def objective(trial):
config = toml.load(args.config)
lr = 1e-3
#lr = trial.suggest_loguniform('learning_rate', 1e-5, 1e-2)
config['encoder']['activation'] = 'gelu'
#config['block'][0]['stride'] = [trial.suggest_int('stride', 4, 6)]
# C1
config['block'][0]['kernel'] = [
int(trial.suggest_discrete_uniform('c1_kernel', 1, 129, 2))
]
config['block'][0]['filters'] = trial.suggest_int(
'c1_filters', 1, 1024)
# B1 - B5
for i in range(1, 6):
config['block'][i]['repeat'] = trial.suggest_int(
'b%s_repeat' % i, 1, 9)
config['block'][i]['filters'] = trial.suggest_int(
'b%s_filters' % i, 1, 512)
config['block'][i]['kernel'] = [
int(trial.suggest_discrete_uniform('b%s_kernel' % i, 1, 129,
2))
]
# C2
config['block'][-2]['kernel'] = [
int(trial.suggest_discrete_uniform('c2_kernel', 1, 129, 2))
]
config['block'][-2]['filters'] = trial.suggest_int(
'c2_filters', 1, 1024)
# C3
config['block'][-1]['kernel'] = [
int(trial.suggest_discrete_uniform('c3_kernel', 1, 129, 2))
]
config['block'][-1]['filters'] = trial.suggest_int(
'c3_filters', 1, 1024)
model = Model(config)
num_params = sum(p.numel() for p in model.parameters())
print("[trial %s]" % trial.number)
if num_params > args.max_params:
print("[pruned] network too large")
raise optuna.exceptions.TrialPruned()
model.to(args.device)
model.train()
os.makedirs(workdir, exist_ok=True)
optimizer = AdamW(model.parameters(), amsgrad=True, lr=lr)
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O1",
verbosity=0)
schedular = CosineAnnealingLR(optimizer,
args.epochs * len(train_loader))
for epoch in range(1, args.epochs + 1):
try:
train_loss, duration = train(model,
device,
train_loader,
optimizer,
use_amp=True)
val_loss, val_mean, val_median = test(model, device,
test_loader)
print(
"[epoch {}] directory={} loss={:.4f} mean_acc={:.3f}% median_acc={:.3f}%"
.format(epoch, workdir, val_loss, val_mean, val_median))
except KeyboardInterrupt:
exit()
except:
print("[pruned] exception")
raise optuna.exceptions.TrialPruned()
if np.isnan(val_loss): val_loss = 9.9
trial.report(val_loss, epoch)
if trial.should_prune():
print("[pruned] unpromising")
raise optuna.exceptions.TrialPruned()
trial.set_user_attr('seed', args.seed)
trial.set_user_attr('val_loss', val_loss)
trial.set_user_attr('val_mean', val_mean)
trial.set_user_attr('val_median', val_median)
trial.set_user_attr('train_loss', train_loss)
trial.set_user_attr('batchsize', args.batch)
trial.set_user_attr('model_params', num_params)
torch.save(model.state_dict(),
os.path.join(workdir, "weights_%s.tar" % trial.number))
toml.dump(
config,
open(os.path.join(workdir, 'config_%s.toml' % trial.number), 'w'))
print("[loss] %.4f" % val_loss)
return val_loss
def main(args):
workdir = os.path.expanduser(args.training_directory)
if os.path.exists(workdir) and not args.force:
print("[error] %s exists, use -f to force continue training." %
workdir)
exit(1)
init(args.seed, args.device)
device = torch.device(args.device)
print("[loading data]")
chunks, chunk_lengths, targets, target_lengths = load_data(
limit=args.chunks, shuffle=True, directory=args.directory)
split = np.floor(chunks.shape[0] * args.validation_split).astype(np.int32)
train_dataset = ChunkDataSet(chunks[:split], chunk_lengths[:split],
targets[:split], target_lengths[:split])
test_dataset = ChunkDataSet(chunks[split:], chunk_lengths[split:],
targets[split:], target_lengths[split:])
train_loader = DataLoader(train_dataset,
batch_size=args.batch,
shuffle=True,
num_workers=4,
pin_memory=True)
test_loader = DataLoader(test_dataset,
batch_size=args.batch,
num_workers=4,
pin_memory=True)
config = toml.load(args.config)
argsdict = dict(training=vars(args))
chunk_config = {}
chunk_config_file = os.path.join(args.directory, 'config.toml')
if os.path.isfile(chunk_config_file):
chunk_config = toml.load(os.path.join(chunk_config_file))
os.makedirs(workdir, exist_ok=True)
toml.dump({
**config,
**argsdict,
**chunk_config
}, open(os.path.join(workdir, 'config.toml'), 'w'))
print("[loading model]")
model = Model(config)
optimizer = AdamW(model.parameters(), amsgrad=False, lr=args.lr)
last_epoch = load_state(workdir,
args.device,
model,
optimizer,
use_amp=args.amp)
lr_scheduler = func_scheduler(optimizer,
