-
Notifications
You must be signed in to change notification settings - Fork 44
/
train.py
304 lines (273 loc) · 11.7 KB
/
train.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
import torch
import torch.nn as nn
from torch.nn.utils import clip_grad_norm_
from models import EncoderDecoder
from data_utils import DataLoader
import constants, time, os, shutil, logging, h5py
def NLLcriterion(vocab_size):
"construct NLL criterion"
weight = torch.ones(vocab_size)
weight[constants.PAD] = 0
## The first dimension is not batch, thus we need
## to average over the batch manually
#criterion = nn.NLLLoss(weight, size_average=False)
criterion = nn.NLLLoss(weight, reduction='sum')
return criterion
def KLDIVcriterion(vocab_size):
"construct KLDIV criterion"
# weight = torch.ones(vocab_size)
# weight[constants.PAD] = 0
# ## The first dimension is not batch, thus we need
# ## to average over the batch manually
# criterion = nn.KLDivLoss(weight, size_average=False)
criterion = nn.KLDivLoss(reduction='sum')
return criterion
def KLDIVloss(output, target, criterion, V, D):
"""
output (batch, vocab_size)
target (batch,)
criterion (nn.KLDIVLoss)
V (vocab_size, k)
D (vocab_size, k)
"""
## (batch, k) index in vocab_size dimension
## k-nearest neighbors for target
indices = torch.index_select(V, 0, target)
## (batch, k) gather along vocab_size dimension
outputk = torch.gather(output, 1, indices)
## (batch, k) index in vocab_size dimension
targetk = torch.index_select(D, 0, target)
return criterion(outputk, targetk)
def KLDIVloss2(output, target, criterion, V, D):
"""
constructing full target distribution, expensive!
"""
indices = torch.index_select(V, 0, target)
targetk = torch.index_select(D, 0, target)
fulltarget = torch.zeros(output.size()).scatter_(1, indices, targetk)
## here: need Variable(fulltarget).cuda() if use gpu
fulltarget = fulltarget.cuda()
return criterion(output, fulltarget)
def dist2weight(D, dist_decay_speed=0.8):
D = D.div(100)
D = torch.exp(-D * dist_decay_speed)
s = D.sum(dim=1, keepdim=True)
D = D / s
## The PAD should not contribute to the decoding loss
D[constants.PAD, :] = 0.0
return D
def genLoss(gendata, m0, m1, lossF, args):
"""
One batch loss
Input:
gendata: a named tuple contains
gendata.src (seq_len1, batch): input tensor
gendata.lengths (1, batch): lengths of source sequences
gendata.trg (seq_len2, batch): target tensor.
m0: map input to output.
m1: map the output of EncoderDecoder into the vocabulary space and do
log transform.
lossF: loss function.
---
Output:
loss
"""
input, lengths, target = gendata.src, gendata.lengths, gendata.trg
if args.cuda and torch.cuda.is_available():
input, lengths, target = input.cuda(), lengths.cuda(), target.cuda()
## (seq_len2, batch, hidden_size)
output = m0(input, lengths, target)
batch = output.size(1)
loss = 0
## we want to decode target in range [BOS+1:EOS]
target = target[1:]
for o, t in zip(output.split(args.generator_batch),
target.split(args.generator_batch)):
## (seq_len, generator_batch, hidden_size) =>
## (seq_len*generator_batch, hidden_size)
o = o.view(-1, o.size(2))
o = m1(o)
## (seq_len*generator_batch,)
t = t.view(-1)
loss += lossF(o, t)
return loss.div(batch)
def disLoss(a, p, n, m0, triplet_loss, args):
"""
