def train(args):
assert args.num_classes
common.make_dir(args.checkout_dir)
nnet = DNN((args.left_context + args.right_context + 1) * args.feat_dim, hidden_layer, \
hidden_size, args.num_classes, dropout=dropout)
print(nnet)
nnet.cuda()
criterion = nn.CrossEntropyLoss()
optimizer = th.optim.Adam(nnet.parameters(), lr=args.learning_rate)
train_dataset = THCHS30(root=args.data_dir, data_type='train', left_context=left_context,
right_context=right_context, model_type='dnn')
train_loader = data.DataLoader(dataset=train_dataset, batch_size=args.min_batch,
shuffle=True, num_workers=6)
test_dataset = THCHS30(root=args.data_dir, data_type='test', left_context=left_context,
right_context=right_context, model_type='dnn')
test_loader = data.DataLoader(dataset=test_dataset, batch_size=args.min_batch,
shuffle=True, num_workers=6)
cross_validate(-1, nnet, test_loader, test_dataset.num_frames)
for epoch in range(args.num_epochs):
common.train_one_epoch(nnet, criterion, optimizer, train_loader)
cross_validate(epoch, nnet, test_loader, test_dataset.num_frames)
th.save(nnet, common.join_path(args.checkout_dir, 'dnn.{}.pkl'.format(epoch + 1)))
def run():
df = pd.read_csv(config.TRAIN_PATH)
kfold = KFold(n_splits=5, random_state=config.SEED, shuffle=True)
fold_losses = []
for i, (train_idx, val_idx) in enumerate(kfold.split(df)):
print("-------------------------------------------------------")
print(f"Training fold {i}")
print("-------------------------------------------------------")
train = df.iloc[train_idx]
validation = df.iloc[val_idx]
train_dataset = PicDataset(train)
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=config.BATCH_SIZE
)
val_dataset = PicDataset(validation)
val_data_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=config.BATCH_SIZE
)
device = 'cuda:0' if torch.cuda.is_available() else "cpu"
model = DNN()
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=config.LR)
loss = 0
for _ in range(config.EPOCHS):
engine.train_fn(train_data_loader, model, optimizer, device)
loss = engine.eval_fn(val_data_loader, model, device)
print(f"Loss on fold {i} is {loss}")
fold_losses.append(loss)
torch.save(model.state_dict(), f'./models/model_{i}.bin')
print(f"Average loss on cross validation is {sum(fold_losses) / 5}")
weight = torch.FloatTensor(list(weight.values())).to(device)
data = sorted(corpus.examples.get('seq'), key=lambda x: len(x), reverse=True)
vocab_size = len(corpus.words2id)
logging.info('vocabulary size: {}'.format(vocab_size))
model = DNN(vocab_size=vocab_size,
embedding_size=200,
hidden_size=512,
embedding=embedding)
model.to(device)
loss_function = nn.CrossEntropyLoss(weight=weight)
optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.9)
# optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
model.train()
total_data = len(data)
batch_size = args['batch_size']
total_step = math.ceil(total_data / batch_size)
last_training_loss = 1000000000000
for epoch in range(args.get('epoch')):
start = 0
training_loss = 0
for _ in tqdm(range(int(total_step)), total=total_step):
batch = data[start:start + batch_size]
start += batch_size
def train(args, config, io):
train_loader, validation_loader = get_loader(args, config)
device = torch.device("cuda" if args.cuda else "cpu")
# print(len(train_loader), len(validation_loader))
#Try to load models
model = DNN(args).to(device)
"""if device == torch.device("cuda"):
model = nn.DataParallel(model)"""
if args.model_path != "":
model.load_state_dict(torch.load(args.model_path))
# for para in list(model.parameters())[:-5]:
# para.requires_grad=False
# print(model)
if args.use_sgd:
# print("Use SGD")
opt = optim.SGD(model.parameters(),
lr=args.lr * 100,
momentum=args.momentum,
weight_decay=1e-4)
else:
# print("Use Adam")
opt = optim.Adam(model.parameters(), lr=args.lr, weight_decay=1e-4)
"""opt = optim.Adam([
{'params': list(model.parameters())[:-1], 'lr':args.lr/50, 'weight_decay': 1e-4},
{'params': list(model.parameters())[-1], 'lr':args.lr, 'weight_decay': 1e-4}
])
"""
scheduler = CosineAnnealingLR(opt, args.epochs, eta_min=args.lr)
criterion = nn.MSELoss()
best_test_loss = 9999999.
