forked from YueDongCS/EditNTS
/
main.py
306 lines (255 loc) · 12 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
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
#!/usr/bin/env python
# coding:utf8
from __future__ import print_function
import os
import argparse
import collections
import logging
import numpy as np
import torch
import torch.nn as nn
import data
from checkpoint import Checkpoint
from editnts import EditNTS
from evaluator import Evaluator
PAD = 'PAD' # This has a vocab id, which is used to represent out-of-vocabulary words [0]
UNK = 'UNK' # This has a vocab id, which is used to represent out-of-vocabulary words [1]
KEEP = 'KEEP' # This has a vocab id, which is used for copying from the source [2]
DEL = 'DEL' # This has a vocab id, which is used for deleting the corresponding word [3]
START = 'START' # this has a vocab id, which is uded for indicating start of the sentence for decoding [4]
STOP = 'STOP' # This has a vocab id, which is used to stop decoding [5]
PAD_ID = 0 # This has a vocab id, which is used to represent out-of-vocabulary words [0]
UNK_ID = 1 # This has a vocab id, which is used to represent out-of-vocabulary words [1]
KEEP_ID = 2 # This has a vocab id, which is used for copying from the source [2]
DEL_ID = 3 # This has a vocab id, which is used for deleting the corresponding word [3]
START_ID = 4 # this has a vocab id, which is uded for indicating start of the sentence for decoding [4]
STOP_ID = 5 # This has a vocab id, which is used to stop decoding [5]
def sort_by_lens(seq, seq_lengths):
seq_lengths_sorted, sort_order = seq_lengths.sort(descending=True)
seq_sorted = seq.index_select(0, sort_order)
return seq_sorted, seq_lengths_sorted, sort_order
def reweigh_batch_loss(target_id_bath):
pad_c = 0
unk_c = 0
keep_c = 0
del_c = 0
start_c = 0
stop_c = 0
other_c = 0
new_edits_ids_l = target_id_bath
for i in new_edits_ids_l:
# start_c += 1
# stop_c += 1
for ed in i:
if ed == PAD_ID:
pad_c += 1
elif ed == UNK_ID:
unk_c += 1
elif ed == KEEP_ID:
keep_c += 1
elif ed == DEL_ID:
del_c += 1
elif ed == START_ID:
start_c +=1
elif ed == STOP_ID:
stop_c +=1
else:
other_c += 1
NLL_weight = np.zeros(30006) + (1 / other_c+1)
NLL_weight[PAD_ID] = 0 # pad
NLL_weight[UNK_ID] = 1. / unk_c+1
NLL_weight[KEEP_ID] = 1. / keep_c+1
NLL_weight[DEL_ID] = 1. / del_c+1
NLL_weight[5] = 1. / stop_c+1
NLL_weight_t = torch.from_numpy(NLL_weight).float().cuda()
# print(pad_c, unk_c, start_c, stop_c, keep_c, del_c, other_c)
return NLL_weight_t
def reweight_global_loss(w_add,w_keep,w_del):
# keep, del, other, (0, 65304, 246768, 246768, 2781648, 3847848, 2016880) pad,start,stop,keep,del,add
NLL_weight = np.ones(30006)+w_add
NLL_weight[PAD_ID] = 0 # pad
NLL_weight[KEEP_ID] = w_keep
NLL_weight[DEL_ID] = w_del
return NLL_weight
def training(edit_net, nepochs, args, vocab, print_every=100, check_every=500, test=False):
if test:
print(args.data_path + 'test.df.filtered.pos')
eval_dataset = data.Dataset(args.data_path + 'test.df.filtered.pos') # load eval dataset
else:
print(args.data_path + 'val.df.filtered.pos')
eval_dataset = data.Dataset(args.data_path + 'val.df.filtered.pos') # load eval dataset
