_sizes[2], size_label=input_sizes[2]),
Normalize(mean=mean, std=std),
LabelMap(label_id_map_city)
])
else: # Standard city
transform_train = Compose([
ToTensor(),
RandomResize(min_size=input_sizes[0], max_size=input_sizes[1]),
RandomCrop(size=input_sizes[0]),
RandomHorizontalFlip(flip_prob=0.5),
Normalize(mean=mean, std=std),
LabelMap(label_id_map_city, outlier=outlier)
])
transform_test = Compose([
ToTensor(),
Resize(size_image=input_sizes[2], size_label=input_sizes[2]),
Normalize(mean=mean, std=std),
LabelMap(label_id_map_city)
])
else:
raise ValueError
# Not the actual test set (i.e. validation set)
test_set = StandardSegmentationDataset(
root=base_city if dataset == 'gtav' or dataset == 'synthia' else base,
image_set='val',
transforms=transform_test,
data_set='city'
if dataset == 'gtav' or dataset == 'synthia' else dataset)
if city_aug == 1:
val_loader = torch.utils.data.DataLoader(dataset=test_set,
batch_size=1,
num_workers=workers,
shuffle=False)
else:
val_loader = torch.utils.data.DataLoader(dataset=test_set,
batch_size=batch_size,
num_workers=workers,
shuffle=False)
# Testing
if state == 1:
return val_loader
else:
# Training
train_set = StandardSegmentationDataset(
root=base,
image_set='trainaug' if dataset == 'voc' else 'train',
transforms=transform_train,
data_set=dataset)
train_loader = torch.utils.data.DataLoader(dataset=train_set,
batch_size=batch_size,
num_workers=workers,
shuffle=True)
return train_loader, val_loader
def _init_fn(self):
# transform_list = [Select(['image', 'bb', 'keypoints'])]
# transform_list.append(RandomRescaleBB(0.9, 1.4))
# transform_list.append(RandomFlipLR())
transform_list = []
transform_list.append(RandomBlur())
#transform_list.append(RandomGrayscale())
transform_list.append(
ColorJitter(brightness=self.options.jitter,
contrast=self.options.jitter,
saturation=self.options.jitter,
hue=self.options.jitter / 4))
if self.options.degrees > 0:
transform_list.append(RandomRotation(degrees=self.options.degrees))
if self.options.max_scale > 1:
transform_list.append(RandomRescaleBB(1.0, self.options.max_scale))
transform_list.append(
CropAndResize(out_size=(self.options.crop_size,
self.options.crop_size)))
transform_list.append(
LocsToHeatmaps(out_size=(self.options.heatmap_size,
self.options.heatmap_size)))
transform_list.append(ToTensor())
transform_list.append(Normalize())
test_transform_list = []
if self.options.max_scale > 1:
test_transform_list.append(
RandomRescaleBB(1.0, self.options.max_scale))
test_transform_list.append(
CropAndResize(out_size=(self.options.crop_size,
self.options.crop_size)))
test_transform_list.append(
LocsToHeatmaps(out_size=(self.options.heatmap_size,
self.options.heatmap_size)))
test_transform_list.append(ToTensor())
test_transform_list.append(Normalize())
self.train_ds = RctaDataset(
root_dir=self.options.dataset_dir,
is_train=True,
transform=transforms.Compose(transform_list))
# print("Keypoints in trainer:", self.train_ds.keypoints[74*6])
# print("Bounding boxes:", self.train_ds.bounding_boxes[74*6])
self.test_ds = RctaDataset(
root_dir=self.options.dataset_dir,
is_train=False,
transform=transforms.Compose(test_transform_list))
self.model = StackedHourglass(self.options.num_keypoints).to(
self.device)
print('Total number of model parameters:',
self.model.num_trainable_parameters())
