-
Notifications
You must be signed in to change notification settings - Fork 9
/
train.py
359 lines (329 loc) · 18.1 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
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
# coding: utf-8
import argparse
from model import *
from util import *
import matplotlib.pyplot as plt
import seaborn as sb
import pandas as pd
from tqdm import tqdm
# from tqdm import tqdm_notebook as tqdm # Comment this line if using jupyter notebook
parser = argparse.ArgumentParser(description='Training HiCE on WikiText-103')
'''
Dataset arguments
'''
parser.add_argument('--w2v_dir', type=str, default='./data/base_w2v/wiki_all.sent.split.model',
help='location of the default node embedding')
parser.add_argument('--corpus_dir', type=str, default='./data/wikitext-103/',
help='location of the training corpus (wikitext-103)')
parser.add_argument('--freq_lbound', type=int, default=16,
help='Lower bound of word frequency in w2v for selecting target words')
parser.add_argument('--freq_ubound', type=int, default=2 ** 16,
help='Upper bound of word frequency in w2v for selecting target words')
parser.add_argument('--cxt_lbound', type=int, default=2,
help='Lower bound of word frequency in corpus for selecting target words')
parser.add_argument('--chimera_dir', type=str, default='./data/chimeras/',
help='location of the testing corpus (Chimeras)')
parser.add_argument('--cuda', type=int, default=-1,
help='Avaiable GPU ID')
'''
Model hyperparameters
'''
parser.add_argument('--maxlen', type=int, default=12,
help='maxlen of context (half, left or right) and character')
parser.add_argument('--use_morph', action='store_true',
help='initial learning rate')
parser.add_argument('--n_head', type=int, default=10,
help='number of hidden units per layer')
parser.add_argument('--n_layer', type=int, default=2,
help='number of encoding layers')
parser.add_argument('--n_epochs', type=int, default=200,
help='upper bound of training epochs')
parser.add_argument('--n_batch', type=int, default=256,
help='batch size')
parser.add_argument('--batch_size', type=int, default=128,
help='batch size')
parser.add_argument('--lr_init', type=float, default=1e-3,
help='initial learning rate for Adam')
parser.add_argument('--n_shot', type=int, default=10,
help='upper bound of training K-shot')
'''
Validation & Test arguments
'''
parser.add_argument('--test_interval', type=int, default=1,
help='report interval')
parser.add_argument('--save_dir', type=str, default='./save/',
help='location for saving the best model')
parser.add_argument('--lr_decay', type=float, default=0.5,
help='Learning Rate Decay using ReduceLROnPlateau Scheduler')
parser.add_argument('--threshold', type=float, default=1e-3,
help='threshold for ReduceLROnPlateau Scheduler judgement')
parser.add_argument('--patience', type=int, default=4,
help='Patience for ReduceLROnPlateau Scheduler judgement')
parser.add_argument('--lr_early_stop', type=float, default=1e-5,
help='the lower bound of training lr. Early stop after lr is below it.')
'''
Adaptation with First-Order MAML arguments
'''
parser.add_argument('--adapt', action='store_true',
help='adapt to target dataset with 1-st order MAML')
parser.add_argument('--inner_batch_size', type=int, default=4,
help='batch for updating using source corpus')
parser.add_argument('--meta_batch_size', type=int, default=16,
help='batch for accumulating meta gradients')
args = parser.parse_args()
def get_batch(words, dataset, w2v, batch_size, k_shot, device):
sample_words = np.random.choice(words, batch_size)
contexts = []
targets = []
vocabs = []
for word in sample_words:
if len(dataset[word]) != 0:
sample_sent_idx = np.random.choice(len(dataset[word]), k_shot)
sample_sents = dataset[word][sample_sent_idx]
contexts += [sample_sents]
targets += [w2v.wv[word]]
vocabs += [[_vocab[vi] for vi in word if vi in _vocab]]
contexts = torch.LongTensor(contexts).to(device)
targets = Variable(torch.FloatTensor(targets).to(device))
vocabs = torch.LongTensor(pad_sequences(vocabs, maxlen = args.maxlen * 2)).to(device)
return contexts, targets, vocabs
def evaluate_on_chimera(model, chimera_data):
'''
Evaluate the model on Chimera datasets
'''
model.eval()
with torch.no_grad():
for k_shot in chimera_data:
data = chimera_data[k_shot]
test_contexts = torch.LongTensor(data['contexts']).to(device)
test_targets = torch.FloatTensor(data['ground_truth_vector']).to(device)
test_vocabs = torch.LongTensor(data['character']).to(device)
test_pred = model.forward(test_contexts, test_vocabs)
cosine = F.cosine_similarity(test_pred, test_targets).mean().cpu().tolist()
test_prb = np.array(list(data["probes"]))
test_scr = np.array(list(data["scores"]))
cors = []
prov = [[base_w2v.wv[pi] for pi in probe] for probe in test_prb]
for p1, p2, p3 in zip(test_pred.cpu().numpy(), prov, test_scr):
cos = cosine_similarity([p1], p2)
cor = spearmanr(cos[0], p3)[0]
cors += [cor]
print('-' * 100)
print("Test with %d shot: Cosine: %.4f; Spearman: %.4f" % (k_shot, cosine, np.average(cors)))
def replace_grad(parameter_gradients, parameter_name):
'''
Creates a backward hook function that replaces the calculated gradient
with a precomputed value when .backward() is called.
