-
Notifications
You must be signed in to change notification settings - Fork 15
/
Copy pathmain.py
253 lines (212 loc) · 7.35 KB
/
main.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
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
from shared import *
from tf_lib import *
from dataset import *
from model import *
#from dialog_gui import *
from classifier import load_classifier
"""
AUTHOR: Xiang Gao ([email protected]) at Microsoft Research
"""
def run_master(mode, args):
if mode not in ['train','continue'] and args.restore != '':
aa = args.restore.split('/')
bb = []
for a in aa:
if len(a) > 0:
bb.append(a)
fld = '/'.join(bb[:-1])
if mode in ['vali','vis']:
vocab_only = False
fld_data, _, _ = get_model_fld(args)
path_bias_vocab = fld_data + '/vocab_bias.txt'
else:
vocab_only = True
fld_data = fld
path_bias_vocab = fld + '/vocab_bias.txt'
else:
vocab_only = False
fld_data, fld_model, subfld = get_model_fld(args)
fld = fld_model + '/' + subfld
path_bias_vocab = fld_data + '/vocab_bias.txt'
if os.path.exists(path_bias_vocab):
allowed_words = [line.strip('\n').strip('\r') for line in open(path_bias_vocab, encoding='utf-8')]
else:
allowed_words = None
model_class = args.model_class.lower()
if model_class.startswith('fuse'):
Master = StyleFusion
elif model_class == 'mtask':
Master = VanillaMTask
elif model_class == 's2s':
Master = Seq2Seq
elif model_class == 'lm':
Master = LanguageModel
elif model_class == 's2s+lm':
pass
else:
raise ValueError
if model_class == 's2s+lm':
master = Seq2SeqLM(args, allowed_words)
else:
dataset = Dataset(fld_data,
max_ctxt_len=args.max_ctxt_len, max_resp_len=args.max_resp_len,
vocab_only=vocab_only, noisy_vocab=args.noisy_vocab)
master = Master(dataset, fld, args, new=(mode=='train'), allowed_words=allowed_words)
if mode != 'train':
if args.restore.endswith('.npz') or model_class == 's2s+lm':
restore_path = args.restore
else:
restore_path = master.fld + '/models/%s.npz'%args.restore
master.load_weights(restore_path)
if mode in ['vis', 'load']:
return master
if args.clf_name.lower() == 'holmes':
CLF_NAMES = ['classifier/Reddit_vs_Holmes/neural', 'classifier/Reddit_vs_Holmes/ngram']
elif args.clf_name.lower() == 'arxiv':
CLF_NAMES = ['classifier/Reddit_vs_arxiv/neural', 'classifier/Reddit_vs_arxiv/ngram']
else:
CLF_NAMES = [args.clf_name]
print('loading classifiers '+str(CLF_NAMES))
master.clf_names = CLF_NAMES
master.classifiers = []
for clf_name in CLF_NAMES:
master.classifiers.append(load_classifier(clf_name))
print('\n'+fld+'\n')
if mode in ['continue', 'train']:
ss = ['', mode + ' @ %i'%time.time()]
for k in sorted(args.__dict__.keys()):
ss.append('%s = %s'%(k, args.__dict__[k]))
with open(master.fld + '/args.txt', 'a') as f:
f.write('\n'.join(ss)+'\n')
if args.debug:
batch_per_load = 1
else:
if PHILLY:
n_sample = 1280 # philly unstable for large memory
else:
n_sample = 2560
batch_per_load = int(n_sample/BATCH_SIZE)
if mode == 'continue':
master.vali()
master.train(batch_per_load)
elif 'summary' == mode:
print(master.model.summary())
elif mode in ['cmd', 'test', 'vali']:
classifiers = []
for clf_name in CLF_NAMES:
classifiers.append(load_classifier(clf_name))
if 'vali' == mode:
data = master.get_vali_data()
s_decoded = eval_decoded(master, data,
classifiers=classifiers, corr_by_tgt=True, r_rand=args.r_rand,
calc_retrieval=('holmes' in args.data_name.lower())
)[0]
s_surrogate = eval_surrogate(master, data)[0]
print(restore_path)
print()
print(s_decoded)
print()
print(s_surrogate)
return
"""
if model_class != 's2s+lm':
with tf.variable_scope('base_rankder', reuse=tf.AUTO_REUSE):