evaluator = Evaluator(loss= nn.NLLLoss(ignore_index=vocab.w2i['PAD'], reduction='none'))
editnet_optimizer = torch.optim.Adam(edit_net.parameters(),
lr=1e-3, weight_decay=1e-6)
# scheduler = MultiStepLR(abstract_optimizer, milestones=[20,30,40], gamma=0.1)
# abstract_scheduler = ReduceLROnPlateau(abstract_optimizer, mode='max')
# uncomment this part to re-weight different operations
# NLL_weight = reweight_global_loss(args.w_add, args.w_keep, args.w_del)
# NLL_weight_t = torch.from_numpy(NLL_weight).float().cuda()
# editnet_criterion = nn.NLLLoss(weight=NLL_weight_t, ignore_index=vocab.w2i['PAD'], reduce=False)
editnet_criterion = nn.NLLLoss(ignore_index=vocab.w2i['PAD'], reduction='none')
best_eval_loss = 0. # init statistics
print_loss = [] # Reset every print_every
for epoch in range(nepochs):
# scheduler.step()
#reload training for every epoch
if os.path.isfile(args.data_path+'train.df.filtered.pos'):
train_dataset = data.Dataset(args.data_path + 'train.df.filtered.pos')
else: # iter chunks and vocab_data
train_dataset = data.Datachunk(args.data_path + 'train.df.filtered.pos')
for i, batch_df in train_dataset.batch_generator(batch_size=args.batch_size, shuffle=True):
# time1 = time.time()
prepared_batch, syn_tokens_list = data.prepare_batch(batch_df, vocab, args.max_seq_len) #comp,scpn,simp
# a batch of complex tokens in vocab ids, sorted in descending order
org_ids = prepared_batch[0]
org_lens = org_ids.ne(0).sum(1)
org = sort_by_lens(org_ids, org_lens) # inp=[inp_sorted, inp_lengths_sorted, inp_sort_order]
# a batch of pos-tags in pos-tag ids for complex
org_pos_ids = prepared_batch[1]
org_pos_lens = org_pos_ids.ne(0).sum(1)
org_pos = sort_by_lens(org_pos_ids, org_pos_lens)
out = prepared_batch[2][:, :]
tar = prepared_batch[2][:, 1:]
simp_ids = prepared_batch[3]
editnet_optimizer.zero_grad()
output = edit_net(org, out, org_ids, org_pos,simp_ids)
##################calculate loss
tar_lens = tar.ne(0).sum(1).float()
tar_flat=tar.contiguous().view(-1)
loss = editnet_criterion(output.contiguous().view(-1, vocab.count), tar_flat).contiguous()
loss[tar_flat == 1] = 0 #remove loss for UNK
loss = loss.view(tar.size())
loss = loss.sum(1).float()
loss = loss/tar_lens
loss = loss.mean()
print_loss.append(loss.item())
loss.backward()
torch.nn.utils.clip_grad_norm_(edit_net.parameters(), 1.)
editnet_optimizer.step()
if i % print_every == 0:
log_msg = 'Epoch: %d, Step: %d, Loss: %.4f' % (
epoch,i, np.mean(print_loss))
print_loss = []
print(log_msg)
# Checkpoint
if i % check_every == 0:
edit_net.eval()
val_loss, bleu_score, sari, sys_out = evaluator.evaluate(eval_dataset, vocab, edit_net,args)
log_msg = "epoch %d, step %d, Dev loss: %.4f, Bleu score: %.4f, Sari: %.4f \n" % (epoch, i, val_loss, bleu_score, sari)
print(log_msg)
if val_loss < best_eval_loss:
best_eval_loss = val_loss
Checkpoint(model=edit_net,
opt=editnet_optimizer,
epoch=epoch, step=i,
).save(args.store_dir)
print("checked after %d steps"%i)
edit_net.train()
print(edit_net)
return edit_net
dataset='newsela'
def set_seed(seed=0):
print("Setting all seeds to stop random..")
