def main():
# Hyper Parameters
parser = argparse.ArgumentParser()
parser.add_argument('--data_path',
default='/A/VSE/data/',
help='path to datasets')
parser.add_argument(
'--data_name',
default='resnet152_precomp',
help='{coco,f8k,f30k,10crop,irv2,resnet152}_precomp|coco|f8k|f30k')
parser.add_argument('--vocab_path',
default='./vocab/',
help='Path to saved vocabulary pickle files.')
parser.add_argument('--margin',
default=0.05,
type=float,
help='Rank loss margin.')
parser.add_argument('--num_epochs',
default=30,
type=int,
help='Number of training epochs.')
parser.add_argument('--batch_size',
default=128,
type=int,
help='Size of a training mini-batch.')
parser.add_argument('--word_dim',
default=300,
type=int,
help='Dimensionality of the word embedding.')
parser.add_argument(
'--embed_size',
default=1024,
type=int,
help=
'Dimensionality of the joint embedding. [NOTE: this is used only if <embed_size> differs from <gru_units>]'
)
parser.add_argument('--gru_units',
default=1024,
type=int,
help='Number of GRU neurons.')
parser.add_argument('--grad_clip',
default=1.,
type=float,
help='Gradient clipping threshold.')
parser.add_argument('--crop_size',
default=224,
type=int,
help='Size of an image crop as the CNN input.')
parser.add_argument('--num_layers',
default=1,
type=int,
help='Number of GRU layers.')
parser.add_argument('--learning_rate',
default=.001,
type=float,
help='Initial learning rate.')
parser.add_argument('--lr_update',
default=15,
type=int,
help='Number of epochs to update the learning rate.')
parser.add_argument('--workers',
default=10,
type=int,
help='Number of data loader workers.')
parser.add_argument('--log_step',
default=10,
type=int,
help='Number of steps to print and record the log.')
parser.add_argument('--val_step',
default=500,
type=int,
help='Number of steps to run validation.')
parser.add_argument('--logger_name',
default='runs/runX',
help='Path to save the model and Tensorboard log.')
parser.add_argument('--resume',
default='',
type=str,
metavar='PATH',
help='path to latest checkpoint (default: none)')
parser.add_argument('--max_violation',
action='store_true',
help='Use max instead of sum in the rank loss.')
parser.add_argument('--img_dim',
default=2048,
type=int,
help='Dimensionality of the image embedding.')
parser.add_argument('--finetune',
action='store_true',
help='Fine-tune the image encoder.')
parser.add_argument('--cnn_type',
default='vgg19',
help="""The CNN used for image encoder
(e.g. vgg19, resnet152)""")
parser.add_argument('--use_restval',
action='store_true',
help='Use the restval data for training on MSCOCO.')
parser.add_argument('--measure',
default='cosine',
help='Similarity measure used (cosine|order)')
parser.add_argument(
'--test_measure',
default=None,
help=
'Similarity used for retrieval (None<same used for training>|cosine|order)'
)
parser.add_argument('--use_abs',
action='store_true',
help='Take the absolute value of embedding vectors.')
parser.add_argument('--no_imgnorm',
action='store_true',
help='Do not normalize the image embeddings.')
parser.add_argument('--text_encoder',
default='seam-e',
choices=text_encoders.text_encoders_alias.keys())
parser.add_argument(
'--att_units',
default=300,
type=int,
help=
'Number of tanh neurons. When using --att_dim=None we apply a tanh directly to the att input. '
)
parser.add_argument('--att_hops',
default=30,
type=int,
help='Number of attention hops (viewpoints).')
parser.add_argument(
'--att_coef',
default=0.,
type=float,
help='Influence of Frobenius divergence in the loss function.')
opt = parser.parse_args()
if opt.test_measure is None:
opt.test_measure = opt.measure
print(opt)
logging.basicConfig(format='%(asctime)s %(message)s', level=logging.INFO)
tb_logger.configure(opt.logger_name, flush_secs=5)
tokenizer, vocab_size = data.get_tokenizer(opt.vocab_path, opt.data_name)
opt.vocab_size = vocab_size
collate_fn = 'collate_fn'
# Load data loaders
train_loader, val_loader = data.get_loaders(opt.data_name, tokenizer,
opt.crop_size, opt.batch_size,
opt.workers, opt, collate_fn)
# Construct the model
model = VSE(opt)
print(model.txt_enc)
# optionally resume from a checkpoint
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
start_epoch = checkpoint['epoch']
best_rsum = checkpoint['best_rsum']
model.load_state_dict(checkpoint['model'])
# Eiters is used to show logs as the continuation of another
# training
model.Eiters = checkpoint['Eiters']
print("=> loaded checkpoint '{}' (epoch {}, best_rsum {})".format(
opt.resume, start_epoch, best_rsum))
validate(opt, val_loader, model)
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
# Train the Model
best_rsum = 0
for epoch in range(opt.num_epochs):
adjust_learning_rate(opt, model.optimizer, epoch)
# train for one epoch
train(opt, train_loader, model, epoch, val_loader)
# evaluate on validation set
rsum = validate(opt, val_loader, model)
# remember best R@ sum and save checkpoint
is_best = rsum > best_rsum
best_rsum = max(rsum, best_rsum)
save_checkpoint(
{
'epoch': epoch + 1,
'model': model.state_dict(),
'best_rsum': best_rsum,
'opt': opt,
'Eiters': model.Eiters,
},
is_best,
prefix=opt.logger_name + '/')
def main():
# Hyper Parameters
parser = argparse.ArgumentParser()
parser.add_argument('--data_path', default='data',
help='path to datasets')
parser.add_argument('--data_name', default='f30k',
help='{coco,f8k,f30k,10crop}_precomp|coco|f8k|f30k')
parser.add_argument('--vocab_path', default='vocab',
help='Path to saved vocabulary pickle files.')
