-
Notifications
You must be signed in to change notification settings - Fork 8
/
RL_opts.py
159 lines (143 loc) · 9.76 KB
/
RL_opts.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
import argparse
def parse_opt():
parser = argparse.ArgumentParser()
# Data input settings
parser.add_argument('--input_json', type=str, default='data/cocotalk.json',
help='path to the json file containing additional info and vocab data/flickr30ktalk.json ')
parser.add_argument('--CNN', type=str, default='Resnet101',
help='CNN features Resnet152 Resnet101')
parser.add_argument('--datasets', type=str, default='mscoco',
help='mscoco, flickr30k')
parser.add_argument('--input_label_h5', type=str, default='data/mscoco_knnlabel.h5',
help='path to the h5file containing the preprocessed dataset data/flickr30ktalk_label.h5')
parser.add_argument('--start_from', type=str, default=None,
help="""continue training from saved model at this path. Path must contain files saved by previous training process:
'infos.pkl' : configuration;
'checkpoint' : paths to model file(s) (created by tf).
Note: this file contains absolute paths, be careful when moving files around;
'model.ckpt-*' : file(s) with model definition (created by tf)
""")
# Model settings
parser.add_argument('--caption_model', type=str, default="show_tell",
help='show_tell, show_attend_tell, all_img, fc, att2in, att2in2, adaatt, adaattmo, topdown')
parser.add_argument('--rnn_size', type=int, default=1024,
help='size of the rnn in number of hidden nodes in each layer')
parser.add_argument('--num_layers', type=int, default=1,
help='number of layers in the RNN')
parser.add_argument('--rnn_type', type=str, default='lstm',
help='rnn, gru, or lstm')
parser.add_argument('--input_encoding_size', type=int, default=512,
help='the encoding size of each token in the vocabulary, and the image.')
parser.add_argument('--att_hid_size', type=int, default=512,
help='the hidden size of the attention MLP; only useful in show_attend_tell; 0 if not using hidden layer')
parser.add_argument('--fc_feat_size', type=int, default=2048,
help='2048 for resnet, 4096 for vgg')
parser.add_argument('--att_feat_size', type=int, default=2048,
help='2048 for resnet, 512 for vgg')
parser.add_argument('--self_critical_metric', type=str, default='CIDEr',
help='RL metric ,CIDEr or BLEU4 ')
# Optimization: General
parser.add_argument('--max_epochs', type=int, default=100,
help='number of epochs')
parser.add_argument('--batch_size', type=int, default=16,
help='minibatch size')
parser.add_argument('--grad_clip', type=float, default=0.1, #5.,
help='clip gradients at this value')
parser.add_argument('--drop_prob_lm', type=float, default=0.5,
help='strength of dropout in the Language Model RNN')
parser.add_argument('--self_critical_start', type=int, default=0,
help='After what iter do we start self_critical training ? (-1 = dont start ; >=0:that epoch )')
parser.add_argument('--seq_per_img', type=int, default=5,
help='number of captions to sample for each image during training. Done for efficiency since CNN forward pass is expensive. E.g. coco has 5 sents/image')
parser.add_argument('--beam_size', type=int, default=1,
help='used when sample_max = 1, indicates number of beams in beam search. Usually 2 or 3 works well. More is not better. Set this to 1 for faster runtime but a bit worse performance.')
#Optimization: for the Language Model
parser.add_argument('--optim', type=str, default='adam',
help='what update to use? rmsprop|sgd|sgdmom|adagrad|adam')
parser.add_argument('--learning_rate', type=float, default=5e-4,
help='learning rate')
parser.add_argument('--learning_rate_decay_start', type=int, default=0,
help='at what iteration to start decaying learning rate? (-1 = dont) (in epoch)')
parser.add_argument('--learning_rate_decay_every', type=int, default=10,
help='every how many iterations thereafter to drop LR?(in epoch)')
parser.add_argument('--learning_rate_decay_rate', type=float, default=0.8,
help='every how many iterations thereafter to drop LR?(in epoch)')
parser.add_argument('--optim_alpha', type=float, default=0.9,
help='alpha for adam')
parser.add_argument('--optim_beta', type=float, default=0.999,
help='beta used for adam')
parser.add_argument('--optim_epsilon', type=float, default=1e-8,
help='epsilon that goes into denominator for smoothing')
parser.add_argument('--weight_decay', type=float, default=0,
help='weight_decay')
parser.add_argument('--hdr_w', type=float, default=1.0,
help='HDR weight factor')
parser.add_argument('--bdr_w', type=float, default=1.0,
help='BDR weight factor')
parser.add_argument('--scheduled_sampling_start', type=int, default=0,
help='at what iteration to start decay gt probability')
parser.add_argument('--scheduled_sampling_increase_every', type=int, default=5,
help='every how many iterations thereafter to gt probability')
parser.add_argument('--scheduled_sampling_increase_prob', type=float, default=0.05,
help='How much to update the prob')
parser.add_argument('--scheduled_sampling_max_prob', type=float, default=0.25,
help='Maximum scheduled sampling prob.')
# Evaluation/Checkpointing
parser.add_argument('--val_images_use', type=int, default=5000,
help='how many images to use when periodically evaluating the validation loss? (-1 = all)')
parser.add_argument('--save_checkpoint_every', type=int, default=5000,
help='how often to save a model checkpoint (in iterations)?')
parser.add_argument('--checkpoint_path', type=str, default='save',
help='directory to store checkpointed models')
parser.add_argument('--language_eval', type=int, default=1,
help='Evaluate language as well (1 = yes, 0 = no)? BLEU/CIDEr/METEOR/ROUGE_L? requires coco-caption code from Github.')
parser.add_argument('--losses_log_every', type=int, default=25,
help='How often do we snapshot losses, for inclusion in the progress dump? (0 = disable)')
parser.add_argument('--load_best_score', type=int, default=1,
help='Do we load previous best score when resuming training.')
# misc
parser.add_argument('--id', type=str, default='',
help='an id identifying this run/job. used in cross-val and appended when writing progress files')
parser.add_argument('--train_only', type=int, default=0,
help='if true then use 80k, else use 110k')
# VSE model
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('--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('--num_gru', default=1, type=int,
help='Number of GRU layers.')
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='resnet152',
help="""The CNN used for image encoder
(e.g. vgg19, resnet152)""")
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.')
args = parser.parse_args()
# Check if args are valid
assert args.rnn_size > 0, "rnn_size should be greater than 0"
assert args.num_layers > 0, "num_layers should be greater than 0"
assert args.input_encoding_size > 0, "input_encoding_size should be greater than 0"
assert args.batch_size > 0, "batch_size should be greater than 0"
assert args.drop_prob_lm >= 0 and args.drop_prob_lm < 1, "drop_prob_lm should be between 0 and 1"
assert args.seq_per_img > 0, "seq_per_img should be greater than 0"
assert args.beam_size > 0, "beam_size should be greater than 0"
assert args.save_checkpoint_every > 0, "save_checkpoint_every should be greater than 0"
assert args.losses_log_every > 0, "losses_log_every should be greater than 0"
assert args.language_eval == 0 or args.language_eval == 1, "language_eval should be 0 or 1"
assert args.load_best_score == 0 or args.load_best_score == 1, "language_eval should be 0 or 1"
assert args.train_only == 0 or args.train_only == 1, "language_eval should be 0 or 1"
return args