def train(lr, l2, momentum, patience, latent_size, n_hidden, hidden_size, n_frames, model, ncoef, dropout_prob, epochs, batch_size, n_workers, cuda, train_hdf_file, valid_hdf_file, valid_n_cycles, cp_path, softmax): if cuda: device=get_freer_gpu() train_dataset=Loader_test(hdf5_name=train_hdf_file, max_nb_frames=int(n_frames)) train_loader=torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=False, num_workers=n_workers, worker_init_fn=set_np_randomseed) valid_dataset = Loader(hdf5_name = valid_hdf_file, max_nb_frames = int(n_frames), n_cycles=valid_n_cycles) valid_loader=torch.utils.data.DataLoader(valid_dataset, batch_size=batch_size, shuffle=False, num_workers=n_workers, worker_init_fn=set_np_randomseed) if args.model == 'resnet_stats': model = model_.ResNet_stats(n_z=int(latent_size), nh=int(n_hidden), n_h=int(hidden_size), proj_size=len(train_dataset.speakers_list), ncoef=ncoef, dropout_prob=dropout_prob, sm_type=softmax) elif args.model == 'resnet_mfcc': model = model_.ResNet_mfcc(n_z=int(latent_size), nh=int(n_hidden), n_h=int(hidden_size), proj_size=len(train_dataset.speakers_list), ncoef=ncoef, dropout_prob=dropout_prob, sm_type=softmax) if args.model == 'resnet_lstm': model = model_.ResNet_lstm(n_z=int(latent_size), nh=int(n_hidden), n_h=int(hidden_size), proj_size=len(train_dataset.speakers_list), ncoef=ncoef, dropout_prob=dropout_prob, sm_type=softmax) elif args.model == 'resnet_small': model = model_.ResNet_small(n_z=int(latent_size), nh=int(n_hidden), n_h=int(hidden_size), proj_size=len(train_dataset.speakers_list), ncoef=ncoef, dropout_prob=dropout_prob, sm_type=softmax) elif args.model == 'resnet_large': model = model_.ResNet_large(n_z=int(latent_size), nh=int(n_hidden), n_h=int(hidden_size), proj_size=len(train_dataset.speakers_list), ncoef=ncoef, dropout_prob=dropout_prob, sm_type=softmax) if cuda: model=model.cuda(device) else: device=None optimizer=optim.SGD(model.parameters(), lr=lr, momentum=momentum, weight_decay=l2) trainer=TrainLoop(model, optimizer, train_loader, valid_loader, patience=int(patience), verbose=-1, device=device, cp_name=get_file_name(cp_path), save_cp=False, checkpoint_path=cp_path, pretrain=False, cuda=cuda) return trainer.train(n_epochs=epochs)
def plot_real(n_tests, data_path):
real_loader = Loader(hdf5_name=data_path)
n_cols, n_rows = (n_tests, 15)
fig, axes = plt.subplots(n_cols, n_rows, figsize=(n_rows, n_cols))
for i in range(n_tests):
img_idx = np.random.randint(len(real_loader))
real_sample = real_loader[img_idx].squeeze()
real_sample = real_sample.transpose(0, 1)
print(real_sample.size())
for ax, img in zip(axes[i, :].flatten(), real_sample):
img = denorm(img.transpose(0, -1))
ax.axis('off')
ax.set_adjustable('box-forced')
ax.imshow(img, aspect='equal')
plt.subplots_adjust(wspace=0, hspace=0)
save_fn = 'real.pdf'
plt.savefig(save_fn)
plt.close()
def plot_real(n_tests, data_path):
real_loader = Loader(hdf5_name=data_path)
n_cols, n_rows = (n_tests, 30)
fig, axes = plt.subplots(n_cols, n_rows, figsize=(n_rows, n_cols))
intra_mse = []
for i in range(n_tests):
img_idx = np.random.randint(len(real_loader))
real_sample = real_loader[img_idx].squeeze()
for ax, img in zip(axes[i, :].flatten(), real_sample):
ax.axis('off')
ax.set_adjustable('box-forced')
ax.imshow(img, cmap="gray", aspect='equal')
plt.subplots_adjust(wspace=0, hspace=0)
mse = 0.0
#mse = []
count = 0
for j in range(1, real_sample.size(0)):
if j % 1 == 0:
a = real_sample[j].numpy().squeeze()
a[a > 0] = 1.