parser.add_argument('--margin', default=0.2, type=float,
help='Rank loss margin.')
parser.add_argument('--num_epochs', default=30, type=int,
help='Number of training epochs.')
parser.add_argument('--batch_size', default=128, type=int,
help='Size of a training mini-batch.')
parser.add_argument('--word_dim', default=300, type=int,
help='Dimensionality of the word embedding.')
parser.add_argument('--embed_size', default=1024, type=int,
help='Dimensionality of the joint embedding.')
parser.add_argument('--grad_clip', default=2., type=float,
help='Gradient clipping threshold.')
parser.add_argument('--crop_size', default=224, type=int,
help='Size of an image crop as the CNN input.')
parser.add_argument('--num_layers', default=1, type=int,
help='Number of GRU layers.')
parser.add_argument('--learning_rate', default=2e-4, type=float,
help='Initial learning rate.')
parser.add_argument('--lr_update', default=15, type=int,
help='Number of epochs to update the learning rate.')
parser.add_argument('--workers', default=10, type=int,
help='Number of data loader workers.')
parser.add_argument('--log_step', default=100, type=int,
help='Number of steps to print and record the log.')
parser.add_argument('--val_step', default=500, type=int,
help='Number of steps to run validation.')
parser.add_argument('--logger_name', default='runs/test',
help='Path to save the model and Tensorboard log.')
parser.add_argument('--resume', default='', type=str, metavar='PATH',
help='path to latest checkpoint (default: none)')
parser.add_argument('--img_dim', default=2048, type=int,
help='Dimensionality of the image embedding.')
parser.add_argument('--finetune', action='store_true',
help='Fine-tune the image encoder.')
parser.add_argument('--use_restval', action='store_true',
help='Use the restval data for training on MSCOCO.')
parser.add_argument('--reset_train', action='store_true',
help='Ensure the training is always done in '
'train mode (Not recommended).')
parser.add_argument('--K', default=2, type=int,help='num of JSR.')
parser.add_argument('--feature_path', default='data/joint-pretrain/flickr30k/region_feat_gvd_wo_bgd/trainval/',
type=str, help='path to the pre-computed image features')
parser.add_argument('--region_bbox_file',
default='data/joint-pretrain/flickr30k/region_feat_gvd_wo_bgd/flickr30k_detection_vg_thresh0.2_feat_gvd_checkpoint_trainvaltest.h5',
type=str, help='path to the region_bbox_file(.h5)')
opt = parser.parse_args()
print(opt)
logging.basicConfig(format='%(asctime)s %(message)s', level=logging.INFO)
tb_logger.configure(opt.logger_name, flush_secs=5)
# Load Vocabulary Wrapper
vocab = pickle.load(open(os.path.join(
opt.vocab_path, '%s_vocab.pkl' % opt.data_name), 'rb'))
opt.vocab_size = len(vocab)
# Load data loaders
train_loader, val_loader = data.get_loaders(
opt.data_name, vocab, opt.crop_size, opt.batch_size, opt.workers, opt)
# Construct the model
model = VSE(opt)
best_rsum = 0
# optionally resume from a checkpoint
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
start_epoch = checkpoint['epoch']
best_rsum = checkpoint['best_rsum']
model.load_state_dict(checkpoint['model'])
# Eiters is used to show logs as the continuation of another
# training
model.Eiters = checkpoint['Eiters']
print("=> loaded checkpoint '{}' (epoch {}, best_rsum {})"
.format(opt.resume, start_epoch, best_rsum))
validate(opt, val_loader, model)
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
del checkpoint
# Train the Model
for epoch in range(opt.num_epochs):
adjust_learning_rate(opt, model.optimizer, epoch)
# train for one epoch
train(opt, train_loader, model, epoch, val_loader)
# evaluate on validation set
rsum = validate(opt, val_loader, model)
# remember best R@ sum and save checkpoint
is_best = rsum > best_rsum
best_rsum = max(rsum, best_rsum)
save_checkpoint({
'epoch': epoch + 1,
'model': model.state_dict(),
'best_rsum': best_rsum,
'opt': opt,
'Eiters': model.Eiters,
}, is_best, epoch, prefix=opt.logger_name + '/')
def main():
# Hyper Parameters
parser = argparse.ArgumentParser()
parser.add_argument('--data_path',
default='/media/shenkev/data/Ubuntu/vsepp/data/data',
help='path to datasets')
parser.add_argument('--data_name',
default='coco_precomp',
help='{coco,f8k,f30k,10crop}_precomp|coco|f8k|f30k')
parser.add_argument('--vocab_path',
default='./vocab/',
help='Path to saved vocabulary pickle files.')
