def main(args):
speaker_cfg = args.speaker_cfg
speech_cfg = args.speech_cfg
args_speaker = read_conf(speaker_cfg, deepcopy(args))
args_speaker.model_path = args.speaker_model
args_speech = read_conf(speech_cfg, deepcopy(args))
args_speech.model_path = args.speech_model
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
print("set seed: ", args_speaker.optimization.seed)
torch.manual_seed(args_speaker.optimization.seed)
np.random.seed(args_speaker.optimization.seed)
random.seed(args_speaker.optimization.seed)
torch.cuda.set_device(args.local_rank)
if not args.no_dist:
torch.distributed.init_process_group(backend="nccl")
train_dataset = TIMIT_speaker(args.data_root,
train=True,
phoneme=True,
norm_factor=True)
test_dataset = TIMIT_speaker(args.data_root,
train=False,
phoneme=True,
norm_factor=True)
pretrained_models = get_pretrained_models(args_speaker, args_speech)
loss_factors = {
"speaker": args.speaker_factor,
"speech": args.speech_factor,
"norm": args.norm_factor
}
if args.target < 0: # non-targeted
speaker_loss = SpeakerLoss(pretrained_models['speaker'])
else: # targeted attack
speaker_loss = SpeakerLossTarget(pretrained_models['speaker'],
args.target)
loss_all = {}
loss_all['speech'] = {
'model':
pretrained_models['speech'],
'factor':
loss_factors['speech'],
'loss_func':
SpeechLoss(pretrained_models['speech'],
factor_kld=args.speech_kld_factor)
}
loss_all['speaker'] = {
'model': pretrained_models['speaker'],
'factor': loss_factors['speaker'],
'loss_func': speaker_loss
}
loss_all['norm'] = {
'loss_func': MSEWithThreshold(args.norm_clip),
'factor': loss_factors['norm']
}
cost = AdversarialLoss(loss_all)
model = SpeechTransformer(args.channel, args.kernel_size, args.dilation,
args.sample, args.noise_scale)
if args.pt_file != 'none':
print("load model from:", args.pt_file)
if os.path.splitext(args.pt_file)[1] == '.pkl':
checkpoint_load = torch.load(args.pt_file)
model.load_raw_state_dict(checkpoint_load)
else:
load_checkpoint(model, args.pt_file)
model = model.cuda()
if args.eval:
assert args.pt_file != 'none', "no pretrained model is provided!"
print('only eval the model')
evaluate(model, test_dataset, cost)
return
if args.test:
assert args.pt_file != 'none', "no pretrained model is provided!"
print("only test the model")
filename_list = open("./data/TIMIT/speaker/test.scp", 'r').readlines()
filename_list = [_f.strip() for _f in filename_list]
label_dict = np.load(
os.path.join(args.data_root, "processed",
"TIMIT_labels.npy")).item()
test_wav(model, filename_list, args.data_root,
os.path.join(args.data_root, "output"),
pretrained_models['speaker'], label_dict, args.target)
return
if args.cpu_test:
assert args.pt_file != 'none', "no pretrained model is provided!"