a (named tuple): anchor data
p (named tuple): positive data
n (named tuple): negative data
"""
a_src, a_lengths, a_invp = a.src, a.lengths, a.invp
p_src, p_lengths, p_invp = p.src, p.lengths, p.invp
n_src, n_lengths, n_invp = n.src, n.lengths, n.invp
if args.cuda and torch.cuda.is_available():
a_src, a_lengths, a_invp = a_src.cuda(), a_lengths.cuda(), a_invp.cuda()
p_src, p_lengths, p_invp = p_src.cuda(), p_lengths.cuda(), p_invp.cuda()
n_src, n_lengths, n_invp = n_src.cuda(), n_lengths.cuda(), n_invp.cuda()
## (num_layers * num_directions, batch, hidden_size)
a_h, _ = m0.encoder(a_src, a_lengths)
p_h, _ = m0.encoder(p_src, p_lengths)
n_h, _ = m0.encoder(n_src, n_lengths)
## (num_layers, batch, hidden_size * num_directions)
a_h = m0.encoder_hn2decoder_h0(a_h)
p_h = m0.encoder_hn2decoder_h0(p_h)
n_h = m0.encoder_hn2decoder_h0(n_h)
## take the last layer as representations (batch, hidden_size * num_directions)
a_h, p_h, n_h = a_h[-1], p_h[-1], n_h[-1]
return triplet_loss(a_h[a_invp], p_h[p_invp], n_h[n_invp])
def init_parameters(model):
for p in model.parameters():
p.data.uniform_(-0.1, 0.1)
def savecheckpoint(state, is_best, args):
torch.save(state, args.checkpoint)
if is_best:
shutil.copyfile(args.checkpoint, os.path.join(args.data, 'best_model.pt'))
def validate(valData, model, lossF, args):
"""
valData (DataLoader)
"""
m0, m1 = model
## switch to evaluation mode
m0.eval()
m1.eval()
num_iteration = valData.size // args.batch
if valData.size % args.batch > 0: num_iteration += 1
total_genloss = 0
for iteration in range(num_iteration):
gendata = valData.getbatch_generative()
with torch.no_grad():
genloss = genLoss(gendata, m0, m1, lossF, args)
total_genloss += genloss.item() * gendata.trg.size(1)
## switch back to training mode
m0.train()
m1.train()
return total_genloss / valData.size
def train(args):
logging.basicConfig(filename=os.path.join(args.data, "training.log"), level=logging.INFO)
trainsrc = os.path.join(args.data, "train.src")
traintrg = os.path.join(args.data, "train.trg")
trainmta = os.path.join(args.data, "train.mta")
trainData = DataLoader(trainsrc, traintrg, trainmta, args.batch, args.bucketsize)
print("Reading training data...")
trainData.load(args.max_num_line)
print("Allocation: {}".format(trainData.allocation))
print("Percent: {}".format(trainData.p))
valsrc = os.path.join(args.data, "val.src")
valtrg = os.path.join(args.data, "val.trg")
valmta = os.path.join(args.data, "val.mta")
if os.path.isfile(valsrc) and os.path.isfile(valtrg):
valData = DataLoader(valsrc, valtrg, valmta, args.batch, args.bucketsize, True)
print("Reading validation data...")
valData.load()
assert valData.size > 0, "Validation data size must be greater than 0"
print("Loaded validation data size {}".format(valData.size))
else:
print("No validation data found, training without validating...")
## create criterion, model, optimizer
if args.criterion_name == "NLL":
criterion = NLLcriterion(args.vocab_size)
lossF = lambda o, t: criterion(o, t)
else:
assert os.path.isfile(args.knearestvocabs),\
"{} does not exist".format(args.knearestvocabs)
print("Loading vocab distance file {}...".format(args.knearestvocabs))
with h5py.File(args.knearestvocabs, "r") as f:
V, D = f["V"][...], f["D"][...]