a[a < 0] = 0.
b = real_sample[j - 1].numpy().squeeze()
b[b > 0] = 1.
b[b < 0] = 0.
count += 1
mse += mean_squared_error(a, b)
#mse+=torch.nn.functional.mse_loss(real_sample[j,:,:], real_sample[j-1,:,:]).item()
mse /= count
intra_mse.append(np.asarray(mse))
save_fn = 'real.pdf'
plt.savefig(save_fn)
plt.close()
return intra_mse
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed(args.seed)
if args.n_classes > 2:
assert args.lists_path is not None, 'Pass the path for the lists of utterances per attack'
train_dataset = Loader_mcc(
hdf5_clean=args.train_hdf_path + 'train_clean.hdf',
hdf5_attack=args.train_hdf_path + 'train_attack.hdf',
max_nb_frames=args.n_frames,
n_cycles=args.n_cycles,
file_lists_path=args.lists_path)
else:
train_dataset = Loader(hdf5_clean=args.train_hdf_path + 'train_clean.hdf',
hdf5_attack=args.train_hdf_path +
'train_attack.hdf',
max_nb_frames=args.n_frames,
n_cycles=args.n_cycles)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers)
if args.valid_hdf_path is not None:
if args.n_classes > 2:
valid_dataset = Loader_mcc(
hdf5_clean=args.valid_hdf_path + 'valid_clean.hdf',
hdf5_attack=args.valid_hdf_path + 'valid_attack.hdf',
max_nb_frames=args.n_frames,
n_cycles=args.valid_n_cycles,
def train(lr, l2, momentum, smoothing, warmup, model, emb_size, n_hidden,
hidden_size, dropout_prob, epochs, batch_size, valid_batch_size,
n_workers, cuda, data_path, valid_data_path, hdf_path,
valid_hdf_path, checkpoint_path, softmax, pretrained,
pretrained_path, max_gnorm, stats, log_dir, eval_every, ablation):
args_dict = locals()
cp_name = get_cp_name(checkpoint_path)
if pretrained_path != 'none':
print('\nLoading pretrained model from: {}\n'.format(
args.pretrained_path))
ckpt = torch.load(pretrained_path,
map_location=lambda storage, loc: storage)
dropout_prob, n_hidden, hidden_size, emb_size = ckpt[
'dropout_prob'], ckpt['n_hidden'], ckpt['hidden_size'], ckpt[
'emb_size']
if 'r_proj_size' in ckpt:
rproj_size = ckpt['r_proj_size']
else:
rproj_size = -1
print('\nUsing pretrained config for discriminator. Ignoring args.')
args_dict['dropout_prob'], args_dict['n_hidden'], args_dict[
'hidden_size'], args_dict[
'emb_size'] = dropout_prob, n_hidden, hidden_size, emb_size
if log_dir != 'none':
writer = SummaryWriter(log_dir=os.path.join(log_dir, cp_name),
comment=model,
purge_step=0)
writer.add_hparams(hparam_dict=args_dict,
metric_dict={'best_eer': 0.0})
else:
writer = None
if stats == 'cars':
mean, std = [0.4461, 0.4329, 0.4345], [0.2888, 0.2873, 0.2946]
elif stats == 'cub':
mean, std = [0.4782, 0.4925, 0.4418], [0.2330, 0.2296, 0.2647]
elif stats == 'sop':
mean, std = [0.5603, 0.5155, 0.4796], [0.2939, 0.2991, 0.3085]
elif stats == 'imagenet':
mean, std = [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]
if hdf_path != 'none':
transform_train = transforms.Compose([
transforms.ToPILImage(),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(10),
transforms.RandomPerspective(p=0.1),
transforms.RandomGrayscale(p=0.1),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)
])
trainset = Loader(hdf_path, transform_train)
else:
transform_train = transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(10),
transforms.RandomPerspective(p=0.1),
transforms.RandomGrayscale(p=0.1),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)
])
trainset = datasets.ImageFolder(data_path, transform=transform_train)
train_loader = torch.utils.data.DataLoader(
trainset,
batch_size=batch_size,
shuffle=True,
num_workers=n_workers,
worker_init_fn=set_np_randomseed,
pin_memory=True)
if valid_hdf_path != 'none':
transform_test = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)