# create optimizer
# if self.options.optimizer == 'sgd':
# self.optimizer = torch.optim.SGD(params=self.model.parameters(), lr=self.options.lr, momentum=self.options.sgd_momentum, weight_decay=self.options.wd)
# elif self.options.optimizer == 'rmsprop':
self.optimizer = torch.optim.RMSprop(params=self.model.parameters(),
lr=self.options.lr,
momentum=0,
weight_decay=self.options.wd)
# else:
# self.optimizer = torch.optim.Adam(params=self.model.parameters(), lr=self.options.lr, betas=(self.options.adam_beta1, 0.999), weight_decay=self.options.wd)
# pack all models and optimizers in dictionaries to interact with the checkpoint saver
self.models_dict = {'stacked_hg': self.model}
self.optimizers_dict = {'optimizer': self.optimizer}
self.criterion = nn.MSELoss(size_average=True).to(self.device)
self.pose = Pose2DEval(detection_thresh=self.options.detection_thresh,
dist_thresh=self.options.dist_thresh)
def init(batch_size, state, split, input_sizes, sets_id, std, mean, keep_scale,
reverse_channels, data_set, valtiny, no_aug):
# Return data_loaders/data_loader
# depending on whether the state is
# 1: semi-supervised training
# 2: fully-supervised training
# 3: Just testing
# Transformations (compatible with unlabeled data/pseudo labeled data)
# ! Can't use torchvision.Transforms.Compose
if data_set == 'voc':
base = base_voc
workers = 4
transform_train = Compose([
ToTensor(keep_scale=keep_scale, reverse_channels=reverse_channels),
RandomResize(min_size=input_sizes[0], max_size=input_sizes[1]),
RandomCrop(size=input_sizes[0]),
RandomHorizontalFlip(flip_prob=0.5),
Normalize(mean=mean, std=std)
])
if no_aug:
transform_train_pseudo = Compose([
ToTensor(keep_scale=keep_scale,
reverse_channels=reverse_channels),
Resize(size_image=input_sizes[0], size_label=input_sizes[0]),
Normalize(mean=mean, std=std)
])
else:
transform_train_pseudo = Compose([
ToTensor(keep_scale=keep_scale,
reverse_channels=reverse_channels),
RandomResize(min_size=input_sizes[0], max_size=input_sizes[1]),
RandomCrop(size=input_sizes[0]),
RandomHorizontalFlip(flip_prob=0.5),
Normalize(mean=mean, std=std)
])
transform_pseudo = Compose([
ToTensor(keep_scale=keep_scale, reverse_channels=reverse_channels),
Resize(size_image=input_sizes[0], size_label=input_sizes[0]),
Normalize(mean=mean, std=std)
])
transform_test = Compose([
ToTensor(keep_scale=keep_scale, reverse_channels=reverse_channels),
ZeroPad(size=input_sizes[2]),
Normalize(mean=mean, std=std)
])
elif data_set == 'city': # All the same size (whole set is down-sampled by 2)
base = base_city
workers = 8
transform_train = Compose([
ToTensor(keep_scale=keep_scale, reverse_channels=reverse_channels),
RandomResize(min_size=input_sizes[0], max_size=input_sizes[1]),
RandomCrop(size=input_sizes[0]),
RandomHorizontalFlip(flip_prob=0.5),
Normalize(mean=mean, std=std),
LabelMap(label_id_map_city)
])
if no_aug:
transform_train_pseudo = Compose([
ToTensor(keep_scale=keep_scale,
reverse_channels=reverse_channels),
Resize(size_image=input_sizes[0], size_label=input_sizes[0]),
Normalize(mean=mean, std=std)
])
else:
transform_train_pseudo = Compose([
ToTensor(keep_scale=keep_scale,
reverse_channels=reverse_channels),
RandomResize(min_size=input_sizes[0], max_size=input_sizes[1]),
RandomCrop(size=input_sizes[0]),
RandomHorizontalFlip(flip_prob=0.5),
Normalize(mean=mean, std=std)
])
transform_pseudo = Compose([
ToTensor(keep_scale=keep_scale, reverse_channels=reverse_channels),
Resize(size_image=input_sizes[0], size_label=input_sizes[0]),
Normalize(mean=mean, std=std),
LabelMap(label_id_map_city)
])
transform_test = Compose([
