np.random.seed(1234)
tf.set_random_seed(1234)
# read the train args from checkpoint
param_path = args['checkpoint'].replace('.tmodel', '_params.json')
with open(param_path, 'r') as file_id:
saved_args = json.load(file_id)
saved_args.update(args)
args = saved_args
args['preload_feats'] = False
# no supervision is needed
args['supervise_attention'] = False
args['batch_size'] = min(args['batch_size'], 10) # adjust for complex models
support.pretty_print_dict(args)
# Data files
root = args['data_root']
if args['use_refer']:
# use refer module
file_name = 'imdb/imdb_%s_refermrf_att.npy' % args['test_split']
else:
file_name = 'imdb/imdb_%s.npy' % args['test_split']
imdb_path_val = os.path.join(root, file_name)
# assemblers for question and caption programs
question_assembler = Assembler(args['prog_vocab_path'])
caption_assembler = Assembler(args['prog_vocab_path'])
assemblers = {'ques': question_assembler, 'cap': caption_assembler}
def read_command_line():
title = 'Train explicit coreference resolution visual dialog model'
parser = argparse.ArgumentParser(description=title)
#-------------------------------------------------------------------------
# data input settings
parser.add_argument('--dataset',
default='visdial_v0.9_tiny',
help='Visdial dataset type')
parser.add_argument('--data_root',
default='data/',
help='Root to the data')
parser.add_argument('--feature_path',
default='data/resnet_res5c/',
help='Path to the image features')
parser.add_argument('--text_vocab_path',
default='',
help='Path to the vocabulary for text')
parser.add_argument('--prog_vocab_path',
default='',
help='Path to the vocabulary for programs')
parser.add_argument('--snapshot_path',
default='checkpoints/',
help='Path to save checkpoints')
#--------------------------------------------------------------------------
# specify encoder/decoder
parser.add_argument('--model',
default='nmn-cap-prog-only',
help='Name of the model, will be changed later')
parser.add_argument('--generator',
default='ques',
help='Name of the generator to use (ques | memory)')
parser.add_argument('--decoder',
default='gen',
help='Name of the decoder to use (gen | disc)')
parser.add_argument('--preload_features',
default=False,
type=bool,
help='Preload visual features on RAM')
#-------------------------------------------------------------------------
# model hyperparameters
parser.add_argument('--h_feat',
default=14,
type=int,
help='Height of visual conv feature')
parser.add_argument('--w_feat',
default=14,
type=int,
help='Width of visual conv feature')
parser.add_argument('--d_feat',
default=2048,
type=int,
help='Size of visual conv feature')
parser.add_argument('--text_embed_size',
default=300,
type=int,
help='Size of embedding for text')
parser.add_argument('--map_size',
default=1024,
type=int,
help='Size of the final mapping')
parser.add_argument('--prog_embed_size',
default=300,
type=int,
help='Size of embedding for program tokens')
parser.add_argument('--lstm_size',
default=1000,
type=int,
help='Size of hidden state in LSTM')
parser.add_argument('--enc_dropout',
default=True,
type=bool,
help='Dropout in encoder')
parser.add_argument('--dec_dropout',
default=True,
type=bool,
help='Dropout in decoder')
parser.add_argument('--num_layers',
default=2,
type=int,
help='Number of layers in LSTM')
parser.add_argument(
'--max_enc_len',
default=24,
type=int,
help='Maximum encoding length for sentences (ques|cap)')
parser.add_argument('--max_dec_len',
default=14,
type=int,
help='Maximum decoding length for programs (ques|cap)')
parser.add_argument('--dec_sampling',
default=False,
type=bool,
help='Sample while decoding programs vs argmax')
#---------------------------------------------------------------------------
parser.add_argument('--use_refer',
dest='use_refer',
action='store_true',
help='Flag to use Refer for coreference resolution')
parser.set_defaults(use_refer=False)
parser.add_argument('--use_fact',
dest='use_fact',
action='store_true',
help='Flag to use the fact in coreference pool')
parser.set_defaults(use_fact=False)
parser.add_argument('--supervise_attention',
dest='supervise_attention',
action='store_true',
help='Flag to supervise attention for the modules')
parser.set_defaults(supervise_attention=False)
parser.add_argument('--amalgam_text_feats',
dest='amalgam_text_feats',
action='store_true',
help='Flag to amalgamate text features')
parser.set_defaults(amalgam_text_feats=False)
parser.add_argument('--no_cap_alignment',
dest='cap_alignment',
action='store_false',
help='Use the auxiliary caption alignment loss')
parser.set_defaults(cap_alignment=True)
#-------------------------------------------------------------------------
# optimization params
parser.add_argument(
'--batch_size',
default=20,
type=int,
help='Training batch size (adjust based on GPU memory)')
parser.add_argument('--learning_rate',
default=1e-3,
type=float,
help='Learning rate for training')
parser.add_argument('--dropout', default=0.5, type=float, help='Dropout')
parser.add_argument('--num_epochs',
default=20,
type=int,
help='Maximum number of epochs to run training')
parser.add_argument('--gpu_id',
type=int,
default=0,
help='GPU id to use for training, -1 for CPU')
#-------------------------------------------------------------------------
try:
parsed_args = vars(parser.parse_args())
except (IOError) as msg:
parser.error(str(msg))
# set the cuda environment variable for the gpu to use
gpu_id = '' if parsed_args['gpu_id'] < 0 else str(parsed_args['gpu_id'])
print('Using GPU id: %s' % gpu_id)
os.environ['CUDA_VISIBLE_DEVICES'] = gpu_id
# pretty print arguments and return
support.pretty_print_dict(parsed_args)
return parsed_args