def main(args):
args = {
"model": "vqvae",
"batch_size": 1,
# "hidden": 64, # 128,
"hidden": 128, # 128,
# "k": 128, # 512,
"k": 256, # 512,
"lr": 5e-7, #5e-6,#2e-4,
"n_recordings": n_recordings,
"vq_coef": 2, # ?
"commit_coef": 2, # ?
"kl_coef": 1, # ?
# "noise_coef" : 1e3,
"dataset": "imagenet",
"epochs": 250 * 4,
"cuda": torch.cuda.is_available(),
"seed": 1,
"gpus": "1",
"log_interval": 50,
"results_dir": "VAE_imagenet",
"save_name": "first",
"data_format": "json",
"num_workers": 4,
"num_folds": 5,
}
lr = args["lr"] # or default_hyperparams[args["dataset"]]['lr']
k = args["k"] # or default_hyperparams[args["dataset"]]['k']
hidden = args["hidden"] or default_hyperparams[args["dataset"]]['hidden']
num_channels = dataset_sizes[args["dataset"]][0]
results, save_path = setup_logging_and_results(args)
torch.manual_seed(args["seed"])
if args["cuda"]:
torch.cuda.manual_seed_all(args["seed"])
args["gpus"] = [int(i) for i in args["gpus"].split(',')]
# torch.cuda.set_device(args["gpus"][0])
cudnn.benchmark = True
torch.cuda.manual_seed(args["seed"])
model = models[args["dataset"]][args["model"]](hidden,
k=k,
num_channels=num_channels)
#model = MyDataParallel(model)
print("Number of Parameters in Model:",
sum(p.numel() for p in model.parameters() if p.requires_grad))
if args["cuda"]:
model.cuda()
# noise_adder = AddNoiseManual(b=.5)
optimizer = optim.Adam(model.parameters(), lr=lr)
scheduler = optim.lr_scheduler.StepLR(
optimizer,
10 if args["dataset"] == 'imagenet' else 30,
0.5,
)
kwargs = {'num_workers': 1, 'pin_memory': True} if args["cuda"] else {}
train_dataset = SeizureDataset(TRAIN_SEIZURE_FILE,
num_folds=args["num_folds"],
cross_val=False,
split='train')
print("len train dataset", len(train_dataset))
#train_dataset = EstimatedDataset(800, transform=dataset_transforms[args["dataset"]])
train_loader = data.DataLoader(dataset=train_dataset,
shuffle=True,
batch_size=args["batch_size"],
num_workers=args["num_workers"])
print("train_loader", len(train_loader))
eval_dataset = SeizureDataset(DEV_SEIZURE_FILE,
num_folds=args["num_folds"],
cross_val=False,
split='dev') # TODO: enter right split
# eval_dataset = EstimatedDataset(20, transform=dataset_transforms[args["dataset"]])
eval_loader = data.DataLoader(dataset=eval_dataset,
shuffle=False,
batch_size=args["batch_size"],
num_workers=args["num_workers"])
print("eval_loader", len(eval_loader))
print("Save path", save_path)
for epoch in range(1, args["epochs"] + 1):
train_losses = train(epoch, model, train_loader, optimizer,
args["cuda"], args["log_interval"], save_path,
args)
test_losses = test_net(epoch, model, eval_loader, args["cuda"],
save_path, args)
writer.flush()
# writer.close()
torch.save(model.state_dict(), "saved_models/" + save_filename + ".pt")
def main(args):
parser = argparse.ArgumentParser(description='Variational AutoEncoders')
model_parser = parser.add_argument_group('Model Parameters')
model_parser.add_argument('--model', default='vae', choices=['vae', 'vqvae'],
help='autoencoder variant to use: vae | vqvae')
model_parser.add_argument('--batch-size', type=int, default=128, metavar='N',
help='input batch size for training (default: 128)')
model_parser.add_argument('--hidden', type=int, metavar='N',
help='number of hidden channels')
model_parser.add_argument('-k', '--dict-size', type=int, dest='k', metavar='K',
help='number of atoms in dictionary')
model_parser.add_argument('--lr', type=float, default=None,
help='learning rate')
model_parser.add_argument('--vq_coef', type=float, default=None,
help='vq coefficient in loss')
model_parser.add_argument('--commit_coef', type=float, default=None,