])
validset = Loader(args.valid_hdf_path, transform_test)
else:
transform_test = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)
])
validset = datasets.ImageFolder(args.valid_data_path,
transform=transform_test)
valid_loader = torch.utils.data.DataLoader(validset,
batch_size=valid_batch_size,
shuffle=True,
num_workers=n_workers,
pin_memory=True)
nclasses = trainset.n_classes if isinstance(trainset, Loader) else len(
trainset.classes)
if model == 'vgg':
model_ = vgg.VGG('VGG19',
nh=n_hidden,
n_h=hidden_size,
dropout_prob=dropout_prob,
sm_type=softmax,
n_classes=nclasses,
emb_size=emb_size,
r_proj_size=rproj_size)
elif model == 'resnet':
model_ = resnet.ResNet50(nh=n_hidden,
n_h=hidden_size,
dropout_prob=dropout_prob,
sm_type=softmax,
n_classes=nclasses,
emb_size=emb_size,
r_proj_size=rproj_size)
elif model == 'densenet':
model_ = densenet.DenseNet121(nh=n_hidden,
n_h=hidden_size,
dropout_prob=dropout_prob,
sm_type=softmax,
n_classes=nclasses,
emb_size=emb_size,
r_proj_size=rproj_size)
if pretrained_path != 'none':
if ckpt['sm_type'] == 'am_softmax':
del (ckpt['model_state']['out_proj.w'])
elif ckpt['sm_type'] == 'softmax':
del (ckpt['model_state']['out_proj.w.weight'])
del (ckpt['model_state']['out_proj.w.bias'])
print(model_.load_state_dict(ckpt['model_state'], strict=False))
print('\n')
if pretrained:
print('\nLoading pretrained encoder from torchvision\n')
if model == 'vgg':
model_pretrained = torchvision.models.vgg19(pretrained=True)
elif model == 'resnet':
model_pretrained = torchvision.models.resnet50(pretrained=True)
elif model == 'densenet':
model_pretrained = torchvision.models.densenet121(pretrained=True)
print(
model_.load_state_dict(model_pretrained.state_dict(),
strict=False))
print('\n')
if cuda:
device = get_freer_gpu()
model_ = model_.cuda(device)
torch.backends.cudnn.benchmark = True
optimizer = TransformerOptimizer(optim.SGD(model_.parameters(),
lr=lr,
momentum=momentum,
weight_decay=l2,
nesterov=True),
lr=lr,
warmup_steps=warmup)
trainer = TrainLoop(model_,
optimizer,
train_loader,
valid_loader,
max_gnorm=max_gnorm,
label_smoothing=smoothing,
verbose=-1,
cp_name=cp_name,
save_cp=True,
checkpoint_path=checkpoint_path,
ablation=ablation,
cuda=cuda,
logger=writer)
for i in range(5):
print(' ')
print('Hyperparameters:')
print('Selected model: {}'.format(model))
print('Embedding size: {}'.format(emb_size))
print('Hidden layer size: {}'.format(hidden_size))
print('Number of hidden layers: {}'.format(n_hidden))
print('Random projection size: {}'.format(rproj_size))
print('Dropout rate: {}'.format(dropout_prob))
print('Batch size: {}'.format(batch_size))
print('LR: {}'.format(lr))
print('Momentum: {}'.format(momentum))
print('l2: {}'.format(l2))
print('Label smoothing: {}'.format(smoothing))
print('Warmup iterations: {}'.format(warmup))
print('Softmax Mode is: {}'.format(softmax))
print('Pretrained: {}'.format(pretrained))
print('Pretrained path: {}'.format(pretrained_path))
print('Evaluate every {} iterations.'.format(eval_every))
print('Ablation Mode: {}'.format(ablation))
print(' ')
if i > 0:
print(' ')
print('Trial {}'.format(i + 1))
print(' ')
try:
cost = trainer.train(n_epochs=epochs,
save_every=epochs + 10,
eval_every=eval_every)
print(' ')
print('Best e2e EER in file ' + cp_name +
' was: {}'.format(cost[0]))
print('Best cos EER in file ' + cp_name +
' was: {}'.format(cost[1]))
print(' ')
if log_dir != 'none':
writer.add_hparams(hparam_dict=args_dict,
metric_dict={'best_eer': cost[0]})
return cost[0]
except:
print("Error:", sys.exc_info())
pass
print('Returning dummy cost due to failures while training.')