fld_base_ranker = 'restore/%s/%s/pretrained/'%(args.model_class.replace('fuse1','fuse'), args.data_name)
dataset_base_ranker = Dataset(fld_base_ranker,
max_ctxt_len=args.max_ctxt_len, max_resp_len=args.max_resp_len,
vocab_only=True, noisy_vocab=False)
base_ranker = Master(dataset_base_ranker, fld_base_ranker, args, new=False, allowed_words=master.allowed_words)
path = fld_base_ranker + '/' + open(fld_base_ranker+'/base_ranker.txt').readline().strip('\n')
base_ranker.load_weights(path)
print('*'*10 + ' base_ranker loaded from: '+path)
else:
"""
base_ranker = None
def print_results(results):
ss = ['total', 'logP', 'logP_c', 'logP_b', 'rep', 'len',] + ['clf%i'%i for i in range(len(CLF_NAMES))]
print('; '.join([' '*(6-len(s))+s for s in ss]))
for score, resp, terms in results:
print('%6.3f; '%score + '; '.join(['%6.3f'%x for x in terms]) + '; ' + resp)
if 'cmd' == mode:
while True:
print('\n---- please input ----')
inp = input()
infer_args = parse_infer_args()
if inp == '':
break
results = infer_rank(inp, master, infer_args, base_ranker=base_ranker)
print_results(results)
elif 'test' == mode:
infer_args = parse_infer_args()
path_out = args.path_test+'.hyp'
open(path_out, 'w', encoding='utf-8')
for line in open(args.path_test, encoding='utf-8'):
line = line.strip('\n')
inp = line.split('\t')[0]
results = infer_rank(inp, master, infer_args, base_ranker=base_ranker)
lines = []
for _, hyp, _ in results[:min(10, len(results))]:
lines.append(line + '\t' + hyp.replace(' _EOS_','').strip())
with open(path_out, 'a', encoding='utf-8') as f:
f.write('\n'.join(lines) + '\n')
"""
path_in = DATA_PATH + '/test/' + args.test_fname
if not PHILLY:
fld_out = master.fld + '/eval2/'
else:
fld_out = OUT_PATH
makedirs(fld_out)
npz_name = args.restore.split('/')[-1].replace('.npz','')
path_out = fld_out + '/' + args.test_fname + '_' + npz_name
test_master(master, path_in, path_out, max_n_src=args.test_n_max, base_ranker=base_ranker, baseline=args.baseline, r_rand=args.r_rand)
"""
else:
raise ValueError
def get_model_fld(args):
data_name = args.data_name
if PHILLY:
data_name = data_name.replace('+','').replace('_','')
fld_data = DATA_PATH +'/' + data_name
master_config = 'width%s_depth%s'%(
(args.token_embed_dim, args.rnn_units),
(args.encoder_depth, args.decoder_depth))
if args.max_ctxt_len != 90 or args.max_resp_len != 30:
master_config += '_len' + str((args.max_ctxt_len, args.max_resp_len))
master_config = master_config.replace("'",'')
fld_model = OUT_PATH
if args.debug:
fld_model += '/debug'
fld_model += '/' + args.data_name.replace('../','') + '_' + master_config
subfld = []
"""
if args.randmix:
s_mix = 'randmix'
if args.ratio05 > 0:
s_mix += '(0.5=%.2f)'%args.ratio05
else:
"""
s_mix = 'mix'
model_class = args.model_class.lower()
if model_class == 's2s':
subfld = ['s2s_%s(%.2f)'%(s_mix, args.conv_mix_ratio)] # no conv data
else:
subfld = ['%s_%s(%.2f,%.2f)'%(model_class, s_mix, args.conv_mix_ratio, args.nonc_mix_ratio)]
if args.noisy_vocab > 0:
subfld.append('unk%.1fk'%(args.noisy_vocab/1000))
if model_class.startswith('fuse'):
subfld.append('std%.1f'%args.stddev)
if args.reld:
subfld.append('reld')
subfld.append('lr'+str(args.lr))
if len(args.fld_suffix) > 0:
subfld.append(args.fld_suffix)
subfld = '_'.join(subfld)
return fld_data, fld_model.replace(' ',''), subfld.replace(' ','')
if __name__ == '__main__':
parser.add_argument('mode')
parser.add_argument('--skip', type=float, default=0.0)
parser.add_argument('--test_fname', default='')
parser.add_argument('--r_rand', '-r', type=float, default=-1)
parser.add_argument('--test_n_max', '-n', type=int, default=2000)
args = parser.parse_args()
run_master(args.mode, args)