See
https://pytorch.org/docs/stable/autograd.html?highlight=hook#torch.Tensor.register_hook
for more info
'''
def replace_grad_(module):
return parameter_gradients[parameter_name]
return replace_grad_
'''
Use the same word embedding model as Nounce2vec and A la Carte for fair comparison.
Note that during training, some of words in Wikitext-103 might not occur in this word embedding.
'''
base_w2v = Word2Vec.load(args.w2v_dir)
_vocab = {v: i+1 for v, i in zip('abcdefghijklmnopqrstuvwxyz', range(26))}
source_train_dataset, source_valid_dataset, dictionary = load_training_corpus(base_w2v, args.corpus_dir, \
maxlen = args.maxlen, freq_lbound = args.freq_lbound, freq_ubound = args.freq_ubound, cxt_lbound = args.cxt_lbound)
chimera_data = load_chimera(dictionary = dictionary, base_w2v = base_w2v, chimera_dir = args.chimera_dir)
device = torch.device("cuda:%d" % args.cuda if args.cuda != -1 else "cpu")
model = HICE(n_head = args.n_head, n_hid = base_w2v.vector_size, n_seq = args.maxlen * 2, \
n_layer= args.n_layer, w2v = dictionary.idx2vec, use_morph=args.use_morph).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr = args.lr_init)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=args.lr_decay, patience = args.patience, mode='max', threshold=args.threshold)
source_train_words = list(source_train_dataset.keys())
source_valid_words = list(source_valid_dataset.keys())
loss_stat = []
best_valid_cosine = -1
for epoch in np.arange(args.n_epochs) + 1:
print('=' * 100)
train_cosine = []
valid_cosine = []
model.train()
with tqdm(np.arange(args.n_batch), desc='Train') as monitor:
for batch in monitor:
k_shot = np.random.randint(args.n_shot) + 1 # randomly sample a context length, and only give the model with this size of contexts
train_contexts, train_targets, train_vocabs = get_batch(words = source_train_words, dataset = source_train_dataset, \
w2v = base_w2v, batch_size = args.batch_size, k_shot = k_shot, device = device)
optimizer.zero_grad()
pred_emb = model.forward(train_contexts, train_vocabs)
loss = -F.cosine_similarity(pred_emb, train_targets).mean()
loss.backward()
optimizer.step()
train_cosine += [[-loss.cpu().detach().numpy(), k_shot]]
monitor.set_postfix(train_status = train_cosine[-1])
model.eval()
with torch.no_grad():
with tqdm(np.arange(args.n_batch // args.n_shot), desc='Valid') as monitor:
for batch in monitor:
for k_shot in np.arange(args.n_shot) + 1: # during evaluation, use all the possible context length
valid_contexts, valid_targets, valid_vocabs = get_batch(words = source_valid_words, dataset = source_valid_dataset, \
w2v = base_w2v, batch_size = args.batch_size, k_shot = k_shot, device = device)
pred_emb = model.forward(valid_contexts, valid_vocabs)
loss = -F.cosine_similarity(pred_emb, valid_targets).mean()
valid_cosine += [[-loss.cpu().numpy(), k_shot]]
monitor.set_postfix(valid_status = valid_cosine[-1])
print('-' * 100)
avg_train, avg_valid = np.average(train_cosine, axis=0)[0], np.average(valid_cosine, axis=0)[0]
print(("Epoch: %d: Train Cosine: %.4f; Valid Cosine: %.4f; LR: %f") \
% (epoch, avg_train, avg_valid, optimizer.param_groups[0]['lr']))
scheduler.step(avg_valid)
if avg_valid > best_valid_cosine:
best_valid_cosine = avg_valid
with open(os.path.join(args.save_dir, 'model.pt'), 'wb') as f:
torch.save(model, f)
with open(os.path.join(args.save_dir, 'optimizer.pt'), 'wb') as f:
torch.save(optimizer.state_dict(), f)
loss_stat += [[epoch, li, 'TRAIN', k_shot] for (li, k_shot) in train_cosine]
loss_stat += [[epoch, li, 'VALID', k_shot] for (li, k_shot) in valid_cosine]
if epoch % args.test_interval == 0:
'''
#This script can plot loss curve and position attention weight for debugging.