print("only cpu test the model")
filename_list = open("./data/TIMIT/speaker/test.scp", 'r').readlines()
filename_list = [_f.strip() for _f in filename_list]
label_dict = np.load(
os.path.join(args.data_root, "processed",
"TIMIT_labels.npy")).item()
test_wav_cpu(model, filename_list, args.data_root,
os.path.join(args.data_root, "output"),
pretrained_models['speaker'], label_dict, args.target)
return
print("train the model")
batch_process = batch_process_speaker
eval_hook = EvalHook()
optimizer = optim.Adam(model.parameters(),
lr=args_speaker.optimization.lr,
betas=(0.95, 0.999))
lr_scheduler = optim.lr_scheduler.StepLR(optimizer, 2, 0.5)
if args.no_dist:
kwarg = {
'shuffle':
True,
'worker_init_fn':
partial(_init_fn, seed=args_speaker.optimization.seed)
}
else:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
kwarg = {
'sampler':
train_sampler,
'worker_init_fn':
partial(_init_fn, seed=args_speaker.optimization.seed)
}
train_dataloader = DataLoader(train_dataset,
args_speaker.optimization.batch_size,
num_workers=args.num_workers,
pin_memory=True,
**kwarg)
test_dataloader = DataLoader(test_dataset,
args_speaker.optimization.batch_size,
shuffle=False,
num_workers=args.num_workers,
pin_memory=True)
trainer = ClassifierTrainer(
model,
train_dataloader,
optimizer,
cost,
batch_process,
args.output_dir,
0,
test_dataloader,
eval_hook=eval_hook,
eval_every=args_speaker.optimization.N_eval_epoch,
print_every=args_speaker.optimization.print_every,
lr_scheduler=lr_scheduler)
trainer.logger.info(args)
trainer.run(args_speaker.optimization.N_epochs)
model_list = [
dict(type='Conv2D',
in_channels=3,
out_channels=32,
kernel_size=5,
stride=1,
padding=2),
dict(type='ReLU'),
dict(type='MaxPooling', kernel_size=2, stride=2),
dict(type='Linear', in_dim=8192, out_dim=10)
]
criterion = dict(type='SoftmaxCrossEntropy')
model = ConvNet(model_list, criterion)
optimizer = SGD(model, learning_rate=0.0001, reg=0.001, momentum=0.9)
trainer = ClassifierTrainer()
loss_history, train_acc_history = trainer.train(X_train[:50],
y_train[:50],
model,
batch_size=10,
num_epochs=10,
verbose=True,
optimizer=optimizer)
plt.plot(train_acc_history)
plt.legend(['train', 'val'], loc='upper left')
plt.xlabel('epoch')
plt.ylabel('accuracy')
plt.savefig('train.png')
def main(args):
# init for distributed training
print("rank:{}, init dataset".format(args.local_rank))
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(
backend="nccl"
) #, rank=0, world_size=1, init_method="tcp://127.0.0.1:29500")
if args.dataset == "birds":
data_root = "./data/birds"
train_dataset = BirdDataset(data_root=data_root, train=True)
val_dataset = BirdDataset(data_root=data_root, train=False)
elif args.dataset == "places":
data_root = "./data/Places_subset"
train_dataset = PlaceSubSet(data_root=data_root, train=True)
val_dataset = PlaceSubSet(data_root=data_root, train=False)
elif args.dataset == "flowers":
data_root = "./data/flowers"
train_dataset = FlowerDataset(data_root=data_root, train=True)
val_dataset = FlowerDataset(data_root=data_root, train=False)
else:
raise NotImplementedError
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
train_dataloader = DataLoader(train_dataset,
batch_size=args.batch_size,
num_workers=args.num_workers,
sampler=train_sampler)
val_dataloader = DataLoader(val_dataset, batch_size=args.batch_size)
# define model
print("rank:{}, define model".format(args.local_rank))
model = CNNRNN(40,
embedding_dim=1024,
drop_prob=args.dropout,
nhidden=1024,
nsent=1024,
bidirectional=args.bidirectional,
rnn_layers=args.rnn_layers)