for epoch in range(args.epochs):
startTime = time.time()
####################
# Train
####################
train_loss = 0.0
train_dis = 0.0
count = 0.0
model.train()
for data, label in train_loader:
data, label = data.to(device), label.to(device)
data = drop(jitter(data, device), device)
# data = jitter(data, device, delta=0.05)
batch_size = data.shape[0]
logits = model(data)
loss = criterion(logits, label)
opt.zero_grad()
loss.backward()
opt.step()
dis = distance(logits, label)
count += batch_size
train_loss += loss.item() * batch_size
train_dis += dis.item() * batch_size
scheduler.step()
outstr = 'Train %d, loss: %.6f, distance: %.6f' % (
epoch, train_loss * 1.0 / count, train_dis * 1.0 / count)
io.cprint(outstr)
####################
# Evaluation
####################
test_loss = 0.0
test_dis = 0.0
count = 0.0
model.eval()
with torch.no_grad():
for data, label in validation_loader:
data, label = data.to(device), label.to(device)
batch_size = data.shape[0]
logits = model(data)
loss = criterion(logits, label)
dis = distance(logits, label)
count += batch_size
test_loss += loss.item() * batch_size
test_dis += dis.item() * batch_size
outstr = 'Test %d, loss: %.6f, distance: %.6f' % (
epoch, test_loss * 1.0 / count, test_dis * 1.0 / count)
io.cprint(outstr)
if test_loss <= best_test_loss:
best_test_loss = test_loss
torch.save(model.state_dict(),
'checkpoints/%s/models/model.t7' % args.exp_name)
torch.save(model, (config.root + config.model_path))
io.cprint('Time: %.3f sec' % (time.time() - startTime))
pass
try:
os.makedirs(args.log)
except OSError:
pass
# 加载训练数据和测试数据
train_loader, test_loader = get_data(args)
# 如果有cuda就用cuda
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
input_size = 28 * 28
output_size = args.num_classes
model = DNN(input_size=input_size, output_size=output_size).to(device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
# 每训练完一个mini-batch就计算一次loss，acc
model.train()
for epoch in range(args.epochs):
correct = 0
total = 0
for idx, (x, y) in enumerate(train_loader):
x, y = x.to(device), y.to(device)
y_pred = model(x)
_, y_pred_t = torch.max(y_pred.data, 1)
total += y.size(0)
# print(y_pred_t, y.data)
# print((y_pred_t == y).sum().item())
def train():
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
data_dict, topic_dict = dh.load_data(
) # data_dict, [group2topic, mem2topic]
train_data, train_label, dev_data, dev_label, test_data, test_label = dh.data_split(
data_dict, topic_dict)
train_dataset = dh.Dataset(train_data, train_label)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True)
dev_dataset = dh.Dataset(dev_data, dev_label)
dev_loader = DataLoader(dev_dataset, batch_size=128, shuffle=True)
lambda1 = lambda epoch: (
epoch / args.warm_up_step
) if epoch < args.warm_up_step else 0.5 * (math.cos(
(epoch - args.warm_up_step) /
(args.n_epoch * len(train_dataset) - args.warm_up_step) * math.pi) + 1)
model = DNN(args).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=args.init_lr)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer,
len(train_loader) * args.n_epoch)
global_step = 0
best_f1 = 0.
loss_deq = collections.deque([], args.report_step)
for epoch in range(args.n_epoch):
for batch in tqdm(train_loader):
optimizer.zero_grad()
inputs = batch['input'].to(device)
group_topic = batch['group_topic'].to(device)
mem_topic = batch['mem_topic'].to(device)
labels = batch['label'].to(device)
output = model(inputs, mem_topic, group_topic, label=labels)
loss = output[0]
loss.backward()
loss_deq.append(loss.item())
optimizer.step()
scheduler.step()
global_step += 1
if global_step % args.report_step == 0:
logger.info('loss: {}, lr: {}, epoch: {}'.format(
np.average(loss_deq).item(),
optimizer.param_groups[0]['lr'],
global_step / len(train_dataset)))
if global_step % args.eval_step == 0:
model.eval()
eval_result = evaluation(model,
data_loader=dev_loader,
device=device)
logger.info(eval_result)
if eval_result['f1'] > best_f1:
torch.save(model,
'./model/{}/torch.pt'.format(args.task_name))
best_f1 = eval_result['f1']
model.train()
for w in st:
tmp.add(w)
word_index = {w: i for i, w in enumerate(tmp)}
#import pickle
#with open("bow.pkl", "wb") as f:
# pickle.dump(word_index, f)
x = torch.zeros(len(train_x1), len(word_index))
for i in range(len(train_x1)):
for w in train_x1[i]:
x[i][word_index[w]] += 1
print(x.size())
print("\nConstructing model...", flush=True)
model = DNN(x.size(1)).to(device)
total_param = sum(p.numel() for p in model.parameters())
trainable_param = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print("{} parameters with {} trainable".format(total_param,
trainable_param),
flush=True)
print("\nStart training...", flush=True)
train_dataset1 = TwitterDataset(x, train_y1)
train_loader1 = torch.utils.data.DataLoader(dataset=train_dataset1,
batch_size=BATCH,
shuffle=True,
num_workers=4)
train_model(train_loader1, model, device, LR)
print("\nStart testing...", flush=True)
def train(args, config, io):
train_loader, validation_loader, unlabelled_loader = get_loader(
args, config)
device = torch.device("cuda" if args.cuda else "cpu")
#Try to load models
model = DNN(args).to(device)
ema_model = DNN(args).to(device)
for param in ema_model.parameters():
param.detach_()
if device == torch.device("cuda"):
model = nn.DataParallel(model)
ema_model = nn.DataParallel(ema_model)
if args.model_path != "":
model.load_state_dict(torch.load(args.model_path))
ema_model.load_state_dict(torch.load(args.model_path))
if args.use_sgd:
print("Use SGD")
opt = optim.SGD(model.parameters(),
lr=args.lr * 100,
momentum=args.momentum,
weight_decay=1e-4)
else:
print("Use Adam")
opt = optim.Adam(model.parameters(), lr=args.lr, weight_decay=1e-4)
scheduler = CosineAnnealingLR(opt, args.epochs, eta_min=args.lr)
criterion = nn.MSELoss()
consistency_criterion = nn.MSELoss()
best_test_loss = 9999999.
