def main():
train_dataset, val_dataset = create_datasets(Config.DATASET_DIR)
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=Config.BATCH_SIZE,
num_workers=Config.DATALOADER_WORKER_NUM,
shuffle=True,
collate_fn=my_collate_fn)
val_dataloader = torch.utils.data.DataLoader(
val_dataset,
batch_size=Config.BATCH_SIZE,
num_workers=Config.DATALOADER_WORKER_NUM,
shuffle=False,
collate_fn=my_collate_fn)
model = Net()
optimizer = torch.optim.Adam(params=model.parameters(),
lr=Config.LEARNING_RATE,
weight_decay=Config.WEIGHT_DECAY)
trainer = Trainer(optimizer,
model,
train_dataloader,
val_dataloader,
resume=Config.RESUME_FROM,
log_dir=Config.LOG_DIR,
persist_stride=Config.MODEL_SAVE_STRIDE)
trainer.train()
def train(args):
dataset_dir = get_dataset_dir(args)
log_dir = get_log_dir(args)
model_class = get_model_class(args)
training_set, validation_set, num_classes = create_datasets(dataset_dir)
training_dataset = Dataset(training_set,
transform_for_training(model_class.IMAGE_SHAPE))
validation_dataset = Dataset(validation_set,
transform_for_infer(model_class.IMAGE_SHAPE))
training_dataloader = torch.utils.data.DataLoader(
training_dataset,
batch_size=args.batch_size,
num_workers=6,
shuffle=True)
validation_dataloader = torch.utils.data.DataLoader(
validation_dataset,
batch_size=args.batch_size,
num_workers=6,
shuffle=False)
model = model_class(num_classes).to(device)
trainables_wo_bn = [
param for name, param in model.named_parameters()
if param.requires_grad and "bn" not in name
]
trainables_only_bn = [
param for name, param in model.named_parameters()
if param.requires_grad and "bn" in name
]
optimizer = torch.optim.SGD(
[
{
"params": trainables_wo_bn,
"weight_decay": 0.0001
},
{
"params": trainables_only_bn
},
],
lr=args.lr,
momentum=0.9,
)
trainer = Trainer(
optimizer,
model,
training_dataloader,
validation_dataloader,
max_epoch=args.epochs,
resume=args.resume,
log_dir=log_dir,
)
trainer.train()
def train(args):
dataset_dir = get_dataset_dir(args)
log_dir = get_log_dir(args)
model_class = get_model_class(args)
# orgainzesz dataset go more into
training_set, validation_set, num_classes = create_datasets(dataset_dir)
#look more into
training_dataset = Dataset(training_set,
transform_for_training(model_class.IMAGE_SHAPE))
validation_dataset = Dataset(validation_set,
transform_for_infer(model_class.IMAGE_SHAPE))
training_dataloader = torch.utils.data.DataLoader(
training_dataset,
batch_size=args.batch_size,
num_workers=6,
shuffle=True)
validation_dataloader = torch.utils.data.DataLoader(
validation_dataset,
batch_size=args.batch_size,
num_workers=6,
shuffle=False)
# what is Numclases and devie
model = model_class(num_classes).to(device)
# tain trainables_wo_bn and trainables_only_bn
trainables_wo_bn = [
param for name, param in model.named_parameters()
if param.requires_grad and 'bn' not in name
]
trainables_only_bn = [
param for name, param in model.named_parameters()
if param.requires_grad and 'bn' in name
]
optimizer = torch.optim.SGD([{
'params': trainables_wo_bn,
'weight_decay': 0.0001
}, {
'params': trainables_only_bn
}],
lr=args.lr,
momentum=0.9)
trainer = Trainer(optimizer,
model,
training_dataloader,
validation_dataloader,
max_epoch=args.epochs,
resume=args.resume,
log_dir=log_dir)
trainer.train()
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('config', type=str, help='Experiment to run')
parser.add_argument('--workers',
type=int,
default=None,
help='set number of workers (default #CPUs - 1)')
parser.add_argument('--log_dir',
type=str,
default="runs/run1",
help='directory to save/load the runs and logs')
parser.add_argument("--dataroot",
type=str,
default="data",
help="change datasets root path")
parser.add_argument("--cuda", action="store_true", help="train on GPUs")
parser.add_argument("--single_gpu",
action="store_true",
help="Disable multi-GPU")
parser.add_argument("--preload_vid",
action="store_true",
help="Load video features into RAM")
parser.add_argument("--no_preload_text",
action="store_true",
help="Do not load text features into RAM")
args = parser.parse_args()
cfg = utils.load_config(args.config)
if cfg.training.random_seed is not None:
print('Random seed: {:d}'.format(cfg.training.random_seed))
utils.set_seed(cfg.training.random_seed)
num_workers = min(10,
cpu_count() -
1) if args.workers is None else args.workers
print(f"{num_workers} parallel dataloader workers")
dataset_path = Path(args.dataroot) / cfg.dataset.name
train_set, val_set = create_datasets(dataset_path, cfg, args.preload_vid,
not args.no_preload_text)