return 0.99
parser.add_argument('--checkpoint-epoch', type=int, default=None, metavar='N', help='epoch to load for checkpointing. If None, training starts from scratch')
parser.add_argument('--checkpoint-path', type=str, default=None, metavar='Path', help='Path for checkpointing')
parser.add_argument('--generator-path', type=str, default=None, metavar='Path', help='Path for frames generator params')
parser.add_argument('--ndiscriminators', type=int, default=8, help='Number of discriminators. Default=8')
parser.add_argument('--nadir-slack', type=float, default=1.5, metavar='nadir', help='factor for nadir-point update. Only used in hyper mode (default: 1.5)')
parser.add_argument('--seed', type=int, default=1, metavar='S', help='random seed (default: 1)')
parser.add_argument('--save-every', type=int, default=5, metavar='N', help='how many batches to wait before logging training status. (default: 5)')
parser.add_argument('--n-workers', type=int, default=4)
parser.add_argument('--gen-arch', choices=['linear', 'conv'], default='linear', help='Linear or convolutional generator')
parser.add_argument('--no-cuda', action='store_true', default=False, help='Disables GPU use')
parser.add_argument('--optimizer', choices=['adam', 'amsgrad', 'rmsprop'], default='adam', help='Select optimizer (Default is adam).')
parser.add_argument('--average-mode', action='store_true', default=False, help='Disables hypervolume maximization and uses average loss instead')
args = parser.parse_args()
args.cuda = True if not args.no_cuda and torch.cuda.is_available() else False
train_data_set = Loader(hdf5_name=args.data_path)
train_loader = DataLoader(train_data_set, batch_size=args.batch_size, shuffle=False, num_workers=args.n_workers)
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed(args.seed)
if args.gen_arch == 'conv':
generator = models_zoo.Generator_conv(args.cuda)
elif args.gen_arch == 'linear':
generator = models_zoo.Generator_linear(args.cuda)
frames_generator = models_zoo.frames_generator().eval()
gen_state = torch.load(args.generator_path, map_location=lambda storage, loc: storage)
frames_generator.load_state_dict(gen_state['model_state'])
print(args, '\n')
if args.cuda:
torch.backends.cudnn.benchmark = True
if args.hdf_path:
transform_train = transforms.Compose([
transforms.ToPILImage(),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(30),
transforms.RandomPerspective(p=0.2),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)
])
trainset = Loader(args.hdf_path, transform_train)
else:
transform_train = transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(30),
transforms.RandomPerspective(p=0.2),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)
])
trainset = datasets.ImageFolder(args.data_path, transform=transform_train)
train_loader = torch.utils.data.DataLoader(trainset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.n_workers,
type=int,
default=1,
metavar='N',
help='Verbose is activated if > 0')
parser.add_argument('--logdir',
type=str,
default=None,
metavar='Path',
help='Path for checkpointing')
args = parser.parse_args()
args.cuda = True if not args.no_cuda and torch.cuda.is_available() else False
if args.cuda:
torch.backends.cudnn.benchmark = True
trainset = Loader(hdf5_name=args.train_hdf_file, max_nb_frames=args.n_frames)
train_loader = torch.utils.data.DataLoader(trainset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.n_workers,
worker_init_fn=set_np_randomseed,
pin_memory=True)
validset = Loader(hdf5_name=args.valid_hdf_file, max_nb_frames=args.n_frames)
valid_loader = torch.utils.data.DataLoader(validset,
batch_size=args.valid_batch_size,
shuffle=True,
