def initialize(args):
if args.device is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = args.device
model = select_model(args)
optimizer = select_optimizer(args, model)
if (args.cuda):
model.cuda()
train_params = {
'batch_size': args.batch_size,
'shuffle': True,
'num_workers': 2
}
test_params = {
'batch_size': args.batch_size,
'shuffle': False,
'num_workers': 1
}
train_loader = COVIDxDataset(mode='train',
n_classes=args.classes,
dataset_path=args.dataset,
dim=(224, 224))
val_loader = COVIDxDataset(mode='test',
n_classes=args.classes,
dataset_path=args.dataset,
dim=(224, 224))
training_generator = DataLoader(train_loader, **train_params)
val_generator = DataLoader(val_loader, **test_params)
return model, optimizer, training_generator, val_generator
def objective(trial):
info = nvmlDeviceGetMemoryInfo(handle)
print("Total memory:", info.total)
print("Free memory:", info.free)
print("Used memory:", info.used)
model, training_generator, val_generator, test_generator = initialize(ARGS)
optim_name = trial.suggest_categorical("optimizer",
["Adam", "RMSprop", "SGD"])
weight_decay = trial.suggest_float("weight_decay", 1e-5, 1e-1, log=True)
lr = trial.suggest_float("learning_rate", 1e-7, 1e-5, log=True)
trial.set_user_attr("worker_id", WORKER_ID)
optimizer = util.select_optimizer(optim_name, model, lr, weight_decay)
scheduler = ReduceLROnPlateau(optimizer,
factor=0.5,
patience=2,
min_lr=1e-5,
verbose=True)
best_pred_loss = 1000.0
for epoch in range(1, EPOCHS + 1):
train(ARGS, model, training_generator, optimizer, epoch)
val_metrics, confusion_matrix = validation(ARGS, model, val_generator,
epoch)
scheduler.step(val_metrics._data.average.loss)
return val_metrics._data.average.recall_mean
def initialize(self, args):
if args.device is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = args.device
model = select_model(args)
optimizer = select_optimizer(args.opt)
if (args.cuda):
model.cuda()
return model, optimizer
def initialize(args):
if args.device is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.device)
model = select_model(args)
optimizer = select_optimizer(args, model)
if (args.cuda):
model.cuda()
train_params = {
'batch_size': args.batch_size,
'shuffle': True,
'num_workers': 2
}
test_params = {
'batch_size': args.batch_size,
'shuffle': False,
'num_workers': 1
}
if args.dataset_name == 'COVIDx':
train_loader = COVIDxDataset(mode='train',
n_classes=args.classes,
dataset_path=args.dataset_name,
dim=(224, 224))
val_loader = COVIDxDataset(mode='test',
n_classes=args.classes,
dataset_path=args.dataset_name,
dim=(224, 224))
test_loader = None
training_generator = DataLoader(train_loader, **train_params)
val_generator = DataLoader(val_loader, **test_params)
test_generator = None
elif args.dataset_name == 'COVID_CT':
train_loader = CovidCTDataset(
'train',
root_dir='./data/covid_ct_dataset',
txt_COVID='./data/covid_ct_dataset/trainCT_COVID.txt',
txt_NonCOVID='./data/covid_ct_dataset/trainCT_NonCOVID.txt')
val_loader = CovidCTDataset(
'val',
root_dir='./data/covid_ct_dataset',
txt_COVID='./data/covid_ct_dataset/valCT_COVID.txt',
txt_NonCOVID='./data/covid_ct_dataset/valCT_NonCOVID.txt')
test_loader = CovidCTDataset(
'test',
root_dir='./data/covid_ct_dataset',
txt_COVID='./data/covid_ct_dataset/testCT_COVID.txt',
txt_NonCOVID='./data/covid_ct_dataset/testCT_NonCOVID.txt')
training_generator = DataLoader(train_loader, **train_params)
val_generator = DataLoader(val_loader, **test_params)
test_generator = DataLoader(test_loader, **test_params)
return model, optimizer, training_generator, val_generator, test_generator
def initialize(args):
if args.device is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.device)
model = select_model(args)
optimizer = select_optimizer(args, model)
if (args.cuda):
model.cuda()
train_loader = COVIDxDataset(mode='train',
n_classes=args.classes,
dataset_path=args.dataset,
dim=(224, 224))
#print(train_loader.)