V, D = torch.LongTensor(V), torch.FloatTensor(D)
D = dist2weight(D, args.dist_decay_speed)
if args.cuda and torch.cuda.is_available():
V, D = V.cuda(), D.cuda()
criterion = KLDIVcriterion(args.vocab_size)
lossF = lambda o, t: KLDIVloss(o, t, criterion, V, D)
triplet_loss = nn.TripletMarginLoss(margin=1.0, p=2)
m0 = EncoderDecoder(args.vocab_size,
args.embedding_size,
args.hidden_size,
args.num_layers,
args.dropout,
args.bidirectional)
m1 = nn.Sequential(nn.Linear(args.hidden_size, args.vocab_size),
nn.LogSoftmax(dim=1))
if args.cuda and torch.cuda.is_available():
print("=> training with GPU")
m0.cuda()
m1.cuda()
criterion.cuda()
#m0 = nn.DataParallel(m0, dim=1)
else:
print("=> training with CPU")
m0_optimizer = torch.optim.Adam(m0.parameters(), lr=args.learning_rate)
m1_optimizer = torch.optim.Adam(m1.parameters(), lr=args.learning_rate)
## load model state and optmizer state
if os.path.isfile(args.checkpoint):
print("=> loading checkpoint '{}'".format(args.checkpoint))
logging.info("Restore training @ {}".format(time.ctime()))
checkpoint = torch.load(args.checkpoint)
args.start_iteration = checkpoint["iteration"]
best_prec_loss = checkpoint["best_prec_loss"]
m0.load_state_dict(checkpoint["m0"])
m1.load_state_dict(checkpoint["m1"])
m0_optimizer.load_state_dict(checkpoint["m0_optimizer"])
m1_optimizer.load_state_dict(checkpoint["m1_optimizer"])
else:
print("=> no checkpoint found at '{}'".format(args.checkpoint))
logging.info("Start training @ {}".format(time.ctime()))
best_prec_loss = float('inf')
#print("=> initializing the parameters...")
#init_parameters(m0)
#init_parameters(m1)
## here: load pretrained wrod (cell) embedding
num_iteration = 67000*128 // args.batch
print("Iteration starts at {} "
"and will end at {}".format(args.start_iteration, num_iteration-1))
## training
for iteration in range(args.start_iteration, num_iteration):
try:
m0_optimizer.zero_grad()
m1_optimizer.zero_grad()
## generative loss
gendata = trainData.getbatch_generative()
genloss = genLoss(gendata, m0, m1, lossF, args)
## discriminative loss
disloss_cross, disloss_inner = 0, 0
if args.use_discriminative and iteration % 10 == 0:
a, p, n = trainData.getbatch_discriminative_cross()
disloss_cross = disLoss(a, p, n, m0, triplet_loss, args)
a, p, n = trainData.getbatch_discriminative_inner()
disloss_inner = disLoss(a, p, n, m0, triplet_loss, args)
loss = genloss + args.discriminative_w * (disloss_cross + disloss_inner)
## compute the gradients
loss.backward()
## clip the gradients
clip_grad_norm_(m0.parameters(), args.max_grad_norm)
clip_grad_norm_(m1.parameters(), args.max_grad_norm)
## one step optimization
m0_optimizer.step()
m1_optimizer.step()
## average loss for one word
avg_genloss = genloss.item() / gendata.trg.size(0)
if iteration % args.print_freq == 0:
print("Iteration: {0:}\tGenerative Loss: {1:.3f}\t"\
"Discriminative Cross Loss: {2:.3f}\tDiscriminative Inner Loss: {3:.3f}"\
.format(iteration, avg_genloss, disloss_cross, disloss_inner))
if iteration % args.save_freq == 0 and iteration > 0:
prec_loss = validate(valData, (m0, m1), lossF, args)
if prec_loss < best_prec_loss:
best_prec_loss = prec_loss
logging.info("Best model with loss {} at iteration {} @ {}"\
.format(best_prec_loss, iteration, time.ctime()))
is_best = True
else:
is_best = False
print("Saving the model at iteration {} validation loss {}"\
.format(iteration, prec_loss))
savecheckpoint({
"iteration": iteration,
"best_prec_loss": best_prec_loss,
"m0": m0.state_dict(),
"m1": m1.state_dict(),
"m0_optimizer": m0_optimizer.state_dict(),
"m1_optimizer": m1_optimizer.state_dict()
}, is_best, args)
except KeyboardInterrupt:
break