batch_param = {'length_required': True, 'sorted_required': True}
model = model.cuda()
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank)
# optim
optimizer = torch.optim.Adam(model.parameters(),
lr=args.learning_rate,
weight_decay=1e-5)
lr_scheduler = torch.optim.lr_scheduler.StepLR(
optimizer,
step_size=args.lr_scheduler_step_size,
gamma=args.lr_scheduler_gamma)
# loss
loss_func = LossFunc(args.loss_diff, args.loss_same, args.jel_flag,
args.l1_flag, args.lambda_l1, args.distill_flag,
args.distill_T, args.lambda_distill)
eval_hook = EvalHook()
test_func = EvalClass()
print("rank:{}, define trainer".format(args.local_rank))
trainer = ClassifierTrainer(model=model,
train_dataloader=train_dataloader,
val_dataloader=val_dataloader,
optimizer=optimizer,
loss_func=loss_func,
batch_process=batch_process,
output_dir=args.output_dir,
local_rank=args.local_rank,
print_every=args.print_every,
eval_every=args.eval_every,
lr_scheduler=lr_scheduler,
eval_hook=eval_hook,
test_func=test_func,
batch_param=batch_param,
resume_from=args.resume)
trainer.run(args.epoch)
process = Preprocessor()
# Dataset
train_data = TrainDataset(process.get_train())
val_data = TrainDataset(process.get_val())
test_data = TestDataset(process.get_test())
# Dataloader
train_loader = DataLoader(train_data, batch_size=32, shuffle=True, drop_last=True)
val_loader = DataLoader(val_data, batch_size=100)
test_loader = DataLoader(test_data, batch_size=16)
model = Classifier().cuda()
# Training
trainers = ClassifierTrainer(model, train_loader, val_loader,
metric=balanced_accuracy_score)
trainers.train(max_epoch=5, patience=1)
# Visualization
trainers.plot_cm()
trainers.plot_metric_curve()
trainers.plot_loss_curve()
# Evaluation
trainers.model.eval()
pred = []
for x in test_loader:
pred.append(trainers.model.predict(x.float().cuda()))
parser.add_argument('-i', '--input_dir', type=str,
default='GTSRB-new\RB\val\0-clear',
help="input image path")
parser.add_argument('-o', '--output_dir', type=str, default='result-2/GTSRB/original/',
help="output image path")
parser.add_argument('-p', '--checkpoint', type=str, default='checkpoints-new-2/0-0.2/outputs/GTSRB/checkpoints/classifier.pt',
help="checkpoint")
parser.add_argument('-l', '--log_name', type=str, default='0-0.2.log', help="log name")
parser.add_argument('-g', '--gpu_id', type=int, default=0, help="gpu id")
opts = parser.parse_args()
# Load experiment setting
config = get_config(opts.config)
# Setup model and data loader
trainer = ClassifierTrainer(config)
state_dict = torch.load(opts.checkpoint, map_location='cuda:{}'.format(opts.gpu_id))
trainer.net.load_state_dict(state_dict['net'])
epochs = state_dict['epochs']
min_loss = state_dict['min_loss']
acc = state_dict['acc'] if 'acc' in state_dict.keys() else 0.0
print("=" * 40)
print('Resume from epoch: {}, min-loss: {} acc: {}'.format(epochs, min_loss, acc))
print("=" * 40)
trainer.cuda()
trainer.eval()
pred_acc_list = []
def main(args):
args = read_conf(args.cfg, args)
torch.manual_seed(args.seed)
np.random.seed(args.seed)
if args.dataset == 'timit':
train_dataset = TIMIT(data_root=args.data_root,
datalist_root=args.datalist_root,
train=True,
oversampling=args.oversampling)
test_dataset = TIMIT(data_root=args.data_root,
datalist_root=args.datalist_root,
train=False)
elif args.dataset == 'libri':
raise NotImplementedError
else:
raise NotImplementedError
cost = nn.NLLLoss()
CNN_arch = {
'input_dim': train_dataset.wlen,
'fs': args.fs,
'cnn_N_filt': args.cnn_N_filt,
'cnn_len_filt': args.cnn_len_filt,
'cnn_max_pool_len': args.cnn_max_pool_len,