ToTensor(keep_scale=keep_scale, reverse_channels=reverse_channels),
Resize(size_image=input_sizes[2], size_label=input_sizes[2]),
Normalize(mean=mean, std=std),
LabelMap(label_id_map_city)
])
else:
base = ''
# Not the actual test set (i.e.validation set)
test_set = StandardSegmentationDataset(
root=base,
image_set='valtiny' if valtiny else 'val',
transforms=transform_test,
label_state=0,
data_set=data_set)
val_loader = torch.utils.data.DataLoader(dataset=test_set,
batch_size=batch_size,
num_workers=workers,
shuffle=False)
# Testing
if state == 3:
return val_loader
else:
# Fully-supervised training
if state == 2:
labeled_set = StandardSegmentationDataset(
root=base,
image_set=(str(split) + '_labeled_' + str(sets_id)),
transforms=transform_train,
label_state=0,
data_set=data_set)
labeled_loader = torch.utils.data.DataLoader(dataset=labeled_set,
batch_size=batch_size,
num_workers=workers,
shuffle=True)
return labeled_loader, val_loader
# Semi-supervised training
elif state == 1:
pseudo_labeled_set = StandardSegmentationDataset(
root=base,
data_set=data_set,
image_set=(str(split) + '_unlabeled_' + str(sets_id)),
transforms=transform_train_pseudo,
label_state=1)
reference_set = SegmentationLabelsDataset(
root=base,
image_set=(str(split) + '_unlabeled_' + str(sets_id)),
data_set=data_set)
reference_loader = torch.utils.data.DataLoader(
dataset=reference_set,
batch_size=batch_size,
num_workers=workers,
shuffle=False)
unlabeled_set = StandardSegmentationDataset(
root=base,
data_set=data_set,
image_set=(str(split) + '_unlabeled_' + str(sets_id)),
transforms=transform_pseudo,
label_state=2)
labeled_set = StandardSegmentationDataset(
root=base,
data_set=data_set,
image_set=(str(split) + '_labeled_' + str(sets_id)),
transforms=transform_train,
label_state=0)
unlabeled_loader = torch.utils.data.DataLoader(
dataset=unlabeled_set,
batch_size=batch_size,
num_workers=workers,
shuffle=False)
pseudo_labeled_loader = torch.utils.data.DataLoader(
dataset=pseudo_labeled_set,
batch_size=int(batch_size / 2),
num_workers=workers,
shuffle=True)
labeled_loader = torch.utils.data.DataLoader(dataset=labeled_set,
batch_size=int(
batch_size / 2),
num_workers=workers,
shuffle=True)
return labeled_loader, pseudo_labeled_loader, unlabeled_loader, val_loader, reference_loader
else:
# Support unsupervised learning here if that's what you want
raise ValueError
def main():
parser = argparse.ArgumentParser()
arg = parser.add_argument
arg('--jaccard-weight', default=0.3, type=float)
arg('--device-ids',
type=str,
default='0',
help='For example 0,1 to run on two GPUs')
arg('--fold', type=int, help='fold', default=0)
arg('--root', default='runs/debug', help='checkpoint root')
arg('--batch-size', type=int, default=1)
arg('--limit', type=int, default=10000, help='number of images in epoch')
arg('--n-epochs', type=int, default=100)
arg('--lr', type=float, default=0.0001)
arg('--workers', type=int, default=12)
arg("--b1",
type=float,
default=0.5,
help="adam: decay of first order momentum of gradient")
arg("--b2",
type=float,
default=0.999,
help="adam: decay of first order momentum of gradient")
args = parser.parse_args()
root = Path(args.root)
root.mkdir(exist_ok=True, parents=True)
# Loss functions
criterion_GAN = GAN_loss(gan_weight=1) #torch.nn.MSELoss()
criterion_pixelwise = torch.nn.L1Loss()
criterion_discrim = Discrim_loss(dircrim_weight=1)
# Loss weight of L1 pixel-wise loss between translated image and real image
lambda_pixel = 100
