-
Notifications
You must be signed in to change notification settings - Fork 11
/
options.py
178 lines (162 loc) · 9.63 KB
/
options.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
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
import os
import argparse
from six import iteritems
from itertools import product
from time import gmtime, strftime
def readCommandLine(argv=None):
parser = argparse.ArgumentParser(description='Train and Test the Visual Dialog model')
#-------------------------------------------------------------------------
# Data input settings
parser.add_argument('-inputImg', default='data/visdial/data_img.h5',
help='HDF5 file with image features')
parser.add_argument('-inputQues', default='data/visdial/chat_processed_data.h5',
help='HDF5 file with preprocessed questions')
parser.add_argument('-inputJson', default='data/visdial/chat_processed_params.json',
help='JSON file with info and vocab')
parser.add_argument('-inputDenseJson', default='data/visdial/visdial_1.0_val_dense_annotations.json',
help='JSON file with dense annotations')
parser.add_argument('-cocoDir', default='',
help='Directory for coco images, optional')
parser.add_argument('-cocoInfo', default='',
help='JSON file with coco split information')
#-------------------------------------------------------------------------
# Logging settings
parser.add_argument('-verbose', type=int, default=1,
help='Level of verbosity (default 1 prints some info)',
choices=[1, 2])
parser.add_argument('-savePath', default='checkpoints/',
help='Path to save checkpoints')
parser.add_argument('-saveName', default='',
help='Name of save directory within savePath')
parser.add_argument('-startFrom', type=str, default='',
help='Copy weights from model at this path')
parser.add_argument('-qstartFrom', type=str, default='',
help='Copy weights from qbot model at this path')
parser.add_argument('-continue', action='store_true',
help='Continue training from last epoch')
parser.add_argument('-enableVisdom', type=int, default=0,
help='Flag for enabling visdom logging')
parser.add_argument('-visdomEnv', type=str, default='',
help='Name of visdom environment for plotting')
parser.add_argument('-visdomServer', type=str, default='127.0.0.1',
help='Address of visdom server instance')
parser.add_argument('-visdomServerPort', type=int, default=8893,
help='Port of visdom server instance')
#-------------------------------------------------------------------------
# Model params for both a-bot and q-bot
parser.add_argument('-randomSeed', default=32, type=int,
help='Seed for random number generators')
parser.add_argument('-imgEmbedSize', default=300, type=int,
help='Size of the multimodal embedding')
parser.add_argument('-imgFeatureSize', default=4096, type=int,
help='Size of the image feature')
parser.add_argument('-embedSize', default=300, type=int,
help='Size of input word embeddings')
parser.add_argument('-rnnHiddenSize', default=512, type=int,
help='Size of the LSTM state')
parser.add_argument('-numLayers', default=2, type=int,
help='Number of layers in LSTM')
parser.add_argument('-imgNorm', default=1, type=int,
help='Normalize the image feature. 1=yes, 0=no')
parser.add_argument('-AbotMCTS', default=0, type=int,
help='Running Rollouts for rewards calculation for Abot. 1=yes, 0=no')
# A-Bot encoder + decoder
parser.add_argument('-encoder', default='hre-ques-lateim-hist',
help='Name of the encoder to use',
choices=['hre-ques-lateim-hist'])
parser.add_argument('-decoder', default='gen',
help='Name of the decoder to use (gen)',
choices=['gen'])
# Q-bot encoder + decoder
parser.add_argument('-qencoder', default='hre-ques-lateim-hist',
help='Name of the encoder to use',
choices=['hre-ques-lateim-hist'])
parser.add_argument('-qdecoder', default='gen',
help='Name of the decoder to use (only gen supported now)',
choices=['gen'])
#-------------------------------------------------------------------------
# Optimization / training params
parser.add_argument('-trainMode', default='rl-full-QAf',