parser.add_argument('--margin',
default=0.2,
type=float,
help='Rank loss margin.')
parser.add_argument('--num_epochs',
default=30,
type=int,
help='Number of training epochs.')
parser.add_argument('--batch_size',
default=128,
type=int,
help='Size of a training mini-batch.')
parser.add_argument('--word_dim',
default=300,
type=int,
help='Dimensionality of the word embedding.')
parser.add_argument('--embed_size',
default=1024,
type=int,
help='Dimensionality of the joint embedding.')
parser.add_argument('--grad_clip',
default=2.,
type=float,
help='Gradient clipping threshold.')
parser.add_argument('--crop_size',
default=224,
type=int,
help='Size of an image crop as the CNN input.')
parser.add_argument('--num_layers',
default=1,
type=int,
help='Number of GRU layers.')
parser.add_argument('--learning_rate',
default=.0002,
type=float,
help='Initial learning rate.')
parser.add_argument('--lr_update',
default=15,
type=int,
help='Number of epochs to update the learning rate.')
parser.add_argument('--workers',
default=10,
type=int,
help='Number of data loader workers.')
parser.add_argument('--log_step',
default=10,
type=int,
help='Number of steps to print and record the log.')
parser.add_argument('--val_step',
default=500,
type=int,
help='Number of steps to run validation.')
parser.add_argument('--logger_name',
default='runs/coco_vse',
help='Path to save the model and Tensorboard log.')
parser.add_argument('--resume',
default='',
type=str,
metavar='PATH',
help='path to latest checkpoint (default: none)')
parser.add_argument('--max_violation',
default=True,
action='store_true',
help='Use max instead of sum in the rank loss.')
parser.add_argument('--img_dim',
default=4096,
type=int,
help='Dimensionality of the image embedding.')
parser.add_argument('--finetune',
action='store_true',
help='Fine-tune the image encoder.')
parser.add_argument('--cnn_type',
default='vgg19',
help="""The CNN used for image encoder
(e.g. vgg19, resnet152)""")
parser.add_argument('--use_restval',
action='store_true',
help='Use the restval data for training on MSCOCO.')
parser.add_argument('--measure',
default='cosine',
help='Similarity measure used (cosine|order)')
parser.add_argument('--use_abs',
action='store_true',
help='Take the absolute value of embedding vectors.')
parser.add_argument('--no_imgnorm',
action='store_true',
help='Do not normalize the image embeddings.')
opt = parser.parse_args()
print(opt)
logging.basicConfig(format='%(asctime)s %(message)s', level=logging.INFO)
tb_logger.configure(opt.logger_name, flush_secs=5)
# Load Vocabulary Wrapper
vocab = pickle.load(
open(os.path.join(opt.vocab_path, '%s_vocab.pkl' % opt.data_name),
'rb'))
opt.vocab_size = len(vocab)
# Load data loaders
train_loader, val_loader = data.get_loaders(opt.data_name, vocab,
opt.crop_size, opt.batch_size,
opt.workers, opt)
# Construct the model
model = VSE(opt)
# optionally resume from a checkpoint
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
start_epoch = checkpoint['epoch']
best_rsum = checkpoint['best_rsum']
model.load_state_dict(checkpoint['model'])
# Eiters is used to show logs as the continuation of another
# training
model.Eiters = checkpoint['Eiters']
print("=> loaded checkpoint '{}' (epoch {}, best_rsum {})".format(
opt.resume, start_epoch, best_rsum))
validate(opt, val_loader, model)
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
# Train the Model
best_rsum = 0
for epoch in range(opt.num_epochs):
adjust_learning_rate(opt, model.optimizer, epoch)
# train for one epoch
train(opt, train_loader, model, epoch, val_loader)
# evaluate on validation set
rsum = validate(opt, val_loader, model)
# remember best R@ sum and save checkpoint
is_best = rsum > best_rsum
best_rsum = max(rsum, best_rsum)
save_checkpoint(
{
'epoch': epoch + 1,
'model': model.state_dict(),
'best_rsum': best_rsum,
'opt': opt,
'Eiters': model.Eiters,
},
is_best,
prefix=opt.logger_name + '/')
def main():
# Hyper Parameters
torch.cuda.set_device(opt.gpu_id)
tb_logger.configure(opt.logger_name, flush_secs=5)
logging.basicConfig(format='%(asctime)s %(message)s', level=logging.INFO, filename=opt.logger_name+'/log.log')
console = logging.StreamHandler()
console.setLevel(logging.INFO)
formatter = logging.Formatter('%(asctime)s %(message)s')
console.setFormatter(formatter)
logging.getLogger('').addHandler(console)
logging.info(opt)
# Load Vocabulary Wrapper
vocab_path = os.path.join(opt.vocab_path, '%s_vocab_total.pkl' % opt.data_name)
print (vocab_path)
vocab = pickle.load(open(vocab_path, 'rb'))
opt.vocab_size = len(vocab)
# Load data loaders
train_loader, val_loader = data.get_loaders(
opt.data_name, vocab, opt.batch_size, opt.workers, opt)
# Construct the model
model = VSE(opt)
print('Print out models:')
print(model.clip_enc)
print(model.txt_enc)
print(model.vid_seq_enc)
print(model.txt_seq_enc)
start_epoch = 0
# optionally resume from a checkpoint
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
start_epoch = checkpoint['epoch']
best_rsum = checkpoint['best_rsum']
model.load_state_dict(checkpoint['model'], opt)
# Eiters is used to show logs as the continuation of another
# training
model.Eiters = checkpoint['Eiters']
print("=> loaded checkpoint '{}' (epoch {}, best_rsum {})"
.format(opt.resume, start_epoch, best_rsum))
validate(opt, val_loader, model)
if opt.eval_only:
return
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
# Train the Model
best_rsum = 0
for epoch in range(start_epoch, opt.num_epochs):
adjust_learning_rate(opt, model.optimizer, epoch)
# train for one epoch
train(opt, train_loader, model, epoch, val_loader)
# evaluate on validation set
rsum = validate(opt, val_loader, model)
# remember best R@ sum and save checkpoint
is_best = rsum > best_rsum
best_rsum = max(rsum, best_rsum)
save_checkpoint({
'epoch': epoch + 1,
'model': model.state_dict(opt),
'best_rsum': best_rsum,
'opt': opt,
'Eiters': model.Eiters,
}, is_best, prefix=opt.logger_name + '/', epoch=epoch)
def main():
# Hyper Parameters
parser = argparse.ArgumentParser()
parser.add_argument('--data_path', default=".", help='path to datasets')
parser.add_argument(
'--data_name',
default='m30k',
help='{coco,f8k,f30k,10crop}_precomp|coco|f8k|f30k|m30k')
parser.add_argument(
'--lang',
default='en',
help='Which language(s) to use from m30k, en-de, trains on en+de.')
parser.add_argument('--sentencepair',
action='store_true',
help='Train caption-caption ranking as well.')
parser.add_argument(
'--sentencepair_p',
default=0.5,
type=float,
help='Probability of training on caption-caption and not image-caption.'
)
parser.add_argument(
'--primary',
default=None,
help=
'Which language to monitor for early stopping. Multiple with l1-l2-l3')
parser.add_argument(
'--undersample',
action='store_true',
help='Pick only one of the 5 possilbe captions for m30k task 2.')
parser.add_argument('--half',
action='store_true',
help='Use only half of the M30K from task 2.')
parser.add_argument('--disaligned',
action='store_true',
help='Use only half of the M30K from task 2.')
parser.add_argument('--vocab_path',
default='.',
help='Path to saved vocabulary pickle files.')
parser.add_argument('--margin',
default=0.2,
type=float,
help='Rank loss margin.')
parser.add_argument('--num_epochs',
default=30,
type=int,
help='Number of training epochs.')
parser.add_argument(
'--patience',
default=10,
type=int,
help='Number of validation steps to tolerate without improvement.')
parser.add_argument('--batch_size',
default=128,
type=int,
help='Size of a training mini-batch.')
parser.add_argument('--word_dim',
default=300,
type=int,
help='Dimensionality of the word embedding.')
parser.add_argument('--embed_size',
default=1024,
type=int,
help='Dimensionality of the joint embedding.')
parser.add_argument('--grad_clip',
default=2.,
type=float,
help='Gradient clipping threshold.')
parser.add_argument('--crop_size',
default=224,
type=int,
help='Size of an image crop as the CNN input.')
parser.add_argument('--num_layers',
default=1,
type=int,
help='Number of GRU layers.')
parser.add_argument('--learning_rate',
default=.0002,
type=float,
help='Initial learning rate.')
parser.add_argument('--lr_update',
default=15,
type=int,
help='Number of epochs to update the learning rate.')
parser.add_argument('--workers',
default=10,
type=int,
help='Number of data loader workers.')
parser.add_argument('--log_step',
default=10,
type=int,
help='Number of steps to print and record the log.')
parser.add_argument('--val_step',
default=500,
type=int,
help='Number of steps to run validation.')
parser.add_argument(
'--logger_path',
default='.',
help='Path where to save the model and Tensorboard log.')
parser.add_argument(
'--logger_name',
help='Name of the folder where to save the model and Tensorboard log.')
parser.add_argument('--resume',
default='',
type=str,
metavar='PATH',
help='path to latest checkpoint (default: none)')
parser.add_argument('--sum_violation',
dest="max_violation",
action='store_false',
help='Use max instead of sum in the rank loss.')
parser.add_argument('--img_dim',
default=2048,
type=int,
help='Dimensionality of the image embedding.')
parser.add_argument('--finetune',
action='store_true',
help='Fine-tune the image encoder.')
parser.add_argument('--cnn_type',
default='vgg19',
help="""The CNN used for image encoder
(e.g. vgg19, resnet152)""")
parser.add_argument('--use_restval',
action='store_true',
help='Use the restval data for training on MSCOCO.')
parser.add_argument('--measure',
default='cosine',
help='Similarity measure used (cosine|order)')
parser.add_argument('--use_abs',
action='store_true',
help='Take the absolute value of embedding vectors.')
parser.add_argument('--no_imgnorm',
action='store_true',
help='Do not normalize the image embeddings.')
parser.add_argument('--reset_train',
action='store_true',
help='Ensure the training is always done in '
'train mode (Not recommended).')
parser.add_argument('--seed', default=42, type=int, help='Random seed.')
opt = parser.parse_args()
if torch.__version__ >= "0.3":
opt.reset_train = True
opt.vocab_path = os.path.join(opt.vocab_path, "vocab")
if opt.logger_name is None:
name = "lang{}_half-{}_undersample-{}_disaligned-{}_sentencepair-{}_primary-{}_epochs-{}"
name = name.format(opt.lang, opt.half, opt.undersample, opt.disaligned,
opt.sentencepair, opt.primary, opt.num_epochs)
opt.logger_name = os.path.join(opt.data_name, name)
opt.logger_name = os.path.join(opt.logger_path, opt.logger_name,
str(opt.seed))
print(opt)
random.seed(rseed + opt.seed)
np.random.seed(rseed + opt.seed)
torch.cuda.manual_seed(rseed + opt.seed)
torch.cuda.manual_seed_all(rseed + opt.seed)
logging.basicConfig(format='%(asctime)s %(message)s', level=logging.INFO)
tb_logger.configure(opt.logger_name, flush_secs=5)
# For multi30k compute vocabulary mappings on the fly.
if opt.data_name == "m30k":
vocab = None
langs = opt.lang.split("-")
# Load Vocabulary Wrapper for COCO or F30K
else:
vocab = pickle.load(
open(os.path.join(opt.vocab_path, '%s_vocab.pkl' % opt.data_name),
'rb'))
opt.vocab_size = len(vocab)
langs = [opt.data_name]
# Load data loaders
train_loader, val_loader = data.get_loaders(opt.data_name, vocab,
opt.crop_size, opt.batch_size,
opt.workers, opt)
# Construct the model
model = VSE(opt)
print(model.txt_enc)
# optionally resume from a checkpoint
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
start_epoch = checkpoint['epoch']
best_rsum = checkpoint['best_rsum']
model.load_state_dict(checkpoint['model'])
# Eiters is used to show logs as the continuation of another
# training
model.Eiters = checkpoint['Eiters']
print("=> loaded checkpoint '{}' (epoch {}, best_rsum {})".format(
opt.resume, start_epoch, best_rsum))
validate(opt, val_loader, model, "")
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
if len(langs) == 1 or opt.data_name != 'm30k':
# Train the Model on a single data set
best_rsum = 0
model.train_start()
for epoch in range(opt.num_epochs):
if opt.reset_train:
# Always reset to train mode, this is not the default behavior
model.train_start()
adjust_learning_rate(opt, model.optimizer, epoch)
# train for one epoch
train(opt, train_loader, model, epoch, val_loader)
# evaluate on validation set
rsum = validate(opt, val_loader, model, langs[0])
# remember best R@ sum and save checkpoint
is_best = rsum > best_rsum
best_rsum = max(rsum, best_rsum)
save_checkpoint(
{
'epoch': epoch + 1,
'model': model.state_dict(),
'best_rsum': best_rsum,
'opt': opt,
'Eiters': model.Eiters,
},
is_best,
prefix=opt.logger_name + '/')
if is_best:
patience_count = 0
print("New best: {}".format(best_rsum))
else:
patience_count += 1
print("No improvement in {}".format(patience_count))
if patience_count == opt.patience:
print("No improvement in {} epochs, stoppin".format(
patience_count))
break
else:
joint_train(opt, train_loader, model, val_loader)
def main():
# Hyper Parameters
parser = argparse.ArgumentParser()
parser.add_argument('--data_path', default='/w/31/faghri/vsepp_data/',
help='path to datasets')
parser.add_argument('--data_name', default='precomp',
help='{coco,f8k,f30k,10crop}_precomp|coco|f8k|f30k')
parser.add_argument('--vocab_path', default='./vocab/',
help='Path to saved vocabulary pickle files.')
parser.add_argument('--margin', default=0.2, type=float,
help='Rank loss margin.')
parser.add_argument('--num_epochs', default=15, type=int,
help='Number of training epochs.')
parser.add_argument('--batch_size', default=128, type=int,
help='Size of a training mini-batch.')
parser.add_argument('--word_dim', default=300, type=int,
help='Dimensionality of the word embedding.')
parser.add_argument('--embed_size', default=1024, type=int,
help='Dimensionality of the joint embedding.')
parser.add_argument('--grad_clip', default=2., type=float,
help='Gradient clipping threshold.')
parser.add_argument('--crop_size', default=224, type=int,
help='Size of an image crop as the CNN input.')
parser.add_argument('--num_layers', default=1, type=int,
help='Number of GRU layers.')
parser.add_argument('--learning_rate', default=.0002, type=float,
help='Initial learning rate.')
parser.add_argument('--lr_update', default=8, type=int,
help='Number of epochs to update the learning rate.')
parser.add_argument('--workers', default=10, type=int,
help='Number of data loader workers.')
parser.add_argument('--log_step', default=10, type=int,
help='Number of steps to print and record the log.')
parser.add_argument('--val_step', default=500, type=int,
help='Number of steps to run validation.')
parser.add_argument('--logger_name', default='runs/runX',
help='Path to save the model and Tensorboard log.')
parser.add_argument('--resume', default='', type=str, metavar='PATH',
help='path to latest checkpoint (default: none)')
parser.add_argument('--max_violation', action='store_true',
help='Use max instead of sum in the rank loss.')
parser.add_argument('--sum_violation', action='store_true')
parser.add_argument('--img_dim', default=4096, type=int,
help='Dimensionality of the image embedding.')
parser.add_argument('--finetune', action='store_true',
help='Fine-tune the image encoder.')
parser.add_argument('--cnn_type', default='vgg19',
help="""The CNN used for image encoder
(e.g. vgg19, resnet152)""")
parser.add_argument('--use_restval', action='store_true',
help='Use the restval data for training on MSCOCO.')
parser.add_argument('--measure', default='cosine',
help='Similarity measure used (cosine|order)')
parser.add_argument('--use_abs', action='store_true',
help='Take the absolute value of embedding vectors.')
parser.add_argument('--no_imgnorm', action='store_true',
help='Do not normalize the image embeddings.')
parser.add_argument('--reset_train', action='store_true',
help='Ensure the training is always done in '
'train mode (Not recommended).')
parser.add_argument('--save_all', action='store_true',
help="Save model after the training of each epoch")
parser.add_argument('--memory_bank', action='store_true',
help="Train model with memory bank")
parser.add_argument('--record_val', action='store_true',
help="Record the rsum values on validation set in file during training")
parser.add_argument('--local_alpha', default=30.0, type=float)
parser.add_argument('--local_ep', default=0.3, type=float)
parser.add_argument('--global_alpha', default=40.0, type=float)
parser.add_argument('--global_beta', default=40.0, type=float)
parser.add_argument('--global_ep_posi', default=0.2, type=float,
help="Global epsilon for positive pairs")
parser.add_argument('--global_ep_nega', default=0.1, type=float,
help="Global epsilon for negative pairs")
parser.add_argument('--mb_k', default=250, type=int,
help="Use top K items in memory bank")
parser.add_argument('--mb_rate', default=0.05, type=float,
help="-")
opt = parser.parse_args()
print(opt)
logging.basicConfig(format='%(message)s', level=logging.INFO)
tb_logger.configure(opt.logger_name, flush_secs=5)
# Load Vocabulary Wrapper
vocab = pickle.load(open(os.path.join(
opt.vocab_path, '%s_vocab.pkl' % opt.data_name), 'rb'))
opt.vocab_size = len(vocab)
print("Vocab Size: %d" % opt.vocab_size)
# Load data loaders
train_loader, val_loader = data.get_loaders(
opt.data_name, vocab, opt.crop_size, opt.batch_size, opt.workers, opt)
# Construct the model
model = VSE(opt)
# optionally resume from a checkpoint
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
start_epoch = checkpoint['epoch']
best_rsum = checkpoint['best_rsum']
model.load_state_dict(checkpoint['model'])
# Eiters is used to show logs as the continuation of another
# training
model.Eiters = checkpoint['Eiters']
print("=> loaded checkpoint '{}' (epoch {}, best_rsum {})"
.format(opt.resume, start_epoch, best_rsum))
validate(opt, val_loader, model)
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
# Train the Model
best_rsum = 0
for epoch in range(opt.num_epochs):
adjust_learning_rate(opt, model.optimizer, epoch)
memory_bank = opt.memory_bank
if memory_bank and epoch > 0:
load_memory_bank(opt, train_loader, model)
# train for one epoch
train(opt, train_loader, model, epoch, val_loader)
# evaluate on validation set
rsum = validate(opt, val_loader, model)
print ("rsum: %.1f" % rsum)
if opt.record_val:
with open("rst_val_" + opt.logger_name[5:], "a") as f:
f.write("Epoch: %d ; rsum: %.1f\n" %(epoch, rsum))
# remember best R@ sum and save checkpoint
is_best = rsum > best_rsum
best_rsum = max(rsum, best_rsum)
save_checkpoint({
'epoch': epoch + 1,
'model': model.state_dict(),
'best_rsum': best_rsum,
'opt': opt,
'Eiters': model.Eiters,
}, is_best, prefix=opt.logger_name + '/', save_all=opt.save_all)
# reset memory bank
model.mb_img = None
model.mb_cap = None
def main():
#训练外参数
parser = argparse.ArgumentParser()
parser.add_argument('--data_type', default='rgb', help='path to datasets')
parser.add_argument('--feature_path', default='', help='path to datasets')
parser.add_argument('--anno_path', default='', help='path to datasets')
parser.add_argument('--feature_prefix',
default='',
help='prefix of feaeture')
parser.add_argument('--dropout', default=0.5, help='prefix of feature')
parser.add_argument('--split_video_file',
default='',
help='prefix of feature')
parser.add_argument('--num_pos_sample',
default=10,
help='prefix of feature')
parser.add_argument('--num_neg_sample',
default=10,
help='prefix of feature')
parser.add_argument('--num_epochs',
default=30,
type=int,
help='Number of training epochs')
parser.add_argument('--batch_size',
default=128,
type=int,
help='Size of a training mini-batch')
parser.add_argument('--embed_size',
default=10240,
typr=int,
help='Dimensionality of the joint embedding.')
parser.add_argument('--grad_clip',
default=2,
type=float,
help='Gradient clipping threshold')
parser.add_argument('--learning_rate',
default=.001,
type=float,
help='Initial learning rate')
parser.add_argument('--lr_update',
default=15,
type=int,
help='Number of epochs to update the learning rate')
parser.add_argument('--wrokers',
default=10,
type=int,
help='Number of data loader workers')
parser.add_argument('--log_step',
default=10,
type=int,
help='Number of steps to print and record the log')
parser.add_argument('--val_step',
default=500,
type=int,
help='Number of steps to run validation')
parser.add_argument('--logger_name',
default='runs/runX',
help='Path to save the model and Tensorboard log')
parser.add_argument('--resume',
default='',
type=str,
metavar='PATH',
help='path to latest checkpoint default=none')
parser.add_argument('--storage_place',
default='',
type=str,
metavar='Path',
help='path to latest checkpoint default=none')
parser.add_argument('--instance_data_path',
default='',
type=str,
metavar='PATH',
help='path to the latest checkpoint')
opt = parser.parse_args()
logging.basicConfig(format='%(asctime)s %(message)s', level=logging.INFO)
tb_logger.configure(opt.logger_name, flush_secs=5)
#Load data loaders
#1
train_loader, val_loader = data.get_loaders(opt)
#Construct the model
#2
model = VSE(opt)
#optionally resume from a checkpoint
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
start_epoch = checkpoint['epoch']
best_rsum = checkpoint['best_rsum']
model.load_state_dict(checkpoint['model'])
#Eiters is used to show logs as the continuation of another
#training
model.Eiters = checkpoint['Eiters']
print("=> loaded checkpoint '{}' (epoch {}, best_rsum {})".format(
opt.resume, start_epoch, best_rsum))
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
# Train the Model
best_rsum = 0
for epoch in range(opt.num_epochs):
adjust_learning_rate(opt, model.optimizer, epoch)
#train for one epoch
#3
train(opt, train_loader, model, epoch)
#evaluate on validation set
rsum = 0
#remeber best R@ sum and save checkpoint
is_best = rsum > best_rsum
best_rsum = max(rsum, best_rsum)
save_checkpoint(
{
'epoch': epoch + 1,
'model': model.state_dict(),
'best_rsum': best_rsum,
'opt': opt,
'Eiters': model.Eiters,
},
is_best,
prefix=opt.logger_name + '/')
#4 val_loader (feat,id,labels)
eval_feat(val_loader, model, opt.storage_place)
def main():
parser = argparse.ArgumentParser()
# Directories.
parser.add_argument('--data_path',
default='/DATA/cvpr19',
help='path to datasets')
parser.add_argument('--vocab_path',
default='../vocab/',
help='Path to saved vocabulary pickle files.')
parser.add_argument('--resume',
default='',
type=str,
metavar='PATH',
help='path to latest checkpoint (default: none)')
parser.add_argument('--model',
type=str,
default='beenburger',
help='model_name')
parser.add_argument(
'--save_dir',
type=str,
default='coco2',
help='save checkpoint and results in DATA_PATH/MODEL_NAME/SAVE_DIR')
# Dataset.
parser.add_argument('--data_name', default='coco', help='{coco|ours}')
parser.add_argument('--use_restval',
default='True',
type=str2bool,
help='Use the restval data for training on MSCOCO.')
# Model configurations.
parser.add_argument('--margin',
default=0.2,
type=float,
help='Rank loss margin.')
parser.add_argument('--K',
default=620,
type=int,
help='Dimensionality of the word embedding.')
parser.add_argument('--num_layers',
default=4,
type=int,
help='Number of SRU layers.')
parser.add_argument('--D',
type=int,
default='2048',
help='dimension of image feature from ResNet')
parser.add_argument('--D_prime',
type=int,
default='2400',
help='dimension of adaptation + pooling')
parser.add_argument('--d',
type=int,
default='2400',
help='Dimensionality of the joint embedding')
# Training configurations.
parser.add_argument('--num_epochs',
default=30,
type=int,
help='Number of training epochs.')
parser.add_argument('--batch_size',
default=160,
type=int,
help='Size of a training mini-batch.')
parser.add_argument('--img_size', default=256, type=int, help='image_size')
parser.add_argument('--crop_size',
default=224,
type=int,
help='Size of an image crop as the CNN input.')
parser.add_argument('--learning_rate',
default=.001,
type=float,
help='Initial learning rate.')
parser.add_argument('--lr_decay',
type=float,
default=0.5,
help='learning rate decay')
parser.add_argument('--workers',
default=4,
type=int,
help='Number of data loader workers.')
# Miscellaneous.
parser.add_argument('--log_step',
default=10,
type=int,
help='Number of steps to print and record the log.')
parser.add_argument('--val_step',
default=500,
type=int,
help='Number of steps to run validation.')
parser.add_argument('--logger_name',
default='../runs/runX',
help='Path to save the model and Tensorboard log.')
opt = parser.parse_args()
print(opt)
logging.basicConfig(format='%(asctime)s %(message)s', level=logging.INFO)
tb_logger.configure(opt.logger_name, flush_secs=5)
# Load Vocabulary Wrapper
vocab = pickle.load(
open(os.path.join(opt.vocab_path, '%s_vocab.pkl' % opt.data_name),
'rb'))
opt.vocab_size = len(vocab)
opt.vocab = vocab
# Create directories
create_directory(opt.data_path, opt.model, opt.save_dir)
# Load data loaders
train_loader, val_loader = data.get_loaders(opt.data_name, vocab,
opt.crop_size, opt.batch_size,
opt.workers, opt)
# Construct the model
model = VSE(opt)
# optionally resume from a checkpoint
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
start_epoch = checkpoint['epoch']
model.change_training_state(0)
if start_epoch > 2:
model.change_training_state(2)
if start_epoch > 8:
model.change_training_state(8)
best_rsum = checkpoint['best_rsum']
model.optimizer = checkpoint['optim']
model.load_state_dict(checkpoint['model'])
# Eiters is used to show logs as the continuation of another
# training
model.Eiters = checkpoint['Eiters']
print("=> loaded checkpoint '{}' (epoch {}, best_rsum {})".format(
opt.resume, start_epoch, best_rsum))
validate(opt, val_loader, model)
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
# Train the Model
best_rsum = 0
start_epoch = 0
for epoch in range(start_epoch, opt.num_epochs):
adjust_learning_rate(opt, model.optimizer, epoch)
model.change_training_state(epoch)
# train for one epoch
train(opt, train_loader, model, epoch, val_loader)
# evaluate on validation set
rsum = validate(opt, val_loader, model)
# remember best R@ sum and save checkpoint
is_best = rsum > best_rsum
best_rsum = max(rsum, best_rsum)
save_checkpoint(
{
'epoch': epoch + 1,
'model': model.state_dict(),
'best_rsum': best_rsum,
'opt': opt,
'Eiters': model.Eiters,
'optim': model.optimizer,
},
is_best,
prefix=os.path.join(opt.data_path, opt.model, opt.save_dir))