train_loader, val_loader = create_loaders(train_set, val_set,
cfg.training.batch_size,
num_workers)
trainer = TrainerVideoText(args.log_dir, cfg, args.cuda, args.cuda
and not args.single_gpu)
trainer.train_loop(train_loader, val_loader)
trainer.logger.info("---------- Results ----------")
utils.print_csv_results(trainer.log_dir / "train_metrics.csv",
cfg,
print_fn=trainer.logger.info)
trainer.close()
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('config', type=str, help='Experiment to run')
parser.add_argument('checkpoint', type=str, help='Checkpoint to load')
parser.add_argument('--workers',
type=int,
default=None,
help='set number of workers (default #CPUs - 1)')
parser.add_argument('--log_dir',
type=str,
default="runs/eval",
help='directory to save/load the runs and logs')
parser.add_argument("--dataroot",
type=str,
default="data",
help="change datasets root path")
parser.add_argument("--cuda", action="store_true", help="train on GPUs")
parser.add_argument("--single_gpu",
action="store_true",
help="Disable multi-GPU")
args = parser.parse_args()
cfg = utils.load_config(args.config)
utils.set_seed(0)
num_workers = min(10,
cpu_count() -
1) if args.workers is None else args.workers
print(f"{num_workers} parallel dataloader workers")
dataset_path = Path(args.dataroot) / cfg.dataset.name
train_set, val_set = create_datasets(dataset_path, cfg, False, False)
train_loader, val_loader = create_loaders(train_set, val_set,
cfg.training.batch_size,
num_workers)
trainer = TrainerVideoText(args.log_dir, cfg, args.cuda, args.cuda
and not args.single_gpu, args.checkpoint, False)
trainer.validate(val_loader)
trainer.close()
def train(config):
train_config = config['train']
global device
device = train_config['device']
if not torch.cuda.is_available(): device = 'cpu'
tqdm.write('Training on {}'.format(device))
writer = SummaryWriter('log')
train_dataset, test_dataset = create_datasets(**config['dataset'])
train_dataloader = DataLoader(train_dataset,
batch_size=train_config['batch_size'],
shuffle=True,
collate_fn=collate_fn)
test_dataloader = DataLoader(test_dataset,
batch_size=train_config['batch_size'],
shuffle=False,
collate_fn=collate_fn)
encoder = Encoder(vocab_size=len(train_dataset.lang1),
**config['encoder'],
device=device).to(device)
decoder = Decoder(vocab_size=len(train_dataset.lang2),
**config['decoder']).to(device)
encoder_optimizer = optim.Adam(encoder.parameters(), lr=train_config['lr'])
decoder_optimizer = optim.Adam(decoder.parameters(), lr=train_config['lr'])
criterion = nn.NLLLoss()
tqdm.write('[-] Start training! ')
epoch_bar = tqdm(range(train_config['n_epochs']),
desc='[Total progress]',
leave=True,
position=0,
dynamic_ncols=True)
for epoch in epoch_bar:
batch_bar = tqdm(range(len(train_dataloader)),
desc='[Train epoch {:2}]'.format(epoch),
leave=True,
position=0,
dynamic_ncols=True)
encoder.train()
decoder.train()
train_loss = 0
for batch in batch_bar:
(source, target_bos, target_eos) = next(iter(train_dataloader))
encoder_optimizer.zero_grad()
decoder_optimizer.zero_grad()
source, target_bos, target_eos = source.to(device), target_bos.to(
device), target_eos.to(device)
encoder_output, encoder_hidden = encoder(source)
decoder_output = decoder(target_bos, encoder_hidden)
loss = criterion(decoder_output.view(-1, decoder_output.size(-1)),
target_eos.view(-1))
train_loss += loss.item()
n_hit, n_total = hitRate(decoder_output, target_eos)
loss.backward()
#print(loss.item())
encoder_optimizer.step()
decoder_optimizer.step()
batch_bar.set_description(
'[Train epoch {:2} | Loss: {:.2f} | Hit: {}/{}]'.format(
epoch, loss, n_hit, n_total))
train_loss /= len(train_dataloader)
batch_bar = tqdm(range(len(test_dataloader)),
desc='[Test epoch {:2}]'.format(epoch),
leave=True,
position=0,
dynamic_ncols=True)
encoder.eval()
decoder.eval()
test_loss = 0
for batch in batch_bar:
(source, target_bos, target_eos) = next(iter(test_dataloader))
source, target_bos, target_eos = source.to(device), target_bos.to(
device), target_eos.to(device)
with torch.no_grad():
encoder_output, encoder_hidden = encoder(source)
decoder_output = decoder(target_bos, encoder_hidden)
loss = criterion(
decoder_output.view(-1, decoder_output.size(-1)),
target_eos.view(-1))
test_loss += loss.item()
n_hit, n_total = hitRate(decoder_output, target_eos)
batch_bar.set_description(
'[Test epoch {:2} | Loss: {:.2f} | Hit: {}/{}]'.format(
epoch, loss, n_hit, n_total))
test_loss /= len(test_dataloader)
writer.add_scalars('Loss', {
'train': train_loss,
'test': test_loss
}, epoch)
sample(test_dataset, encoder, decoder)
tqdm.write('[-] Done!')