num_workers=args.n_workers,
pin_memory=True)
args.nclasses = trainset.nclasses
import cupy
cupy.cuda.Device(int(str(device).split(':')[-1])).use()
else:
device = None
if args.logdir:
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter(log_dir=os.path.join(args.logdir, args.cp_name),
comment=args.model,
purge_step=0)
writer.add_hparams(hparam_dict=args_dict, metric_dict={'best_eer': 0.0})
else:
writer = None
train_dataset = Loader(hdf5_name=args.train_hdf_file,
max_nb_frames=args.n_frames,
delta=args.delta)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
worker_init_fn=set_np_randomseed)
valid_dataset = Loader_valid(hdf5_name=args.valid_hdf_file,
max_nb_frames=args.n_frames,
delta=args.delta)
valid_loader = torch.utils.data.DataLoader(valid_dataset,
batch_size=args.valid_batch_size,
shuffle=True,
num_workers=args.workers,
worker_init_fn=set_np_randomseed)
def train(lr, l2, momentum, max_gnorm, warmup, input_size, n_hidden,
hidden_size, dropout_prob, smoothing, n_cycles, epochs, batch_size,
valid_batch_size, n_workers, cuda, train_hdf_path, valid_hdf_path,
cp_path, logdir):
hp_dict = {
'lr': lr,
'l2': l2,
'momentum': momentum,
'max_gnorm': max_gnorm,
'warmup': warmup,
'input_size': input_size,
'n_hidden': n_hidden,
'hidden_size': hidden_size,
'dropout_prob': dropout_prob,
'smoothing': smoothing,
'n_cycles': n_cycles,
'epochs': epochs,
'batch_size': batch_size,
'valid_batch_size': valid_batch_size,
'n_workers': n_workers,
'cuda': cuda,
'train_hdf_path': train_hdf_path,
'valid_hdf_path': valid_hdf_path,
'cp_path': cp_path
}
cp_name = get_file_name(cp_path)
if args.logdir:
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter(log_dir=logdir + cp_name, purge_step=True)
writer.add_hparams(hparam_dict=hp_dict, metric_dict={'best_eer': 0.5})
else:
writer = None
train_dataset = Loader(hdf5_clean=train_hdf_path + 'train_clean.hdf',
hdf5_attack=train_hdf_path + 'train_attack.hdf',
label_smoothing=smoothing,
n_cycles=n_cycles)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=n_workers)
valid_dataset = Loader(hdf5_clean=valid_hdf_path + 'valid_clean.hdf',
hdf5_attack=valid_hdf_path + 'valid_attack.hdf',
n_cycles=1)
valid_loader = torch.utils.data.DataLoader(valid_dataset,
batch_size=valid_batch_size,
shuffle=False,
num_workers=n_workers)
model = model_.MLP(n_in=input_size,
nh=n_hidden,
n_h=hidden_size,
dropout_prob=dropout_prob)
if cuda:
device = get_freer_gpu()
model = model.cuda(device)
optimizer = TransformerOptimizer(optim.SGD(model.parameters(),
lr=lr,
momentum=momentum,
weight_decay=l2,
nesterov=True),
lr=lr,
warmup_steps=warmup)
trainer = TrainLoop(model,
optimizer,
train_loader,
valid_loader,
max_gnorm=max_gnorm,
verbose=-1,
cp_name=cp_name,
save_cp=True,
checkpoint_path=cp_path,
cuda=cuda,
logger=writer)
for i in range(5):
if i > 0:
print(' ')
print('Trial {}'.format(i + 1))
print(' ')
try:
cost = trainer.train(n_epochs=epochs, save_every=epochs + 10)
print(' ')
print('Best EER in file ' + cp_name + ' was: {}'.format(cost))
print(' ')
print('With hyperparameters:')
print('Hidden layer size size: {}'.format(int(hidden_size)))
print('Number of hidden layers: {}'.format(int(n_hidden)))
print('Dropout rate: {}'.format(dropout_prob))
print('Batch size: {}'.format(batch_size))
print('LR: {}'.format(lr))
print('Warmup iterations: {}'.format(warmup))
print('Momentum: {}'.format(momentum))
print('l2: {}'.format(l2))
print('Max. Grad. norm: {}'.format(max_gnorm))
print('Label smoothing: {}'.format(smoothing))
print(' ')
if args.logdir:
writer.add_hparams(hparam_dict=hp_dict,
metric_dict={'best_eer': cost})
return cost
except:
pass
print('Returning dummy cost due to failures while training.')