'cnn_use_laynorm_inp': args.cnn_use_laynorm_inp,
'cnn_use_batchnorm_inp': args.cnn_use_batchnorm_inp,
'cnn_use_laynorm': args.cnn_use_laynorm,
'cnn_use_batchnorm': args.cnn_use_batchnorm,
'cnn_act': args.cnn_act,
'cnn_drop': args.cnn_drop,
}
DNN1_arch = {
'fc_lay': args.fc_lay,
'fc_drop': args.fc_drop,
'fc_use_batchnorm': args.fc_use_batchnorm,
'fc_use_laynorm': args.fc_use_laynorm,
'fc_use_laynorm_inp': args.fc_use_laynorm_inp,
'fc_use_batchnorm_inp': args.fc_use_batchnorm_inp,
'fc_act': args.fc_act,
}
DNN2_arch = {
'input_dim': args.fc_lay[-1],
'fc_lay': args.class_lay,
'fc_drop': args.class_drop,
'fc_use_batchnorm': args.class_use_batchnorm,
'fc_use_laynorm': args.class_use_laynorm,
'fc_use_laynorm_inp': args.class_use_laynorm_inp,
'fc_use_batchnorm_inp': args.class_use_batchnorm_inp,
'fc_act': args.class_act,
}
model = SpeakerIDNet(CNN_arch, DNN1_arch, DNN2_arch)
if args.pt_file != '':
print("load model from:", args.pt_file)
checkpoint_load = torch.load(args.pt_file)
ext = os.path.splitext(args.pt_file)[1]
if ext == '.pkl':
model.load_raw_state_dict(checkpoint_load)
elif ext == '.pickle':
model.load_state_dict(checkpoint_load)
elif ext == '.pth':
load_checkpoint(model, args.pt_file)
else:
raise NotImplementedError
model = model.cuda()
if args.eval:
print('only eval the model')
evaluate(model, test_dataset, cost)
return
else:
print("train the model")
optimizer = optim.RMSprop(model.parameters(),
lr=args.lr,
alpha=0.95,
eps=1e-8)
train_dataloader = DataLoader(train_dataset,
args.batch_size,
shuffle=True,
num_workers=8,
pin_memory=True)
test_dataloader = DataLoader(test_dataset,
1,
shuffle=False,
num_workers=8,
pin_memory=True)
trainer = ClassifierTrainer(model,
train_dataloader,
optimizer,
cost,
batch_process,
args.output_dir,
0,
test_dataloader,
eval_every=args.N_eval_epoch,
print_every=args.print_every)
trainer.run(args.N_epochs)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--batch_size',
help='Batch size',
required=False,
type=int,
default=256)
parser.add_argument('--pickle_file',
help='Pickle file',
required=False,
default='./traffic.pickle')
parser.add_argument('--learning_rate',
help='Learning rate',
required=False,
type=float,
default=0.0001)
parser.add_argument('--image_size',
help='Image size',
required=False,
type=int,
default=32)
parser.add_argument('--num_classes',
help="Number of classes",
required=False,
type=int,
default=43)
parser.add_argument('--color_channels',
help="Color channels",
required=False,
type=int,
default=1)
parser.add_argument('--ckpt_dir',
help="Check point directory",
required=False,
default='./modelSave/')
parser.add_argument('--num_iter_per_epoch',
help="Number of iterations per epoch",
required=False,
type=int,
default=1)
parser.add_argument('--num_epochs',
help="Number of epochs",
required=False,
type=int,
default=1)
parser.add_argument('--load_model',
help="Load model or train model from scratch",
required=False,
type=bool,
default=True)
args = parser.parse_args()
class config:
batch_size = args.batch_size
pickle_file = args.pickle_file
learning_rate = args.learning_rate
image_size = args.image_size
num_classes = args.num_classes
num_channels = args.color_channels
num_iter_per_epoch = args.num_iter_per_epoch
num_epochs = args.num_epochs
checkpoint_dir = args.ckpt_dir
load_model = args.load_model
sess = tf.Session()
data_loader = DataLoader(config=config)
model = ClassifierModel(data_loader=data_loader, config=config)
trainer = ClassifierTrainer(sess=sess,
model=model,
config=config,
logger=None,
dataLoader=data_loader)
trainer.train()
return