logger.info('Running in offline mode...')
# Access workspace.
logger.info('Accessing workspace...')
workspace = Workspace.from_config()
experiment = Experiment(workspace, "training-junkyard")
run = experiment.start_logging(outputs=None, snapshot_directory=None)
# Online run. Use dataset provided by training notebook.
else:
logger.info('Running in online mode...')
experiment = run.experiment
workspace = experiment.workspace
# Prepare the datasets.
dataset_train, dataset_validate, dataset_anomaly = create_datasets(
workspace, experiment, run, offline_run, CONFIG)
# Create the model.
model = Autoencoder(family=CONFIG.MODEL_FAMILY,
input_shape=(CONFIG.IMAGE_TARGET_HEIGHT,
CONFIG.IMAGE_TARGET_WIDTH,
CONFIG.IMAGE_TARGET_DEPTH),
filters=CONFIG.FILTERS,
latent_dim=CONFIG.LATENT_DIM,
size=CONFIG.MODEL_SIZE)
#model.summary()
# Make sure that output path exists.
outputs_path = "outputs"
if not os.path.exists(outputs_path):
os.mkdir(outputs_path)
from lang import Lang
from utils import *
from dataset import SentencePairDataset, collate_fn, create_datasets
from torch.utils.data import DataLoader
if __name__ == '__main__':
lang1 = Lang('english',
[['hello', 'fda', 'fsdf'], ['okf', 'fda', 'fdsfa']])
print(lang1)
# dataset = SentencePairDataset('cmn.txt', 'eng', 'chi')
# print(len(dataset))
# print(dataset.lang1)
# print(dataset.lang2)
# print(dataset[100])
# print(dataset)
train_dataset, test_dataset = create_datasets('../data/cmn.txt', 'english',
'chinese', 5, 0.1)
print(len(train_dataset), len(test_dataset))
dataloader = DataLoader(train_dataset,
batch_size=32,
shuffle=True,
collate_fn=collate_fn)
for i, (source, target_bos, target_eos) in enumerate(dataloader):
if i == 0: print(source, target_bos, target_eos)
print(len(train_dataset.lang1), len(train_dataset.lang2))
print(len(test_dataset.lang1), len(test_dataset.lang2))
type=bool,
default=False,
help='Perform inference on the whole test set')
parser.add_argument('-i',
'--image',
type=str,
default=None,
help='To the image file')
FLAGS, unparsed = parser.parse_known_args()
# Create all datasets
if not os.path.exists(TRAIN_PATH) or not os.path.exists(VAL_PATH) or \
not os.path.exists(TEST_PATH):
create_datasets()
if not FLAGS.train:
if FLAGS.image == None and FLAGS.on_test_batch == False:
raise Exception('Path to image file is not provided!')
elif FLAGS.on_test_batch:
test_model(FLAGS.graph_path, FLAGS.on_test_batch)
else:
test_model(FLAGS.graph_path, FLAGS.on_test_batch, FLAGS.image)
exit(0)
# Initialize the ImageDataGenerator
train_aug = ImageDataGenerator(rescale=1.0 / 255,
rotation_range=20,
zoom_range=0.05,
width_shift_range=0.05,