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
np.random.seed(seed)
# random.seed(seed) # not using random now
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
def main():
# torch.manual_seed(233)
set_seed(233)
logging.basicConfig(level=logging.INFO, format='%(asctime)s [INFO] %(message)s')
parser = argparse.ArgumentParser()
parser.add_argument('--data_path', type=str,dest='data_path',
default='/home/ml/ydong26/data/EditNTS_data/editnet_data/%s/'%dataset,
help='Path to train vocab_data')
parser.add_argument('--store_dir', action='store', dest='store_dir',
default='/home/ml/ydong26/tmp_store/editNTS_%s'%dataset,
help='Path to exp storage directory.')
parser.add_argument('--vocab_path', type=str, dest='vocab_path',
default='../vocab_data/',
help='Path contains vocab, embedding, postag_set')
parser.add_argument('--load_model', type=str, dest='load_model',
default=None,
help='Path for loading pre-trained model for further training')
parser.add_argument('--vocab_size', dest='vocab_size', default=30000, type=int)
parser.add_argument('--batch_size', dest='batch_size', default=32, type=int)
parser.add_argument('--max_seq_len', dest='max_seq_len', default=100)
parser.add_argument('--epochs', type=int, default=50)
parser.add_argument('--hidden', type=int, default=200)
parser.add_argument('--lr', type=float, default=1e-4)
parser.add_argument('--device', type=int, default=1,
help='select GPU')
parser.add_argument('--test', action='store_true', default=False, dest='test_enabled')
parser.add_argument('--run_eval', action='store_true', default=False, dest='run_eval')
parser.add_argument('--run_training', action='store_true', default=False, dest='run_training')
#train_file = '/media/vocab_data/yue/TS/editnet_data/%s/train.df.filtered.pos'%dataset
# test='/media/vocab_data/yue/TS/editnet_data/%s/test.df.pos' % args.dataset
args = parser.parse_args()
print(args)
torch.cuda.set_device(args.device)
# load vocab-related files and init vocab
print('*'*10)
vocab = data.Vocab()
vocab.add_vocab_from_file(args.vocab_path+'vocab.txt', args.vocab_size)
vocab.add_embedding(gloveFile=args.vocab_path+'glove.6B.100d.txt')
pos_vocab = data.POSvocab(args.vocab_path) #load pos-tags embeddings
print('*' * 10)
print(args)
print("generating config")
hyperparams=collections.namedtuple(
'hps', #hyper=parameters
['vocab_size', 'embedding_dim',
'word_hidden_units', 'sent_hidden_units',
'pretrained_embedding', 'word2id', 'id2word',
'pos_vocab_size', 'pos_embedding_dim']
)
hps = hyperparams(
vocab_size=vocab.count,
embedding_dim=100,
word_hidden_units=args.hidden,
sent_hidden_units=args.hidden,
pretrained_embedding=vocab.embedding,
word2id=vocab.w2i,
id2word=vocab.i2w,
pos_vocab_size=pos_vocab.count,
pos_embedding_dim=30
)
print('init editNTS model')
edit_net = EditNTS(hps, n_layers=1)
edit_net.cuda()
if args.load_model is not None:
print("load edit_net for further training")
ckpt_path = args.load_model
ckpt = Checkpoint.load(ckpt_path)
print("Epoch: {} | Step: {}".format(ckpt.epoch, ckpt.step))
edit_net = ckpt.model
edit_net.cuda()
edit_net.train()
if args.run_eval:
print("Running Evaluation..")
eval_standalone(edit_net, args, vocab, ckpt)
elif args.run_training:
print("Running Training..")
training(edit_net, args.epochs, args, vocab, test=args.test_enabled)
else:
print("ERROR: No running mode selected")
def eval_standalone(edit_net, args, vocab, ckpt):
eval_dataset = data.Dataset(args.data_path + 'test.df.filtered.pos') # load eval dataset
edit_net.eval()
evaluator = Evaluator(loss=nn.NLLLoss(ignore_index=vocab.w2i['PAD'], reduction='none'))
path = args.data_path
dataset_name = path[path.rindex("/") + 1:len(path)]
val_loss, bleu_score, sari, sys_out = evaluator.evaluate(eval_dataset, vocab, edit_net, args,
dataset_name=dataset_name, eval_dir=args.load_model)
log_msg = "epoch %d, step %d, Dev loss: %.4f, Bleu score: %.4f, Sari: %.4f \n" \
% (ckpt.epoch, ckpt.step, val_loss, bleu_score, sari)
print("Output: val_loss: {}, bleu_score: {}, sari:{}"
.format(val_loss, bleu_score, sari))
print(log_msg)
if __name__ == '__main__':
import os
cwd = os.getcwd()
print(cwd)
main()