help='commitment coefficient in loss')
model_parser.add_argument('--kl_coef', type=float, default=None,
help='kl-divergence coefficient in loss')
training_parser = parser.add_argument_group('Training Parameters')
training_parser.add_argument('--dataset', default='cifar10', choices=['mnist', 'cifar10', 'imagenet'],
help='dataset to use: mnist | cifar10')
training_parser.add_argument('--data-dir', default='datasets/',
help='directory containing the dataset')
training_parser.add_argument('--epochs', type=int, default=20, metavar='N',
help='number of epochs to train (default: 10)')
training_parser.add_argument('--no-cuda', action='store_true', default=False,
help='enables CUDA training')
training_parser.add_argument('--seed', type=int, default=1, metavar='S',
help='random seed (default: 1)')
training_parser.add_argument('--gpus', default='0',
help='gpus used for training - e.g 0,1,3')
logging_parser = parser.add_argument_group('Logging Parameters')
logging_parser.add_argument('--log-interval', type=int, default=10, metavar='N',
help='how many batches to wait before logging training status')
logging_parser.add_argument('--results-dir', metavar='RESULTS_DIR', default='./results',
help='results dir')
logging_parser.add_argument('--save-name', default='',
help='saved folder')
logging_parser.add_argument('--data-format', default='json',
help='in which format to save the data')
args = parser.parse_args(args)
args.cuda = not args.no_cuda and torch.cuda.is_available()
lr = args.lr or default_hyperparams[args.dataset]['lr']
k = args.k or default_hyperparams[args.dataset]['k']
hidden = args.hidden or default_hyperparams[args.dataset]['hidden']
num_channels = dataset_sizes[args.dataset][0]
results, save_path = setup_logging_and_results(args)
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed_all(args.seed)
args.gpus = [int(i) for i in args.gpus.split(',')]
torch.cuda.set_device(args.gpus[0])
cudnn.benchmark = True
torch.cuda.manual_seed(args.seed)
model = models[args.dataset][args.model](hidden, k=k, num_channels=num_channels)
if args.cuda:
model.cuda()
optimizer = optim.Adam(model.parameters(), lr=lr)
scheduler = optim.lr_scheduler.StepLR(optimizer, 10 if args.dataset == 'imagenet' else 30, 0.5,)
kwargs = {'num_workers': 1, 'pin_memory': True} if args.cuda else {}
dataset_train_dir = os.path.join(args.data_dir, args.dataset)
dataset_test_dir = os.path.join(args.data_dir, args.dataset)
if 'imagenet' in args.dataset:
dataset_train_dir = os.path.join(dataset_train_dir, 'train')
dataset_test_dir = os.path.join(dataset_test_dir, 'val')
train_loader = torch.utils.data.DataLoader(
datasets_classes[args.dataset](dataset_train_dir,
transform=dataset_transforms[args.dataset],
**dataset_train_args[args.dataset]),
batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
datasets_classes[args.dataset](dataset_test_dir,
transform=dataset_transforms[args.dataset],
**dataset_test_args[args.dataset]),
batch_size=args.batch_size, shuffle=True, **kwargs)
for epoch in range(1, args.epochs + 1):
train_losses = train(epoch, model, train_loader, optimizer, args.cuda, args.log_interval, save_path, args)
test_losses = test_net(epoch, model, test_loader, args.cuda, save_path, args)
results.add(epoch=epoch, **train_losses, **test_losses)
for k in train_losses:
key = k[:-6]
results.plot(x='epoch', y=[key + '_train', key + '_test'])
results.save()
scheduler.step()
def main(args):
args = {
"model": "vqvae",
"batch_size": 8, # 7 * 9 = 63 (close to desired 64)
"hidden": 128,
"k": 512 * 1,
"lr": 2e-4,
"n_recordings" : n_recordings,
"vq_coef": 1, # ?
"commit_coef": 1, # ?
"kl_coef": 1, # ?