# Initialize generator and discriminator
model = AlbuNet34(num_classes=1, pretrained=True)
discrim_model = discriminator()
if torch.cuda.is_available():
if args.device_ids:
device_ids = list(map(int, args.device_ids.split(',')))
else:
device_ids = None
model = nn.DataParallel(model, device_ids=device_ids).cuda()
discrim_model = nn.DataParallel(discrim_model,
device_ids=device_ids).cuda()
# Load pretrained models
root = Path(args.root)
model_path = root / 'model_{fold}.pt'.format(fold=args.fold)
if model_path.exists():
state = torch.load(str(model_path))
epoch = state['epoch']
step = state['step']
model.load_state_dict(state['model'])
print('Restored model, epoch {}, step {:,}'.format(epoch, step))
else:
epoch = 1
step = 0
save = lambda ep: torch.save(
{
'model': model.state_dict(),
'epoch': ep,
'step': step,
}, str(model_path))
# Optimizers
optimizer_G = Adam(model.parameters(),
lr=args.lr,
betas=(args.b1, args.b2))
optimizer_D = Adam(discrim_model.parameters(),
lr=args.lr,
betas=(args.b1, args.b2))
# Configure dataloaders
def make_loader(file_names, shuffle=False, transform=None, limit=None):
return DataLoader(dataset=Polyp(file_names,
transform=transform,
limit=limit),
shuffle=shuffle,
num_workers=args.workers,
batch_size=args.batch_size,
pin_memory=torch.cuda.is_available())
train_file_names, val_file_names = get_split(args.fold)
print('num train = {}, num_val = {}'.format(len(train_file_names),
len(val_file_names)))
train_transform = DualCompose([
CropCVC612(),
img_resize(512),
HorizontalFlip(),
VerticalFlip(),
Rotate(),
Rescale(),
Zoomin(),
ImageOnly(RandomHueSaturationValue()),
ImageOnly(Normalize())
])
train_loader = make_loader(train_file_names,
shuffle=True,
transform=train_transform,
limit=args.limit)
root.joinpath('params.json').write_text(
json.dumps(vars(args), indent=True, sort_keys=True))
report_each = 10
log = root.joinpath('train_{fold}.log'.format(fold=args.fold)).open(
'at', encoding='utf8')
for epoch in range(epoch, args.n_epochs + 1):
model.train()
discrim_model.train()
random.seed()
tq = tqdm.tqdm(total=(len(train_loader) * args.batch_size))
tq.set_description('Epoch {}, lr {}'.format(epoch, args.lr))
losses = []
tl = train_loader
try:
mean_loss = 0
for i, (inputs, targets) in enumerate(tl):
# Model inputs
inputs, targets = variable(inputs), variable(targets)
# ------------------
# Train Generators
# ------------------
optimizer_G.zero_grad()
# Generate output
outputs = model(inputs)
# fake loss
predict_fake = discrim_model(inputs, outputs)
# Pixel-wise loss
loss_pixel = criterion_pixelwise(outputs, targets)
# Generator loss
loss_GAN = criterion_GAN(predict_fake)
# Total loss of GAN
loss_G = loss_GAN + lambda_pixel * loss_pixel
loss_G.backward()
optimizer_G.step()
# ---------------------
# Train Discriminator
# ---------------------
optimizer_D.zero_grad()
# Real loss
predict_real = discrim_model(inputs, targets)
predict_fake = discrim_model(inputs, outputs.detach())
# Discriminator loss
loss_D = criterion_discrim(predict_real, predict_fake)
loss_D.backward()
optimizer_D.step()
step += 1
batch_size = inputs.size(0)
tq.update(batch_size)
losses.append(float(loss_G.data))
mean_loss = np.mean(losses[-report_each:])
tq.set_postfix(loss='{:.5f}'.format(mean_loss))
if i and i % report_each == 0:
write_event(log, step, loss=mean_loss)
write_event(log, step, loss=mean_loss)
tq.close()
save(epoch + 1)
except KeyboardInterrupt:
tq.close()
print('Ctrl+C, saving snapshot')
save(epoch)
print('done.')