cosine_decay_schedule(1.0, 0.1),
args.epochs * len(train_loader),
warmup_steps=500,
start_step=last_epoch * len(train_loader))
if args.multi_gpu:
from torch.nn import DataParallel
model = DataParallel(model)
model.decode = model.module.decode
model.stride = model.module.stride
model.alphabet = model.module.alphabet
for epoch in range(1 + last_epoch, args.epochs + 1 + last_epoch):
try:
train_loss, duration = train(model,
device,
train_loader,
optimizer,
use_amp=args.amp,
lr_scheduler=lr_scheduler)
val_loss, val_mean, val_median = test(model, device, test_loader)
except KeyboardInterrupt:
break
print(
"[epoch {}] directory={} loss={:.4f} mean_acc={:.3f}% median_acc={:.3f}%"
.format(epoch, workdir, val_loss, val_mean, val_median))
model_state = model.state_dict(
) if not args.multi_gpu else model.module.state_dict()
torch.save(model_state, os.path.join(workdir,
"weights_%s.tar" % epoch))
torch.save(optimizer.state_dict(),
os.path.join(workdir, "optim_%s.tar" % epoch))
with open(os.path.join(workdir, 'training.csv'), 'a',
newline='') as csvfile:
csvw = csv.writer(csvfile, delimiter=',')
if epoch == 1:
csvw.writerow([
'time', 'duration', 'epoch', 'train_loss',
'validation_loss', 'validation_mean', 'validation_median'
])
csvw.writerow([
datetime.today(),
int(duration),
epoch,
train_loss,
val_loss,
val_mean,
val_median,
])
def main(args):
workdir = os.path.expanduser(args.training_directory)
if os.path.exists(workdir) and not args.force:
print("[error] %s exists." % workdir)
exit(1)
init(args.seed, args.device)
device = torch.device(args.device)
print("[loading data]")
chunks, chunk_lengths, targets, target_lengths = load_data(
limit=args.chunks, shuffle=True, directory=args.directory)
split = np.floor(chunks.shape[0] * args.validation_split).astype(np.int32)
train_dataset = ChunkDataSet(chunks[:split], chunk_lengths[:split],
targets[:split], target_lengths[:split])
test_dataset = ChunkDataSet(chunks[split:], chunk_lengths[split:],
targets[split:], target_lengths[split:])
train_loader = DataLoader(train_dataset,
batch_size=args.batch,
shuffle=True,
num_workers=4,
pin_memory=True)
test_loader = DataLoader(test_dataset,
batch_size=args.batch,
num_workers=4,
pin_memory=True)
config = toml.load(args.config)
argsdict = dict(training=vars(args))
chunk_config = {}
chunk_config_file = os.path.join(
args.directory if args.directory else __data__, 'config.toml')
if os.path.isfile(chunk_config_file):
chunk_config = toml.load(os.path.join(chunk_config_file))
print("[loading model]")
model = Model(config)
weights = os.path.join(workdir, 'weights.tar')
if os.path.exists(weights): model.load_state_dict(torch.load(weights))
model.to(device)
model.train()
os.makedirs(workdir, exist_ok=True)
toml.dump({
**config,
**argsdict,
**chunk_config
}, open(os.path.join(workdir, 'config.toml'), 'w'))
optimizer = AdamW(model.parameters(), amsgrad=True, lr=args.lr)
if args.amp:
try:
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O1",
verbosity=0)
except NameError:
print(
"[error]: Cannot use AMP: Apex package needs to be installed manually, See https://github.com/NVIDIA/apex"
)
exit(1)
schedular = CosineAnnealingLR(optimizer, args.epochs * len(train_loader))
for epoch in range(1, args.epochs + 1):
try:
train_loss, duration = train(model,
device,
train_loader,
optimizer,
use_amp=args.amp)
val_loss, val_mean, val_median = test(model, device, test_loader)
except KeyboardInterrupt:
break
print(
"[epoch {}] directory={} loss={:.4f} mean_acc={:.3f}% median_acc={:.3f}%"
.format(epoch, workdir, val_loss, val_mean, val_median))
torch.save(model.state_dict(),
os.path.join(workdir, "weights_%s.tar" % epoch))
with open(os.path.join(workdir, 'training.csv'), 'a',
newline='') as csvfile:
csvw = csv.writer(csvfile, delimiter=',')
if epoch == 1:
csvw.writerow([
'time', 'duration', 'epoch', 'train_loss',
'validation_loss', 'validation_mean', 'validation_median'
])
csvw.writerow([
datetime.today(),
int(duration),
epoch,
train_loss,
val_loss,
val_mean,
val_median,
])
schedular.step()
def main(args):
model = Model(toml.load(args.config))
print(model)
print("Total parameters in model",
sum(p.numel() for p in model.parameters()))