# "noise_coef" : 1e3,
"dataset": "imagenet",
"epochs": 250 * 4,
"cuda": torch.cuda.is_available(),
"seed": 1,
"gpus": "1",
"log_interval": 50,
"results_dir": "VAE_imagenet",
"save_name": "first",
"data_format": "json"
}
#
# parser = argparse.ArgumentParser(description='Variational AutoEncoders')
#
# model_parser = parser.add_argument_group('Model Parameters')
# model_parser.add_argument('--model', default='vae', choices=['vae', 'vqvae'],
# help='autoencoder variant to use: vae | vqvae')
# model_parser.add_argument('--batch-size', type=int, default=128, metavar='N',
# help='input batch size for training (default: 128)')
# model_parser.add_argument('--hidden', type=int, metavar='N',
# help='number of hidden channels')
# model_parser.add_argument('-k', '--dict-size', type=int, dest='k', metavar='K',
# help='number of atoms in dictionary')
# model_parser.add_argument('--lr', type=float, default=None,
# help='learning rate')
# model_parser.add_argument('--vq_coef', type=float, default=None,
# help='vq coefficient in loss')
# model_parser.add_argument('--commit_coef', type=float, default=None,
# help='commitment coefficient in loss')
# model_parser.add_argument('--kl_coef', type=float, default=None,
# help='kl-divergence coefficient in loss')
#
# training_parser = parser.add_argument_group('Training Parameters')
# training_parser.add_argument('--dataset', default='cifar10', choices=['mnist', 'cifar10', 'imagenet'],
# help='dataset to use: mnist | cifar10')
# training_parser.add_argument('--data-dir', default='/media/ssd/Datasets',
# help='directory containing the dataset')
# training_parser.add_argument('--epochs', type=int, default=20, metavar='N',
# help='number of epochs to train (default: 10)')
# training_parser.add_argument('--no-cuda', action='store_true', default=False,
# help='enables CUDA training')
# training_parser.add_argument('--seed', type=int, default=1, metavar='S',
# help='random seed (default: 1)')
# training_parser.add_argument('--gpus', default='0',
# help='gpus used for training - e.g 0,1,3')
#
# logging_parser = parser.add_argument_group('Logging Parameters')
# logging_parser.add_argument('--log-interval', type=int, default=10, metavar='N',
# help='how many batches to wait before logging training status')
# logging_parser.add_argument('--results-dir', metavar='RESULTS_DIR', default='./results',
# help='results dir')
# logging_parser.add_argument('--save-name', default='',
# help='saved folder')
# logging_parser.add_argument('--data-format', default='json',
# help='in which format to save the data')
# args = parser.parse_args(args)
# args["cuda"] = not args["no_cuda"] and torch.cuda.is_available()
lr = args["lr"] # or default_hyperparams[args["dataset"]]['lr']
k = args["k"] # or default_hyperparams[args["dataset"]]['k']
hidden = args["hidden"] or default_hyperparams[args["dataset"]]['hidden']
num_channels = dataset_sizes[args["dataset"]][0]
results, save_path = setup_logging_and_results(args)
torch.manual_seed(args["seed"])
if args["cuda"]:
torch.cuda.manual_seed_all(args["seed"])
args["gpus"] = [int(i) for i in args["gpus"].split(',')]
# torch.cuda.set_device(args["gpus"][0])
cudnn.benchmark = True
torch.cuda.manual_seed(args["seed"])
model = models[args["dataset"]][args["model"]](hidden, k=k, num_channels=num_channels)
#model = MyDataParallel(model)
print("Number of Parameters in Model:", sum(p.numel() for p in model.parameters() if p.requires_grad))
if args["cuda"]:
model.cuda()
# noise_adder = AddNoiseManual(b=.5)
optimizer = optim.Adam(model.parameters(), lr=lr)
scheduler = optim.lr_scheduler.StepLR(optimizer, 10 if args["dataset"] == 'imagenet' else 30, 0.5, )
kwargs = {'num_workers': 1, 'pin_memory': True} if args["cuda"] else {}
train_dataset = EstimatedDataset(8 * 20 * 5, transform=dataset_transforms[args["dataset"]])
train_loader = torch.utils.data.DataLoader(
dataset=train_dataset,
shuffle=True,
batch_size=args["batch_size"],
)