global_step = 0
for epoch in range(args.epochs):
startTime = time.time()
####################
# Train
####################
train_loss = 0.0
count = 0.0
model.train()
ema_model.train()
i = -1
for (data, label), (u, _) in zip(cycle(train_loader),
unlabelled_loader):
i = i + 1
if data.shape[0] != u.shape[0]:
bt_size = np.minimum(data.shape[0], u.shape[0])
data = data[0:bt_size]
label = label[0:bt_size]
u = u[0:bt_size]
data, label, u = data.to(device), label.to(device), u.to(device)
batch_size = data.shape[0]
logits = model(data)
class_loss = criterion(logits, label)
u_student = jitter(u, device)
u_teacher = jitter(u, device)
logits_unlabeled = model(u_student)
ema_logits_unlabeled = ema_model(u_teacher)
ema_logits_unlabeled = Variable(ema_logits_unlabeled.detach().data,
requires_grad=False)
consistency_loss = consistency_criterion(logits_unlabeled,
ema_logits_unlabeled)
if epoch < args.consistency_rampup_starts:
consistency_weight = 0.0
else:
consistency_weight = get_current_consistency_weight(
args, args.final_consistency, epoch, i,
len(unlabelled_loader))
consistency_loss = consistency_weight * consistency_loss
loss = class_loss + consistency_loss
opt.zero_grad()
loss.backward()
opt.step()
global_step += 1
# print(global_step)
update_ema_variables(model, ema_model, args.ema_decay, global_step)
count += batch_size
train_loss += loss.item() * batch_size
scheduler.step()
outstr = 'Train %d, loss: %.6f' % (epoch, train_loss * 1.0 / count)
io.cprint(outstr)
####################
# Evaluation
####################
test_loss = 0.0
count = 0.0
model.eval()
ema_model.eval()
for data, label in validation_loader:
data, label = data.to(device), label.to(device)
batch_size = data.shape[0]
logits = ema_model(data)
loss = criterion(logits, label)
count += batch_size
test_loss += loss.item() * batch_size
outstr = 'Test %d, loss: %.6f' % (epoch, test_loss * 1.0 / count)
io.cprint(outstr)
if test_loss <= best_test_loss:
best_test_loss = test_loss
torch.save(ema_model.state_dict(),
'checkpoints/%s/models/model.t7' % args.exp_name)
torch.save(ema_model, (config.root + config.model_path))
io.cprint('Time: %.3f sec' % (time.time() - startTime))
def main():
print('> Starting execution...')