cost = 0.99
if args.logdir:
writer.add_hparams(hparam_dict=hp_dict, metric_dict={'best_eer': cost})
return cost
if args.cuda:
device = get_freer_gpu()
else:
device = torch.device('cpu')
if args.logdir:
writer = SummaryWriter(log_dir=args.logdir, purge_step=True)
else:
writer = None
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed(args.seed)
train_dataset = Loader(hdf5_clean = args.train_hdf_path+'train_clean.hdf', hdf5_attack = args.train_hdf_path+'train_attack.hdf', label_smoothing=args.smoothing, n_cycles=args.n_cycles)
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers)
valid_dataset = Loader(hdf5_clean = args.valid_hdf_path+'valid_clean.hdf', hdf5_attack = args.valid_hdf_path+'valid_attack.hdf', n_cycles=1)
valid_loader = torch.utils.data.DataLoader(valid_dataset, batch_size=args.valid_batch_size, shuffle=False, num_workers=args.workers)
model = model_.MLP(n_in=args.input_size, nh=args.n_hidden, n_h=args.hidden_size, dropout_prob=args.dropout_prob)
if args.pretrained_path is not None:
ckpt = torch.load(args.pretrained_path, map_location = lambda storage, loc: storage)
model = model_.MLP(n_in=ckpt['input_size'], nh=ckpt['n_hidden'], n_h=ckpt['hidden_size'], dropout_prob=ckpt['dropout_prob'])
try:
model.load_state_dict(ckpt['model_state'], strict=True)
except RuntimeError as err:
print("Runtime Error: {0}".format(err))
model.load_state_dict(ckpt['model_state'], strict=False)
parser.add_argument('--n-workers', type=int, default=4, metavar='N', help='Workers for data loading. Default is 4')
parser.add_argument('--model', choices=['vgg', 'resnet', 'densenet'], default='resnet')
parser.add_argument('--save-every', type=int, default=1, metavar='N', help='how many epochs to wait before logging training status. Default is 1')
parser.add_argument('--no-cuda', action='store_true', default=False, help='Disables GPU use')
parser.add_argument('--no-cp', action='store_true', default=False, help='Disables checkpointing')
parser.add_argument('--verbose', type=int, default=1, metavar='N', help='Verbose is activated if > 0')
parser.add_argument('--softmax', choices=['softmax', 'am_softmax'], default='softmax', help='Softmax type')
args = parser.parse_args()
args.cuda = True if not args.no_cuda and torch.cuda.is_available() else False
if args.cuda:
torch.backends.cudnn.benchmark=True
if args.hdf_path:
transform_train = transforms.Compose([transforms.ToPILImage(), transforms.RandomCrop(84, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
trainset = Loader(args.hdf_path, transform_train)
else:
transform_train = transforms.Compose([transforms.RandomCrop(84, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
trainset = datasets.ImageFolder(args.data_path, transform=transform_train)
train_loader = torch.utils.data.DataLoader(trainset, batch_size=args.batch_size, shuffle=True, num_workers=args.n_workers, worker_init_fn=set_np_randomseed, pin_memory=True)
transform_test = transforms.Compose([transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
validset = datasets.ImageFolder(args.valid_data_path, transform=transform_test)
valid_loader = torch.utils.data.DataLoader(validset, batch_size=args.valid_batch_size, shuffle=True, num_workers=args.n_workers, pin_memory=True)
if args.model == 'vgg':
model = vgg.VGG('VGG16', sm_type=args.softmax)
elif args.model == 'resnet':
model = resnet.ResNet18(sm_type=args.softmax)
elif args.model == 'densenet':
def train(lr, l2, max_gnorm, momentum, margin, lambda_, swap, latent_size, n_frames, model, ncoef, epochs, batch_size, valid_batch_size, n_workers, cuda, train_hdf_file, valid_hdf_file, cp_path, softmax, delta, logdir):
if cuda:
device=get_freer_gpu()
if args.model == 'resnet_qrnn':
import cupy
cupy.cuda.Device(int(str(device).split(':')[-1])).use()
cp_name = get_file_name(cp_path)
if args.logdir:
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter(log_dir=logdir+cp_name, comment=args.model, purge_step=True)
else:
writer = None
train_dataset = Loader(hdf5_name = train_hdf_file, max_nb_frames = int(n_frames), delta = delta)
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=n.workers, worker_init_fn=set_np_randomseed)
valid_dataset = Loader_valid(hdf5_name = valid_hdf_file, max_nb_frames = int(n_frames), delta = delta)