test_dataset = EstimatedDataset(1, transform=dataset_transforms[args["dataset"]])
test_loader = torch.utils.data.DataLoader(
dataset=test_dataset,
batch_size=args["batch_size"],
shuffle=True
)
print("Save path", save_path)
for epoch in range(1, args["epochs"] + 1):
train_losses = train(epoch, model, train_loader, optimizer, args["cuda"], args["log_interval"], save_path, args)
torch.save(model.state_dict(), "saved_models/" + save_filename + ".pt")
def main(args):
args = {
"model": "vqvae",
"batch_size": 16, # anything too large and runs out of vRAM
"hidden": 128, # 128,
"k": 256, # 512,
"lr": 5e-7, #5e-6,#2e-4,
"vq_coef": 2,
"commit_coef": 2,
"kl_coef": 1,
"dataset": "imagenet",
"epochs": 250 * 4,
"cuda": torch.cuda.is_available(),
"seed": 1,
"gpus": "1",
"log_interval": 50,
"results_dir": "VAE_imagenet",
"save_name": "first",
"data_format": "json",
"num_workers": 4,
"num_folds": 5,
"num_estimated_batches": 16
}
lr = args["lr"] # or default_hyperparams[args["dataset"]]['lr']
k = args["k"] # or default_hyperparams[args["dataset"]]['k']
hidden = args["hidden"] or default_hyperparams[args["dataset"]]['hidden']
num_channels = dataset_sizes[args["dataset"]][0]
results, save_path = setup_logging_and_results(args)
torch.manual_seed(args["seed"])
if args["cuda"]:
torch.cuda.manual_seed_all(args["seed"])
args["gpus"] = [int(i) for i in args["gpus"].split(',')]
# torch.cuda.set_device(args["gpus"][0])
cudnn.benchmark = True
torch.cuda.manual_seed(args["seed"])
model = models[args["dataset"]][args["model"]](hidden,
k=k,
num_channels=num_channels)
if args["cuda"]:
model.cuda()
print("Number of Parameters in Model:",
sum(p.numel() for p in model.parameters() if p.requires_grad))
optimizer = optim.Adam(model.parameters(), lr=lr)
scheduler = optim.lr_scheduler.StepLR(
optimizer,
10 if args["dataset"] == 'imagenet' else 30,
0.5,
)
kwargs = {'num_workers': 1, 'pin_memory': True} if args["cuda"] else {}
### Load Training Data ###
# ideally use a dataset of certin artificact not just the standard recordings
train_dataset_artifacts = SeizureDataset(TRAIN_SEIZURE_FILE,
num_folds=args["num_folds"],
cross_val=False,
split='train',
transform=normalize)
train_artifacts_loader = data.DataLoader(dataset=train_dataset_artifacts,
shuffle=True,
batch_size=args["batch_size"],
num_workers=args["num_workers"])
num_examples = len(train_dataset_artifacts)
print("len train dataset", num_examples)
# more examples would occupy more RAM and take a while to compute
num_examples_estimated = args["batch_size"] * args["num_estimated_batches"]
train_dataset_estimated = EstimatedDataset(num_examples_estimated,
transform=normalize)
train_estimated_loader = data.DataLoader(dataset=train_dataset_estimated,
shuffle=True,
batch_size=args["batch_size"],
num_workers=args["num_workers"])
### Load Eval Data ###
eval_artifcats_dataset = SeizureDataset(DEV_SEIZURE_FILE,
num_folds=args["num_folds"],
cross_val=False,
split='dev',
transform=normalize)
eval_artificats_loader = data.DataLoader(
dataset=eval_artifcats_dataset,
shuffle=False,
batch_size=args["batch_size"],
num_workers=args["num_workers"],
)
print("eval loader", len(eval_artificats_loader))
eval_dataset_estimated = EstimatedDataset(num_examples_estimated,
transform=normalize)
eval_estimated_loader = data.DataLoader(dataset=eval_dataset_estimated,
shuffle=True,
batch_size=args["batch_size"],
num_workers=args["num_workers"])
print("Save path", save_path)
for epoch in range(1, args["epochs"] + 1):
try:
train_losses = train(epoch, model, train_estimated_loader,
train_artifacts_loader, optimizer,
args["cuda"], args["log_interval"], save_path,
args)
test_losses = test_net(epoch, model, eval_estimated_loader,
eval_artificats_loader, args["cuda"],
save_path, args)
writer.flush()
torch.save(model.state_dict(),
"saved_models/" + save_filename + ".pt")
except KeyboardInterrupt:
print("Early Ending")
break
writer.flush()
writer.close()
torch.save(model.state_dict(), "saved_models/" + save_filename + ".pt")