plot_stat = pd.DataFrame(loss_stat, columns=['Epoch', 'Cosine', 'Data', 'K-shot'])
print(model.bal)
print(model.pos_att.pos_att)
for k_shot in [2,4,6]:
data = plot_stat[plot_stat['K-shot'] == k]
sb.lineplot(x='Epoch', y='Cosine', hue='Data', data = data)
plt.title('K-shot = ' + str(k))
plt.savefig(args.save_dir + 'training_curve_%d.png' % k)
'''
evaluate_on_chimera(model, chimera_data)
if optimizer.param_groups[0]['lr'] < args.lr_early_stop:
print('Finish Training')
break
'''
Evaluate on the best model:
'''
model = torch.load(os.path.join(args.save_dir, 'model.pt')).to(device)
print('=' * 100)
print('Evaluate on the best model with supervised training on source corpus:')
evaluate_on_chimera(model, chimera_data)
if args.adapt:
best_score = -1
'''
Use the other words that are not OOV in the target corpus for adapting the previous learned model into target domain.
'''
target_train_dataset, target_valid_dataset, target_dictionary = load_training_corpus(base_w2v, args.chimera_dir, maxlen = args.maxlen,\
freq_lbound = args.freq_lbound, freq_ubound = args.freq_ubound, cxt_lbound = args.cxt_lbound, dictionary = dictionary)
target_train_words = list(target_train_dataset.keys())
target_valid_words = list(target_valid_dataset.keys())
model.update_embedding(target_dictionary.idx2vec)
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr = args.lr_init * args.lr_decay)
optimizer.zero_grad()
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=args.lr_decay, patience = args.patience, mode='max', threshold=args.threshold)
'''
Use a temp model to calculate update on source task, then calculate the gradient with updated weights on target task.
Finally pass the gradient to original model and conduct optimization (gradient descent)
'''
model_tmp = copy.deepcopy(model)
for meta_epoch in np.arange(args.n_epochs):
print('=' * 100)
source_cosine = []
target_cosine = []
meta_grads = []
with tqdm(np.arange(args.meta_batch_size), desc='Meta Train') as monitor:
for meta_batch in monitor:
'''
Initialized from model, which will be learned to facilitate fast adaptation.