#------ Class weigths for sampling and for loss function -----------------------------------
labels = np.unique(train_loader.labels)
print(labels)
class_weight = compute_class_weight('balanced', labels,
train_loader.labels)
#weights_sample = compute_sample_weight('balanced',train_loader.labels)
#print(np.unique(weights_sample))
#---------- Alphabetical order in labels does not correspond to class order in COVIDxDataset-----
class_weight = class_weight[::-1]
#---------------------------------------------------------------------------------
#weights_sample = torch.DoubleTensor(weights_sample)
#sampler = torch.utils.data.sampler.WeightedRandomSampler(weights_sample, len(weights_sample))
if (args.cuda):
class_weight = torch.from_numpy(class_weight.astype(float)).cuda()
else:
class_weight = torch.from_numpy(class_weight.astype(float))
#print(class_weight.shape)
#-------------------------------------------
val_loader = COVIDxDataset(mode='test',
n_classes=args.classes,
dataset_path=args.dataset,
dim=(224, 224))
#------------------------------------------------------------------------------------
train_params = {
'batch_size': args.batch_size,
'shuffle': True,
'num_workers': 4
} #'sampler' : sampler
test_params = {
'batch_size': args.batch_size,
'shuffle': True,
'num_workers': 4
}
#------------------------------------------------------------------------------------------
training_generator = DataLoader(train_loader, **train_params)
val_generator = DataLoader(val_loader, **test_params)
return model, optimizer, training_generator, val_generator, class_weight
def main():
plt.rcParams['figure.figsize'] = (10.0, 8.0) # set default size of plots
plt.rcParams['image.interpolation'] = 'nearest'
plt.rcParams['image.cmap'] = 'gray'
save_h5 = False
#save_h5 = True
args = get_arguments()
SEED = args.seed
torch.manual_seed(SEED)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
np.random.seed(SEED)
if torch.cuda.is_available():
torch.cuda.manual_seed(SEED)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# Set the path for the dataset
args.root_path = 'G:/datasets/covidx_v3_data'
weight_path = './save/COVIDNet20200520_0709/COVIDNet_large_best_checkpoint.pt'
if save_h5:
args.h5 = False
dataset, val_dataset, test_generator = initialize_datasets(
args, use_transform=False)
#dataset, val_dataset, test_generator = initialize_datasets(args, train_size=10, val_size=100, use_transform=False)
# Save of the datasets as h5 is desired
for ds in (dataset, val_dataset, test_generator):
if ds:
imgs = []
labels = []
for n, imgTen in enumerate(ds):
# Create an h5 version of the image
image = imgTen[0].numpy()
label = imgTen[1].numpy()
imgs.append(image)
labels.append(label)
imgs = np.array(imgs)
labels = np.array(labels)
# Create the dataset
hf = h5py.File(ds.mode + '.h5', 'w')
hf.create_dataset('images', data=imgs)
hf.create_dataset('labels', data=labels)
hf.close()
# Load the h5 datasets if available
args.h5 = True
#args.h5 = False
#dataset, val_dataset, test_dataset = initialize_datasets(args, train_size=300, val_size=100)
dataset, val_dataset, test_dataset = initialize_datasets(args)
# Reload weights if desired
args.batch_size = 256
args.log_interval = 10
args.nEpochs = 100
args.nEpochs = 3
test_list = [
#[True, False, 'COVIDNet_small', False, None, 100, 36, 50, 5e-5],
#[True, False, 'COVIDNet_large', False, None, 100, 28, 50, 5e-5],
#[True, True, 'resnet18', False, None, 100, 256, 50, 2e-5],
#[True, True, 'mobilenet_v2', False, None, 100, 256, 50, 2e-5],
#[True, True, 'densenet169', False, None, 100, 256, 50, 2e-5],
#[True, True, 'resnext50_32x4d', False, None, 100, 256, 50, 2e-5]
# Run Inference only
#[False, False, 'resnext50_32x4d', True, "F:/Stanford/CS231N/Project/CS231N/save/resnext50_32x4d_Transfer20200605_0720/best_checkpoint.pt", 100, 16, 50, 2e-5]
[
False, False, 'densenet169', True,