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
group = parser.add_mutually_exclusive_group()
group.add_argument('--fit',
action='store_true',
help='fit the tuned model on digits 0-4')
group.add_argument('--transfer',
action='store_true',
help='train a pretrained model on digits 5-9')
parser.add_argument('--batch-size',
type=int,
default=256,
metavar='N',
help='input batch size for training (default: 256)')
parser.add_argument('--epochs',
type=int,
default=50,
metavar='E',
help='number of epochs to train (default: 50)')
parser.add_argument('--lr',
type=float,
default=1e-3,
metavar='L',
help='learning rate (default: 1e-3)')
parser.add_argument('--early-stopping',
type=int,
default=7,
metavar='E',
help='early stopping (default: 7 epochs)')
parser.add_argument(
'--size',
type=int,
default=100,
metavar='S',
help='size of the training data for transfer learning (default: 100)')
parser.add_argument('--seed',
type=int,
default=23,
metavar='S',
help='random seed (default: 23)')
args = parser.parse_args()
use_cuda = torch.cuda.is_available() # use cuda if available
device = torch.device("cuda" if use_cuda else "cpu")
torch.manual_seed(args.seed) # random seed
print('> Loading MNIST data')
train_set = datasets.MNIST(MNIST_DATA_DIR,
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
]))
test_set = datasets.MNIST(MNIST_DATA_DIR,
train=False,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
]))
train_digits_04 = np.where(train_set.train_labels < 5)[0]
train_digits_59 = np.where(train_set.train_labels > 4)[0]
test_digits_04 = np.where(test_set.test_labels < 5)[0]
test_digits_59 = np.where(test_set.test_labels > 4)[0]
if args.fit:
# Training the tuned model on digits 0-4
print('> Training a new model on MNIST digits 0-4')
X_train_04, y_train_04, X_valid_04, y_valid_04 = data_to_numpy(
train_set, test_set, INPUT_DIM, train_digits_04, test_digits_04)
torch.manual_seed(args.seed)
print('> Initializing the model')
model = DNN(INPUT_DIM, OUTPUT_DIM, HIDDEN_DIM, batch_norm=True)
model.apply(init_he_normal) # He initialization
model = model.to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=args.lr)
print('> Training the model')
model, _, _ = train_model(model,
device,
X_train_04,
y_train_04,
criterion,
optimizer,
X_valid=X_valid_04,
y_valid=y_valid_04,
batch_size=args.batch_size,
n_epochs=args.epochs,
early_stopping=args.early_stopping)
print(f'> Saving the model state at {MODEL_04_PATH}')
torch.save(model.state_dict(), MODEL_04_PATH)
elif args.transfer:
# Transfer learning
print(
'> Training a model on MNIST digits 5-9 from a pretrained model for digits 0-4'
)
if os.path.isfile(MODEL_04_PATH):
print('> Loading the pretrained model')
model = DNN(INPUT_DIM, OUTPUT_DIM, HIDDEN_DIM,
batch_norm=True).to(device)
model.load_state_dict(torch.load(MODEL_04_PATH))
for param in model.parameters():
param.requires_grad = False
# Parameters of newly constructed modules have requires_grad=True by default
model.fc4 = nn.Linear(HIDDEN_DIM, HIDDEN_DIM)
model.fc5 = nn.Linear(HIDDEN_DIM, HIDDEN_DIM)
model.out = nn.Linear(HIDDEN_DIM, OUTPUT_DIM)
print('> Using saved model state')
else:
print(
'> Model state file is not found, fit a model before the transfer learning'
)
print('> Stopping execution')
return
X_train_59, y_train_59, X_valid_59, y_valid_59 = data_to_numpy(
train_set, test_set, INPUT_DIM, train_digits_59[:args.size],
test_digits_59)
# fixing the issues with labels
y_train_59 = y_train_59 - 5
y_valid_59 = y_valid_59 - 5
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=args.lr)
print('> Training the model')
model, _, _ = train_model(model,
device,
X_train_59,
y_train_59,
criterion,
optimizer,
X_valid=X_valid_59,
y_valid=y_valid_59,
batch_size=args.batch_size,
n_epochs=args.epochs,
early_stopping=args.early_stopping)
print(f'> Saving the model state at {MODEL_59_PATH}')
torch.save(model.state_dict(), MODEL_59_PATH)
else:
print('> Incorrect mode, try either `--fit` or `--transfer`')
print('> Stopping execution')
model = DNN(input_size, hidden_size, out_size)
model = torch.nn.DataParallel(model.to(device), device_ids=use_devices) # use GPU
print(model)
# load data
print('loading data...')
sample_generator = AudioSampleGenerator(os.path.join(in_path, ser_data_fdr))
random_data_loader = DataLoader(
dataset=sample_generator,
batch_size=batch_size, # specified batch size here
shuffle=True,
num_workers=1,
drop_last=True, # drop the last batch that cannot be divided by batch_size
pin_memory=True)
print('DataLoader created')
#optimizer = optim.Adam(model.parameters(), lr=lr, betas=(0.5, 0.999))
optimizer = optim.Adam(model.parameters(), lr=1e-4, betas=(0.9, 0.999), eps=1e-08, weight_decay=0)
# create tensorboard writer
# The logs will be stored NOT under the run_time, but under segan_data_out/'tblog_fdr'.
# This way, tensorboard can show graphs for each experiment in one board
tbwriter = SummaryWriter(log_dir=tblog_path)
print('TensorboardX summary writer created')
print('Starting Training...')
total_steps = 1
MSE = nn.MSELoss()
scaler_path_input = os.path.join(scaler_dir, "scaler_input.p")
scaler_input = pickle.load(open(scaler_path_input, 'rb'))
scaler_path_label = os.path.join(scaler_dir, "scaler_label.p")
scaler_label = pickle.load(open(scaler_path_label, 'rb'))
for epoch in range(num_epochs):