valid_loader = torch.utils.data.DataLoader(valid_dataset, batch_size=valid_batch_size, shuffle=True, num_workers=n_workers, worker_init_fn=set_np_randomseed)
if model == 'resnet_mfcc':
model=model_.ResNet_mfcc(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=ncoef, sm_type=softmax, delta=delta)
elif model == 'resnet_34':
model=model_.ResNet_34(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=ncoef, sm_type=softmax, delta=delta)
elif model == 'resnet_lstm':
model=model_.ResNet_lstm(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=ncoef, sm_type=softmax, delta=delta)
elif model == 'resnet_qrnn':
model=model_.ResNet_qrnn(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=ncoef, sm_type=softmax, delta=delta)
elif model == 'resnet_stats':
model=model_.ResNet_stats(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=ncoef, sm_type=softmax, delta=delta)
elif args.model == 'resnet_large':
model = model_.ResNet_large(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=args.ncoef, sm_type=softmax, delta=delta)
elif args.model == 'resnet_small':
model = model_.ResNet_small(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=args.ncoef, sm_type=softmax, delta=delta)
elif args.model == 'resnet_2d':
model = model_.ResNet_2d(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=args.ncoef, sm_type=softmax, delta=delta)
elif args.model == 'TDNN':
model = model_.TDNN(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=args.ncoef, sm_type=softmax, delta=delta)
elif args.model == 'TDNN_att':
model = model_.TDNN_att(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=args.ncoef, sm_type=softmax, delta=delta)
elif args.model == 'TDNN_multihead':
model = model_.TDNN_multihead(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=args.ncoef, sm_type=softmax, delta=delta)
elif args.model == 'TDNN_lstm':
model = model_.TDNN_lstm(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=args.ncoef, sm_type=softmax, delta=delta)
elif args.model == 'TDNN_aspp':
model = model_.TDNN_aspp(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=args.ncoef, sm_type=softmax, delta=delta)
elif args.model == 'TDNN_mod':
model = model_.TDNN_mod(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=args.ncoef, sm_type=softmax, delta=delta)
elif args.model == 'TDNN_multipool':
model = model_.TDNN_multipool(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=args.ncoef, sm_type=softmax, delta=delta)
elif args.model == 'transformer':
model = model_.transformer_enc(n_z=int(latent_size), proj_size=train_dataset.n_speakers, ncoef=args.ncoef, sm_type=softmax, delta=delta)
if cuda:
model=model.to(device)
else:
device=None
optimizer=optim.SGD(model.parameters(), lr=lr, momentum=momentum, weight_decay=l2)
trainer=TrainLoop(model, optimizer, train_loader, valid_loader, max_gnorm=max_gnorm, margin=margin, lambda_=lambda_, verbose=-1, device=device, cp_name=cp_name, save_cp=True, checkpoint_path=cp_path, swap=swap, softmax=True, pretrain=False, mining=True, cuda=cuda, logger=writer)
return trainer.train(n_epochs=epochs)
default=False,
help='Enables PASE updates')
parser.add_argument('--verbose',
type=int,
default=1,
metavar='N',
help='Verbose is activated if > 0')
args = parser.parse_args()
args.cuda = True if not args.no_cuda and torch.cuda.is_available() else False
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed(args.seed)
train_dataset = Loader(hdf5_name=args.train_hdf_file,
max_len=args.max_len,
vad=args.vad,
ncoef=args.ncoef if args.mfcc else None)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
drop_last=True,
num_workers=args.workers,
worker_init_fn=set_np_randomseed)
if args.valid_hdf_file is not None:
valid_dataset = Loader_valid(hdf5_name=args.valid_hdf_file,
max_len=args.max_len,
vad=args.vad,
ncoef=args.ncoef if args.mfcc else None)
valid_loader = torch.utils.data.DataLoader(
valid_dataset,
parser.add_argument('--sgd',
action='store_true',
default=False,
help='enables SGD - *MGD only* ')
parser.add_argument('--job-id',
type=str,
default=None,
help='Arbitrary id to be written on checkpoints')
args = parser.parse_args()
args.cuda = True if not args.no_cuda and torch.cuda.is_available() else False
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed(args.seed)