"F:/Stanford/CS231N/Project/CS231N/save/densenet169_Transfer20200605_0127/best_checkpoint.pt",
100, 32, 50, 2e-5
]
]
# Iterate over tests
for trainme, transfer, model_name, reload_weights, weight_path, args.nEpochs, args.batch_size, args.log_interval, args.lr in test_list:
if transfer:
code = "_Transfer"
else:
code = ""
id = model_name + code + util.datestr()
model = select_model(model_name, args.classes)
if reload_weights:
print("Loading model with weights from: {}".format(weight_path))
checkpoint = torch.load(weight_path)
model.load_state_dict(checkpoint['state_dict'])
model.to(device)
#print(model)
# Freeze model if transfer learning
if transfer:
set_parameter_requires_grad(model)
if args.tensorboard and trainme:
writer = SummaryWriter('./runs/' + id)
data_loader = torch.utils.data.DataLoader(
dataset,
batch_size=args.batch_size,
shuffle=True,
pin_memory=False,
num_workers=num_workers)
images, labels = next(iter(data_loader))
writer.add_graph(model, images.to(device))
img_grid = torchvision.utils.make_grid(images)
# show images
newimg = matplotlib_imshow(img_grid, one_channel=True)
## write to tensorboard
writer.add_image('Xrays', img_grid)
else:
writer = None
if trainme:
best_score = 0
optimizer = select_optimizer(args, model)
train_loader = torch.utils.data.DataLoader(
dataset,
batch_size=args.batch_size,
shuffle=True,
pin_memory=False,
num_workers=num_workers)
val_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=args.batch_size,
shuffle=True,
pin_memory=False,
num_workers=num_workers)
for epoch in range(1, args.nEpochs + 1):
# Train 1 epoch
train_metrics, writer_step = train(model, args, device, writer,
optimizer, train_loader,
epoch)
# Run Inference on val set
val_score, confusion_matrix = inference(
args, model, val_loader, epoch, writer, device,
writer_step)
best_score = util.save_model(model, id, args, val_score, epoch,
best_score, confusion_matrix)
# Just evaluate a trained model
else:
val_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=args.batch_size,
shuffle=True,
pin_memory=False,
num_workers=num_workers)
#inference (args, model, val_loader, 1, writer, device, 0)
images, labels = next(iter(val_loader))
images = images.to(device)
labels = labels.to(device)
show_saliency_maps(model, images, labels)
def main():
args = get_arguments()
myargs = [] # getopts(sys.argv)
now = datetime.datetime.now()
cwd = os.getcwd()
if len(myargs) > 0:
if 'c' in myargs:
config_file = myargs['c']
else:
config_file = 'config/trainer_config.yml'
config = OmegaConf.load(os.path.join(cwd, config_file))['trainer']
config.cwd = str(cwd)
reproducibility(config)
dt_string = now.strftime("%d_%m_%Y_%H.%M.%S")
cpkt_fol_name = os.path.join(
config.cwd,
f'checkpoints/model_{config.model.name}/dataset_{config.dataset.name}/date_{dt_string}'
)
log = Logger(path=cpkt_fol_name, name='LOG').get_logger()
best_pred_loss = 1000.0
log.info(f"Checkpoint folder {cpkt_fol_name}")
log.info(f"date and time = {dt_string}")
log.info(f'pyTorch VERSION:{torch.__version__}', )
log.info(f'CUDA VERSION')
log.info(f'CUDNN VERSION:{torch.backends.cudnn.version()}')
log.info(f'Number CUDA Devices: {torch.cuda.device_count()}')
if args.tensorboard:
writer_path = os.path.join(cpkt_fol_name + 'runs/')
writer = SummaryWriter(writer_path + util.datestr())
else:
writer = None
use_cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if use_cuda else "cpu")
log.info(f'device: {device}')
training_generator, val_generator, test_generator, class_dict = select_dataset(
config)
n_classes = len(class_dict)
model = select_model(config, n_classes)
log.info(f"{model}")
if (config.load):
pth_file, _ = load_checkpoint(config.pretrained_cpkt,
model,
strict=True,
load_seperate_layers=False)
else:
pth_file = None
if (config.cuda and use_cuda):
if torch.cuda.device_count() > 1:
log.info(f"Let's use {torch.cuda.device_count()} GPUs!")