'''
model_tmp.load_state_dict(model.state_dict())
model_tmp.train()
optimizer_tmp = torch.optim.Adam(model_tmp.parameters(), lr = 5e-4)
'''
Cumulate Inner Gradient on source corpus: \theta^* = \theta - \alpha \nabla_\theta \mathcal{L}_{D_T}(\theta)
'''
for inner_batch in range(args.inner_batch_size):
k_shot = np.random.randint(args.n_shot) + 1
source_train_contexts, source_train_targets, source_train_vocabs = get_batch(words = source_train_words, \
dataset = source_train_dataset, w2v = base_w2v, batch_size = args.batch_size, k_shot = k_shot, device = device)
optimizer_tmp.zero_grad()
pred_emb = model_tmp.forward(source_train_contexts, source_train_vocabs)
loss = -F.cosine_similarity(pred_emb, source_train_targets).mean()
loss.backward()
optimizer_tmp.step()
'''
Calculate the gradient with \theta^* on target corpus: \nabla_\theta \mathcal{L}_{D_N}(\theta^*)
'''
model_tmp.eval()
optimizer_tmp.zero_grad()
k_shot = np.random.randint(args.n_shot) + 1
target_train_contexts, target_train_targets, target_train_vocabs = get_batch(words = target_train_words, \
dataset = target_train_dataset, w2v = base_w2v, batch_size = args.batch_size, k_shot = k_shot, device = device)
pred_emb = model_tmp.forward(target_train_contexts, target_train_vocabs)
loss = -F.cosine_similarity(pred_emb, target_train_targets).mean()
loss.backward()
meta_grads += [{name: param.grad for (name, param) in model_tmp.named_parameters() if param.requires_grad}]
# end for meta_batch in tqdm:
# end with tqdm(np.arange(args.meta_batch_size), desc='Meta Train') as monitor:
'''
Meta-Update with the average of meta gradients: \theta' = \theta - \beta \nabla_\theta \mathcal{L}_{D_N}(\theta^*)
Use replace_grad function as a hook to assign the gradients. Such operation will omit the second-order gradient,
Nichol et al. (https://arxiv.org/abs/1803.02999) shows that this first-order cutoff will not sacrifice too much
performance while gaining good training efficiency.
'''
meta_grads = {name: torch.stack([name_grad[name] for name_grad in meta_grads]).mean(dim=0)
for name in meta_grads[0].keys()}
hooks = []
for name, param in model.named_parameters():
if param.requires_grad:
hooks.append(
param.register_hook(replace_grad(meta_grads, name))
)
model.train()
optimizer.zero_grad()
pred_emb = model.forward(target_train_contexts, target_train_vocabs)
# Here the data (forwad, loss) doesn't matter at all, as the gradient will be replaced with meta_grads when "loss.backward()" is called.
loss = -F.cosine_similarity(pred_emb, target_train_targets).mean()
loss.backward()
optimizer.step()
'''
After such optimization, the model will later serve as initialization for model_tmp, making it generalize to
target corpus by only updating on source corpus.
'''
for h in hooks:
h.remove()
'''
Validate using either of the updated model_tmp (we use the last one for convenience)
'''
model_tmp.eval()
with tqdm(np.arange(args.n_batch // args.n_shot), desc='Meta Valid') as monitor:
for batch in monitor:
for k_shot in np.arange(args.n_shot) + 1:
source_valid_contexts, source_valid_targets, source_valid_vocabs = get_batch(words = source_valid_words, \
dataset = source_valid_dataset, w2v = base_w2v, batch_size = args.batch_size, k_shot = k_shot, device = device)
pred_emb = model_tmp.forward(source_valid_contexts, source_valid_vocabs)
source_cosine += [F.cosine_similarity(pred_emb, source_valid_targets).mean().cpu().detach().numpy()]
target_valid_contexts, target_valid_targets, target_valid_vocabs = get_batch(words = target_valid_words, \
dataset = target_valid_dataset, w2v = base_w2v, batch_size = args.batch_size, k_shot = k_shot, device = device)
pred_emb = model_tmp.forward(target_valid_contexts, target_valid_vocabs)
target_cosine += [F.cosine_similarity(pred_emb, target_valid_targets).mean().cpu().detach().numpy()]
print('-' * 100)
avg_train, avg_valid = np.average(source_cosine), np.average(target_cosine)
print(("Epoch: %d: Meta Train Cosine: %.4f; Meta Valid Cosine: %.4f; LR: %f") \
% (meta_epoch, avg_train, avg_valid, optimizer.param_groups[0]['lr']))
score = avg_train + avg_valid * 2
scheduler.step(score)
if score > best_score:
best_score = score
with open(os.path.join(args.save_dir, 'meta_model.pt'), 'wb') as f:
torch.save(model_tmp, f)
evaluate_on_chimera(model_tmp, chimera_data)
print('-' * 100)
if optimizer.param_groups[0]['lr'] < args.lr_early_stop:
print('Finish Training')
break
# end for meta_epoch in np.arange(args.n_epochs):
'''
Evaluate on the best meta model:
'''
model = torch.load(os.path.join(args.save_dir, 'meta_model.pt')).to(device)
print('=' * 100)
print('Evaluate on the best model with meta training on both source and target corpus:')
evaluate_on_chimera(model, chimera_data)
# end if args.adapt: