parser.add_argument(
'--dataset_file',
default=
'/data6/users/xuran7/myresearch/dense_videocap/third_party/myfork/densecap/data/anet/anet_annotations_trainval.json',
type=str)
parser.add_argument('--val_data_folder',
default='validation',
help='validation data folder')
parser.set_defaults(cuda=False, learn_mask=False, gated_mask=False)
args = parser.parse_args()
#pdb.set_trace()
with open(args.cfgs_file, 'r') as handle:
options_yaml = yaml.load(handle)
update_values(options_yaml, vars(args))
print(args)
def eval_results(args):
subprocess.Popen(["python2", args.densecap_eval_file, "-s", \
args.densecap_res, \
"-v", "-r"] + \
args.densecap_references \
)
anet_proposal = ANETproposal(args.dataset_file,
args.prop_res,
tiou_thresholds=np.linspace(0.5, 0.95, 10),
max_avg_nr_proposals=100,
def parse_opt(train=True):
parser = argparse.ArgumentParser()
# Data input settings
parser.add_argument('--cfgs_file',
default='cfgs/yc2.yml',
type=str,
help='dataset specific settings. anet | yc2')
parser.add_argument('--dataset',
default='yc2',
type=str,
help='which dataset to use. two options: anet | yc2')
parser.add_argument('--dataset_file', default='', type=str)
parser.add_argument('--feature_root',
default='',
type=str,
help='the feature root')
parser.add_argument('--dur_file', default='', type=str)
parser.add_argument(
'--start_from',
default='',
help=
'path to a model checkpoint to initialize model weights from. Empty = dont'
)
if not train:
parser.set_defaults(start_from='checkpoint/best_model.t7')
parser.add_argument('--max_sentence_len', default=20, type=int)
parser.add_argument('--num_workers', default=1, type=int)
if train:
parser.add_argument('--train_data_folder',
default=['training'],
type=str,
nargs='+',
help='training data folder')
parser.add_argument('--val_data_folder',
default=['validation'],
help='validation data folder')
else:
parser.add_argument('--test_data_folder',
default=['validation'],
help='testing data folder')
# Model settings: General
# waiting for completion
parser.add_argument('--temporal_scale',
default=200,
type=int,
help='length of observation window')
parser.add_argument('--max_duration',
default=100,
type=int,
help='max proposal duration')
parser.add_argument('--prop_boundary_ratio', type=int, default=0.5)
parser.add_argument('--num_sample', type=int, default=32)
parser.add_argument('--num_sample_perbin', type=int, default=3)
parser.add_argument('--feat_dim', type=int, default=3072)
# Optimization: General
parser.add_argument(
'--batch_size',
default=8,
type=int,
help=
'what is the batch size in number of images per batch? (there will be x seq_per_img sentences)'
)
parser.add_argument('--cuda',
dest='cuda',
action='store_true',
help='use gpu')
parser.add_argument(
'--id',
default='',
help=
'an id identifying this run/job. used in cross-val and appended when writing progress files'
)
if train:
# Optimization
parser.add_argument('--max_epochs',
default=10,
type=int,
help='max number of epochs to run for')
parser.add_argument(
'--optim',
default='sgd',
help='what update to use? rmsprop|sgd|sgdmom|adagrad|adam')
parser.add_argument('--learning_rate',
default=0.02,
type=float,
help='learning rate')
parser.add_argument('--alpha',
default=0.95,
type=float,
help='alpha for adagrad/rmsprop/momentum/adam')
parser.add_argument('--beta',
default=0.999,
type=float,
help='beta used for adam')
parser.add_argument(
'--epsilon',
default=1e-8,
help='epsilon that goes into denominator for smoothing')
parser.add_argument('--loss_alpha_r',
default=10,
type=int,
help='The weight for regression loss')
parser.add_argument('--patience_epoch',
default=1,
type=int,
help='Epoch to wait to determine a pateau')
parser.add_argument('--reduce_factor',
default=0.5,
type=float,
help='Factor of learning rate reduction')
parser.add_argument('--grad_norm',
default=1.5,
type=float,
help='Gradient clipping norm')
# Evaluation/Checkpointing
parser.add_argument('--calc_pos_neg',
dest='calc_pos_neg',
action='store_true')
parser.add_argument('--save_checkpoint_every',
default=1,
type=int,
help='how many epochs to save a model checkpoint?')
parser.add_argument(
'--checkpoint_path',
default='./checkpoint',
help='folder to save checkpoints into (empty = this folder)')
parser.add_argument(
'--losses_log_every',
default=1,
type=int,
help=
'How often do we snapshot losses, for inclusion in the progress dump? (0 = disable)'
)
parser.add_argument('--seed',
default=213,
type=int,
help='random number generator seed to use')
else:
# Model settings: Proposal and mask
parser.add_argument('--soft_nms_alpha', default=0.4, type=float)
parser.add_argument('--soft_nms_low_thres', default=0.5, type=float)
parser.add_argument('--soft_nms_high_thres', default=0.9, type=float)
parser.set_defaults(cuda=True)
args = parser.parse_args()
if train:
args.learning_rate /= args.batch_size
args.grad_norm *= args.batch_size
print('lr=', args.learning_rate)
import os
path = os.path.dirname(__file__)
with open(os.path.join(path, args.cfgs_file), 'r') as handle:
options_yaml = yaml.load(handle)
update_values(options_yaml, vars(args))
return args
def parse_args():
parser = argparse.ArgumentParser()
# Data input settings
parser.add_argument('--cfgs_file',
default='cfgs/anet.yml',
type=str,
help='dataset specific settings. anet | yc2')
parser.add_argument('--dataset',
default='',
type=str,
help='which dataset to use. two options: anet | yc2')
parser.add_argument('--dataset_file', default='', type=str)
parser.add_argument('--feature_root',
default='',
type=str,
help='the feature root')
parser.add_argument('--dur_file', default='', type=str)
parser.add_argument('--train_data_folder',
default=['training'],
type=str,
nargs='+',
help='training data folder')
parser.add_argument('--val_data_folder',
default=['validation'],
help='validation data folder')
parser.add_argument('--save_train_samplelist', action='store_true')
parser.add_argument('--load_train_samplelist', action='store_true')
parser.add_argument(
'--train_samplelist_path',
type=str,
default='/z/home/luozhou/subsystem/densecap_vid/train_samplelist.pkl')
parser.add_argument('--save_valid_samplelist', action='store_true')
parser.add_argument('--load_valid_samplelist', action='store_true')
parser.add_argument(
'--valid_samplelist_path',
type=str,
default='/z/home/luozhou/subsystem/densecap_vid/valid_samplelist.pkl')
parser.add_argument(
'--start_from',
default='',
help=
'path to a model checkpoint to initialize model weights from. Empty = dont'
)
parser.add_argument('--max_sentence_len', default=20, type=int)
parser.add_argument('--num_workers', default=1, type=int)
# Model settings: General
parser.add_argument(
'--d_model',
default=1024,
type=int,
help='size of the rnn in number of hidden nodes in each layer')
parser.add_argument('--d_hidden', default=2048, type=int)
parser.add_argument('--n_heads', default=8, type=int)
parser.add_argument('--in_emb_dropout', default=0.1, type=float)
parser.add_argument('--attn_dropout', default=0.2, type=float)
parser.add_argument('--vis_emb_dropout', default=0.1, type=float)
parser.add_argument('--cap_dropout', default=0.2, type=float)
parser.add_argument('--image_feat_size',
default=3072,
type=int,
help='the encoding size of the image feature')
parser.add_argument('--n_layers',
default=2,
type=int,
help='number of layers in the sequence model')
parser.add_argument(
'--train_sample',
default=20,
type=int,
help='total number of positive+negative training samples (2*U)')
parser.add_argument(
'--sample_prob',
default=0,
type=float,
help='probability for use model samples during training')
# Model settings: Proposal and mask
parser.add_argument('--slide_window_size',
default=480,
type=int,
help='the (temporal) size of the sliding window')
parser.add_argument('--slide_window_stride',
default=20,
type=int,
help='the step size of the sliding window')
parser.add_argument(
'--sampling_sec',
default=0.5,
help='sample frame (RGB and optical flow) with which time interval')
parser.add_argument('--kernel_list',
default=[
1, 2, 3, 4, 5, 7, 9, 11, 15, 21, 29, 41, 57, 71,
111, 161, 211, 251
],
type=int,
nargs='+')
parser.add_argument('--pos_thresh', default=0.7, type=float)
parser.add_argument('--neg_thresh', default=0.3, type=float)
parser.add_argument(
'--stride_factor',
default=50,
type=int,
help=
'the proposal temporal conv kernel stride is determined by math.ceil(kernel_len/stride_factor)'
)
# Optimization: General
parser.add_argument('--max_epochs',
default=20,
type=int,
help='max number of epochs to run for')
parser.add_argument(
'--batch_size',
default=32,
type=int,
help=
'what is the batch size in number of images per batch? (there will be x seq_per_img sentences)'
)
parser.add_argument('--valid_batch_size', default=64, type=int)
parser.add_argument('--cls_weight', default=1.0, type=float)
parser.add_argument('--reg_weight', default=10, type=float)
parser.add_argument('--sent_weight', default=0.25, type=float)
parser.add_argument('--scst_weight', default=0.0, type=float)
parser.add_argument('--mask_weight', default=0.0, type=float)
parser.add_argument('--gated_mask', action='store_true', dest='gated_mask')
# Optimization
parser.add_argument(
'--optim',
default='sgd',
help='what update to use? rmsprop|sgd|sgdmom|adagrad|adam')
parser.add_argument('--learning_rate',
default=0.1,
type=float,
help='learning rate')
parser.add_argument('--alpha',
default=0.95,
type=float,
help='alpha for adagrad/rmsprop/momentum/adam')
parser.add_argument('--beta',
default=0.999,
type=float,
help='beta used for adam')
parser.add_argument(
'--epsilon',
default=1e-8,
help='epsilon that goes into denominator for smoothing')
parser.add_argument('--loss_alpha_r',
default=2,
type=int,
help='The weight for regression loss')
parser.add_argument('--patience_epoch',
default=1,
type=int,
help='Epoch to wait to determine a pateau')
parser.add_argument('--reduce_factor',
default=0.5,
type=float,
help='Factor of learning rate reduction')
parser.add_argument('--grad_norm',
default=1,
type=float,
help='Gradient clipping norm')
# Data parallel
parser.add_argument('--dist_url',
default='file:///home/luozhou/nonexistent_file',
type=str,
help='url used to set up distributed training')
parser.add_argument('--dist_backend',
default='gloo',
type=str,
help='distributed backend')
parser.add_argument('--world_size',
default=1,
type=int,
help='number of distributed processes')
################### add ################
parser.add_argument('--rank',
default=0,
type=int,
help='rank of distributefd process')
#########################################
# Evaluation/Checkpointing
parser.add_argument('--save_checkpoint_every',
default=1,
type=int,
help='how many epochs to save a model checkpoint?')
parser.add_argument(
'--checkpoint_path',
default='./checkpoint',
help='folder to save checkpoints into (empty = this folder)')
parser.add_argument(
'--losses_log_every',
default=1,
type=int,
help=
'How often do we snapshot losses, for inclusion in the progress dump? (0 = disable)'
)
parser.add_argument('--seed',
default=213,
type=int,
help='random number generator seed to use')
parser.add_argument('--cuda',
dest='cuda',
action='store_true',
help='use gpu')
parser.add_argument('--enable_visdom',
action='store_true',
dest='enable_visdom')
parser.set_defaults(cuda=False,
save_train_samplelist=False,
load_train_samplelist=False,
save_valid_samplelist=False,
load_valid_samplelist=False,
gated_mask=False,
enable_visdom=False)
args = parser.parse_args()
with open(args.cfgs_file, 'r') as handle:
options_yaml = yaml.load(handle)
update_values(options_yaml, vars(args))
print(args)
# arguments inspection
assert (args.slide_window_size >= args.slide_window_stride)
assert (args.sampling_sec == 0.5
) # attention! sampling_sec is hard coded as 0.5
torch.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
if args.cuda:
torch.cuda.manual_seed_all(args.seed)
return args
def parse_opt_rnn():
parser = argparse.ArgumentParser()
# Data input settings
parser.add_argument('--cfgs_file',
default='cfgs/yc2.yml',
type=str,
help='dataset specific settings. anet | yc2')
parser.add_argument('--dataset',
default='yc2',
type=str,
help='which dataset to use. two options: anet | yc2')
parser.add_argument('--dataset_file', default='', type=str)
parser.add_argument('--feature_root',
default='',
type=str,
help='the feature root')
parser.add_argument('--dur_file', default='', type=str)
parser.add_argument('--bmn_model',
default='checkpoint/best_model.t7',
help='path to a pretrained bmn model')
parser.add_argument('--ptr_model',
default='checkpoint_rnn/best_model.t7',
help='path to a pretrained bmn model')
parser.add_argument('--max_sentence_len', default=20, type=int)
parser.add_argument('--num_workers', default=3, type=int)
parser.add_argument('--train_data_folder',
default=['training'],
type=str,
nargs='+',
help='training data folder')
parser.add_argument('--val_data_folder',
default=['validation'],
help='validation data folder')
parser.add_argument('--test_data_folder',
default=['validation'],
help='testing data folder')
# Model settings: General
parser.add_argument('--temporal_scale',
default=200,
type=int,
help='length of observation window')
parser.add_argument('--max_duration',
default=100,
type=int,
help='max proposal duration')
parser.add_argument('--prop_boundary_ratio', type=int, default=0.5)
parser.add_argument('--prop_num', type=int, default=100)
parser.add_argument('--max_seq_len', type=int, default=20)
parser.add_argument('--num_sample', type=int, default=32)
parser.add_argument('--num_sample_perbin', type=int, default=3)
parser.add_argument('--feat_dim', type=int, default=3072)
# Model settings: rnn_ptr
parser.add_argument('--hidden_size', type=int, default=256)
parser.add_argument('--prop_feat',
type=list,
default=['score', 'loc', 'vis'])
# Model settings: Proposal and mask
parser.add_argument('--soft_nms_alpha', default=0.4, type=float)
parser.add_argument('--soft_nms_low_thres', default=0.5, type=float)
parser.add_argument('--soft_nms_high_thres', default=0.9, type=float)
# Optimization: General
parser.add_argument('--batch_size', default=8, type=int, help='batch size')
parser.add_argument('--cuda',
dest='cuda',
action='store_true',
help='use gpu')
# Optimization
parser.add_argument('--max_epochs',
default=5,
type=int,
help='max number of epochs to run for')
parser.add_argument(
'--optim',
default='adam',
help='what update to use? rmsprop|sgd|sgdmom|adagrad|adam')
parser.add_argument('--learning_rate',
default=1e-3,
type=float,
help='learning rate')
parser.add_argument('--alpha',
default=0.95,
type=float,
help='alpha for adagrad/rmsprop/momentum/adam')
parser.add_argument('--beta',
default=0.999,
type=float,
help='beta used for adam')
parser.add_argument(
'--epsilon',
default=1e-8,
help='epsilon that goes into denominator for smoothing')
parser.add_argument('--patience_epoch',
default=1,
type=int,
help='Epoch to wait to determine a pateau')
parser.add_argument('--reduce_factor',
default=0.5,
type=float,
help='Factor of learning rate reduction')
parser.add_argument('--grad_norm',
default=1,
type=float,
help='Gradient clipping norm')
# Evaluation/Checkpointing
parser.add_argument('--save_checkpoint_every',
default=1,
type=int,
help='how many epochs to save a model checkpoint?')
parser.add_argument(
'--checkpoint_path',
default='./checkpoint_rnn',
help='folder to save checkpoints into (empty = this folder)')
parser.add_argument(
'--losses_log_every',
default=1,
type=int,
help=
'How often do we snapshot losses, for inclusion in the progress dump? (0 = disable)'
)
parser.add_argument('--seed',
default=213,
type=int,
help='random number generator seed to use')
parser.set_defaults(cuda=True)
parser.set_defaults(cuda=True)
args = parser.parse_args()
import os
path = os.path.dirname(__file__)
with open(os.path.join(path, args.cfgs_file), 'r') as handle:
options_yaml = yaml.load(handle)
update_values(options_yaml, vars(args))
return args