return
def main():
parser = argparse.ArgumentParser()
arg = parser.add_argument
arg('--jaccard-weight', default=1, type=float)
arg('--device-ids',
type=str,
default='0',
help='For example 0,1 to run on two GPUs')
arg('--fold', type=int, help='fold', default=0)
arg('--root', default='runs/debug', help='checkpoint root')
arg('--batch-size', type=int, default=8)
arg('--n-epochs', type=int, default=14)
arg('--lr', type=float, default=0.000001)
arg('--workers', type=int, default=8)
arg('--type',
type=str,
default='binary',
choices=['binary', 'parts', 'instruments'])
arg('--model',
type=str,
default='TernausNet',
choices=['UNet', 'UNet11', 'LinkNet34', 'TernausNet'])
args = parser.parse_args()
root = Path(args.root)
root.mkdir(exist_ok=True, parents=True)
if args.type == 'parts':
num_classes = 3
elif args.type == 'instruments':
num_classes = 8
else:
num_classes = 1
if args.model == 'TernausNet':
model = TernausNet34(num_classes=num_classes)
else:
model = TernausNet34(num_classes=num_classes)
if torch.cuda.is_available():
if args.device_ids:
device_ids = list(map(int, args.device_ids.split(',')))
else:
device_ids = None
model = nn.DataParallel(model, device_ids=device_ids).cuda()
if args.type == 'binary':
loss = LossBinary(jaccard_weight=args.jaccard_weight)
else:
loss = LossMulti(num_classes=num_classes,
jaccard_weight=args.jaccard_weight)
cudnn.benchmark = True
def make_loader(file_names,
shuffle=False,
transform=None,
mode='train',
problem_type='binary'):
return DataLoader(dataset=MapDataset(file_names,
transform=transform,
problem_type=problem_type,
mode=mode),
shuffle=shuffle,
num_workers=args.workers,
batch_size=args.batch_size,
pin_memory=torch.cuda.is_available())
# labels = pd.read_csv('data/stage1_train_labels.csv')
# labels = os.listdir('data/stage1_train_')
# train_file_names, val_file_names = train_test_split(labels, test_size=0.2, random_state=42)
# print('num train = {}, num_val = {}'.format(len(train_file_names), len(val_file_names)))
# train_transform = DualCompose([
# HorizontalFlip(),
# VerticalFlip(),
# RandomCrop([256, 256]),
# RandomRotate90(),
# ShiftScaleRotate(),
# ImageOnly(RandomHueSaturationValue()),
# ImageOnly(RandomBrightness()),
# ImageOnly(RandomContrast()),
# ImageOnly(Normalize())
# ])
train_transform = DualCompose([
OneOrOther(*(OneOf([
Distort1(distort_limit=0.05, shift_limit=0.05),
Distort2(num_steps=2, distort_limit=0.05)
]),
ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.10,
rotate_limit=45)),
prob=0.5),
RandomRotate90(),
RandomCrop([256, 256]),
RandomFlip(prob=0.5),
Transpose(prob=0.5),
ImageOnly(RandomContrast(limit=0.2, prob=0.5)),
ImageOnly(RandomFilter(limit=0.5, prob=0.2)),
ImageOnly(RandomHueSaturationValue(prob=0.2)),
ImageOnly(RandomBrightness()),
ImageOnly(Normalize())
])
val_transform = DualCompose([
# RandomCrop([256, 256]),
Rescale([256, 256]),
ImageOnly(Normalize())
])
train_loader = make_loader(TRAIN_ANNOTATIONS_PATH,
shuffle=True,
transform=train_transform,
problem_type=args.type)
valid_loader = make_loader(VAL_ANNOTATIONS_PATH,
transform=val_transform,
mode='valid',
problem_type=args.type)
root.joinpath('params.json').write_text(
json.dumps(vars(args), indent=True, sort_keys=True))
if args.type == 'binary':
valid = validation_binary
else:
valid = validation_multi
utils.train(init_optimizer=lambda lr: Adam(model.parameters(), lr=lr),
args=args,
model=model,
criterion=loss,
train_loader=train_loader,
valid_loader=valid_loader,
validation=valid,
fold=args.fold,
num_classes=num_classes)
from criterion import CrossEntropyLoss2d
from transform import Relabel, ToLabel, Colorize
import deeplab_resnet
import torch.nn.functional as F
from accuracy_metrics import pixel_accuracy, mean_accuracy, mean_IU
NUM_CHANNELS = 3
NUM_CLASSES = 22 #6 for brats
color_transform = Colorize()
image_transform = ToPILImage()
input_transform = Compose([
CenterCrop(256),
#Scale(240),
ToTensor(),
Normalize([.485, .456, .406], [.229, .224, .225]),
])
input_transform1 = Compose([
CenterCrop(256),
#Scale(136),
#ToTensor(),
])
target_transform = Compose([
CenterCrop(256),
#Scale(240),
ToLabel(),
#Relabel(255, NUM_CLASSES-1),
])