help='What should train.py do?',
choices=['sl-abot', 'sl-qbot', 'rl-full-QAf'])
parser.add_argument('-numRounds', default=10, type=int,
help='Number of rounds of dialog (max 10)')
parser.add_argument('-batchSize', default=20, type=int,
help='Batch size (number of threads) '
'(Adjust base on GPU memory)')
parser.add_argument('-learningRate', default=1e-3, type=float,
help='Learning rate')
parser.add_argument('-minLRate', default=5e-5, type=float,
help='Minimum learning rate')
parser.add_argument('-dropout', default=0.0, type=float, help='Dropout')
parser.add_argument('-numEpochs', default=85, type=int, help='Epochs')
parser.add_argument('-lrDecayRate', default=0.9997592083, type=float,
help='Decay for learning rate')
parser.add_argument('-CELossCoeff', default=1, type=float,
help='Coefficient for cross entropy loss')
parser.add_argument('-RLLossCoeff', default=20000, type=float,
help='Coefficient for cross entropy loss')
parser.add_argument('-useCosSimilarityLoss', default=1, type=int,
help='whether to use similarity loss')
parser.add_argument('-CosSimilarityLossCoeff', default=0.1, type=float,
help='Coefficient for similarity loss')
parser.add_argument('-useHuberLoss', default=1, type=int,
help='whether to use Huber loss')
parser.add_argument('-HuberLossCoeff', default=1, type=float,
help='Coefficient for Huber loss')
parser.add_argument('-featLossCoeff', default=1000, type=float,
help='Coefficient for feature regression loss')
parser.add_argument('-useCurriculum', default=1, type=int,
help='Use curriculum or for RL training (1) or not (0)')
parser.add_argument('-freezeQFeatNet', default=0, type=int,
help='Freeze weights of Q-bot feature network')
parser.add_argument('-rlAbotReward', default=1, type=int,
help='Choose whether RL reward goes to A-Bot')
# annealing params"
parser.add_argument('-annealingEndRound', default=3, type=int, help='Round at which annealing ends')
parser.add_argument('-annealingReduceEpoch',default=1,type=int, help='Num epochs at which annealing happens')
# Other training environmnet settings
parser.add_argument('-useGPU', action='store_true', help='Use GPU or CPU')
parser.add_argument('-numWorkers', default=2, type=int,
help='Number of worker threads in dataloader')
#-------------------------------------------------------------------------
# Evaluation params
parser.add_argument('-beamSize', default=1, type=int,
help='Beam width for beam-search sampling')
parser.add_argument('-evalModeList', default=[], nargs='+',
help='What task should the evaluator perform?',
choices=['ABotRank', 'QBotRank', 'QABotsRank', 'dialog','human_study'])
parser.add_argument('-evalSplit', default='val',
choices=['train', 'val', 'test'])
parser.add_argument('-evalTitle', default='eval',
help='If generating a plot, include this in the title')
parser.add_argument('-startEpoch',default=1, type=int,help='Starting epoch for evaluation')
parser.add_argument('-endEpoch',default=1, type=int,help='Last epoch for evaluation')
parser.add_argument('-useNDCG', action='store_true',
help='Whether to use NDCG in evaluation')
parser.add_argument('-discountFactor',default=0.5,type=float,help="discount factor for future rewards")
#-------------------------------------------------------------------------
try:
parsed = vars(parser.parse_args(args=argv))
except IOError as msg:
parser.error(str(msg))
if parsed['saveName']:
# Custom save file path
parsed['savePath'] = os.path.join(parsed['savePath'],
parsed['saveName'])
else:
# Standard save path with time stamp
import random
timeStamp = strftime('%d-%b-%y-%X-%a', gmtime())
parsed['savePath'] = os.path.join(parsed['savePath'], timeStamp)
parsed['savePath'] += '_{:0>6d}'.format(random.randint(0, 10e6))
# check if history is needed
parsed['useHistory'] = True if 'hist' in parsed['encoder'] else False
# check if image is needed
if 'lateim' in parsed['encoder']:
parsed['useIm'] = 'late'
elif 'im' in parsed['encoder']:
parsed['useIm'] = True
else:
parsed['useIm'] = False
return parsed