model = torch.nn.DataParallel(model)
model.to(device)
optimizer, scheduler = select_optimizer(model, config['model'], None)
log.info(f'{model}')
log.info(f"Checkpoint Folder {cpkt_fol_name} ")
shutil.copy(os.path.join(config.cwd, config_file), cpkt_fol_name)
trainer = Trainer(config,
model=model,
optimizer=optimizer,
data_loader=training_generator,
writer=writer,
logger=log,
valid_data_loader=val_generator,
test_data_loader=test_generator,
class_dict=class_dict,
lr_scheduler=scheduler,
checkpoint_dir=cpkt_fol_name)
trainer.train()
def initialize(args):
if args.device is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.device)
model = select_model(args)
optimizer = select_optimizer(args, model)
if (args.cuda):
model.cuda()
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_transformer = transforms.Compose([
transforms.Resize(256),
transforms.RandomResizedCrop((224), scale=(0.5, 1.0)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(), normalize
])
val_transformer = transforms.Compose([
transforms.Resize(224),
transforms.CenterCrop(224),
transforms.ToTensor(), normalize
])
train_params = {
'batch_size': args.batch_size,
'shuffle': True,
'num_workers': 2
}
test_params = {
'batch_size': args.batch_size,
'shuffle': False,
'num_workers': 1
}
if args.dataset_name == 'COVIDx':
train_loader = COVIDxDataset(mode='train',
n_classes=args.classes,
dataset_path=args.dataset,
dim=(224, 224))
val_loader = COVIDxDataset(mode='test',
n_classes=args.classes,
dataset_path=args.dataset,
dim=(224, 224))
test_loader = None
training_generator = DataLoader(train_loader, **train_params)
val_generator = DataLoader(val_loader, **test_params)
test_generator = None
elif args.dataset_name == 'COVID_CT':
train_loader = CovidCTDataset(
root_dir='data/covid_ct_dataset',
txt_COVID='data/covid_ct_dataset/trainCT_COVID.txt',
txt_NonCOVID='data/covid_ct_dataset/trainCT_NonCOVID.txt',
transform=train_transformer)
val_loader = CovidCTDataset(
root_dir='data/covid_ct_dataset',
txt_COVID='data/covid_ct_dataset/valCT_COVID.txt',
txt_NonCOVID='data/covid_ct_dataset/valCT_NonCOVID.txt',
transform=val_transformer)
test_loader = CovidCTDataset(
root_dir='data/covid_ct_dataset',
txt_COVID='data/covid_ct_dataset/testCT_COVID.txt',
txt_NonCOVID='data/covid_ct_dataset/testCT_NonCOVID.txt',
transform=val_transformer)
training_generator = DataLoader(train_loader, **train_params)
val_generator = DataLoader(val_loader, **test_params)
test_generator = DataLoader(test_loader, **test_params)
return model, optimizer, training_generator, val_generator, test_generator