trainset = Loader(args.data_path)
train_loader = torch.utils.data.DataLoader(trainset,
batch_size=args.batch_size,
num_workers=args.workers)
generator = Generator().train()
disc_list = []
for i in range(args.ndiscriminators):
if args.optimizer == 'adam':
disc = Discriminator(optim.Adam, args.optimizer, args.lr,
(args.beta1, args.beta2)).train()
elif args.optimizer == 'amsgrad':
disc = Discriminator(optim.Adam,
args.optimizer,
if args.cuda:
torch.cuda.manual_seed(args.seed)
if args.mine_triplets:
train_dataset = Loader_mining(hdf5_name=args.train_hdf_file,
max_nb_frames=args.n_frames,
n_cycles=args.n_cycles,
augment=True)
elif args.softmax != 'none':
train_dataset = Loader_softmax(hdf5_name=args.train_hdf_file,
max_nb_frames=args.n_frames,
n_cycles=args.n_cycles,
augment=True)
else:
train_dataset = Loader(hdf5_name=args.train_hdf_file,
max_nb_frames=args.n_frames,
n_cycles=args.n_cycles,
augment=True)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
worker_init_fn=set_np_randomseed)
if args.valid_hdf_file is not None:
valid_dataset = Loader_softmax(hdf5_name=args.valid_hdf_file,
max_nb_frames=args.n_frames,
n_cycles=args.valid_n_cycles,
augment=False)
valid_loader = torch.utils.data.DataLoader(
valid_dataset,
else:
device = None
train_dataset = Loader_mining(hdf5_name=args.train_hdf_file,
max_nb_frames=args.n_frames,
n_cycles=args.n_cycles,
delta=args.delta)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
worker_init_fn=set_np_randomseed)
valid_dataset = Loader(hdf5_name=args.valid_hdf_file,
max_nb_frames=args.n_frames,
n_cycles=args.valid_n_cycles,
delta=args.delta)
valid_loader = torch.utils.data.DataLoader(valid_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
worker_init_fn=set_np_randomseed)
if args.model == 'resnet_mfcc':
model = model_.ResNet_mfcc(n_z=args.latent_size,
proj_size=len(train_dataset.speakers_list) if
args.softmax != 'none' or args.pretrain else 0,
ncoef=args.ncoef,
sm_type=args.softmax,
delta=args.delta)
elif args.model == 'resnet_lstm':
choices=['cnn', 'vgg', 'resnet', 'densenet', 'tdnn'],
default='resnet')
parser.add_argument('--no-cuda',
action='store_true',
default=False,
help='Disables GPU use')
parser.add_argument('--workers',
type=int,
default=4,
metavar='N',
help='Data load workers (default: 4)')
args = parser.parse_args()
args.cuda = True if not args.no_cuda and torch.cuda.is_available(
) else False
testset = Loader(hdf5_name=args.data_path, max_nb_frames=args.n_frames)
test_loader = torch.utils.data.DataLoader(testset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers)
args.nclasses = testset.nclasses
args_dict = parse_args_for_log(args)
print('\n')
for key in args_dict:
print('{}: {}'.format(key, args_dict[key]))
print('\n')
idx_to_class = {}
parser.add_argument('--classifier-path', type=str, default=None, metavar='Path', help='Path to pretrained classifier on MNIST')
parser.add_argument('--data-path', type=str, default='./test.hdf', metavar='Path', help='Path to hdf file containing stacked MNIST. Can be generated with gen_data.py')
parser.add_argument('--out-file', type=str, default='./test_data_coverage.p', metavar='Path', help='files for dumping coverage data')
parser.add_argument('--batch-size', type=int, default=512, metavar='Path', help='batch size')
parser.add_argument('--workers', type=int, help='number of data loading workers', default=4)
parser.add_argument('--no-cuda', action='store_true', default=False, help='Disables GPU use')
args = parser.parse_args()
args.cuda = True if not args.no_cuda and torch.cuda.is_available() else False
if args.classifier_path is None:
raise ValueError('There is no classifier path. Use arg --classifier-path to indicate the path!')
classifier = cnn().eval()
classifier_state = torch.load(args.classifier_path, map_location=lambda storage, loc: storage)
classifier.load_state_dict(classifier_state['model_state'])
if args.cuda:
classifier = classifier.cuda()
print('Cuda Mode is: {}'.format(args.cuda))
testset = Loader(args.data_path)
test_loader = torch.utils.data.DataLoader(testset, batch_size=args.batch_size, num_workers=args.workers)
freqs = compute_freqs_real_data(test_loader, classifier, cuda=args.cuda)
freqs/=freqs.sum()
pfile = open(args.out_file, "wb")
pickle.dump(freqs, pfile)
pfile.close()
