-
Notifications
You must be signed in to change notification settings - Fork 0
/
lm.py
373 lines (317 loc) · 17.8 KB
/
lm.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
360
361
362
363
364
365
366
367
368
369
370
371
372
373
import tqdm
import random
from collections import defaultdict
import torch
from torch.nn.functional import log_softmax
import transformers
from graphmlp import positive_reinforcement_nllloss, GraphMLP
from util import reg_result, graph_result, combined_result
from evaluation import eval_combined
EXP_NEG_40 = torch.exp(torch.tensor(-40.))
def bhattacharyya_c(p, q, is_log=True):
'''
Computes the Bhattacharyya coefficient of (log) softmax distributions p and q.
'''
assert p.size() == q.size()
p = torch.clamp_min(p, min=-40 if is_log else EXP_NEG_40)
q = torch.clamp_min(q, min=-40 if is_log else EXP_NEG_40)
return torch.nansum(torch.sqrt(torch.exp(p + q) if is_log else torch.mul(p, q)), dim=-1)
def diff(p, q, is_log=True):
'''
Computes the simple difference of (log) softmax distributions p and q.
'''
assert p.size() == q.size()
p = torch.clamp_min(p, min=-40 if is_log else EXP_NEG_40)
q = torch.clamp_min(q, min=-40 if is_log else EXP_NEG_40)
return torch.nansum(torch.abs(torch.exp(p) - torch.exp(q) if is_log else p - q), dim=-1)
def log_dist(coeff):
return lambda *args, **kwargs: -torch.log(coeff(*args, **kwargs))
bhattacharyya_d = log_dist(bhattacharyya_c)
diff_d = log_dist(diff)
class RegularizedLM(torch.nn.Module):
def __init__(self, pretrained_lm: transformers.GPT2LMHeadModel, graph_model: GraphMLP):
super(RegularizedLM, self).__init__()
self.pretrained_lm = pretrained_lm
self.graph_model = graph_model
self.vocab_size = self.pretrained_lm.config.vocab_size
def forward(self, plm_inputs, gm_inputs=None, plm_gm_tokens=None, loss_fxn=positive_reinforcement_nllloss(), lbda=.5,
eps=.5, verbose=0, **kwargs):
del kwargs
plm_inputs = plm_inputs.clone()
plm_labels = plm_inputs.clone()
plm_inputs[plm_inputs == -100] = 0
plm_outputs = self.pretrained_lm(plm_inputs, labels=plm_labels)
lm_loss = plm_outputs.loss
lm_logits = log_softmax(plm_outputs.logits, dim=-1)
asserts = [not lm_loss.isnan(), lm_loss >= 0]
assert_prints = [f'lm_loss == {lm_loss}']
assert plm_outputs.logits.size(-1) == self.vocab_size
loss = lm_loss.clone()
if gm_inputs is not None:
gm_labels = torch.gather(plm_labels, 1, plm_gm_tokens)
gm_labels[gm_labels == self.graph_model.tokenizer.eos_token_id] = -100
gm_labels = gm_labels.view(-1)
gm_mask = gm_labels != -100
gm_inputs = gm_inputs.view(-1, gm_inputs.size(-1))[gm_mask]
gm_labels = gm_labels[gm_mask]
gm_outputs = self.graph_model(gm_inputs,
labels=gm_labels,
softmax=False, loss_fxn=loss_fxn)
gm_loss = gm_outputs.loss
asserts.append(not gm_loss.isnan())
asserts.append(gm_loss >= 0)
assert_prints.append(f'gm_loss == {gm_loss}')
assert gm_outputs.logits.size(-1) == self.vocab_size
q = log_softmax(gm_outputs.logits, dim=-1)
p = torch.gather(lm_logits, 1, torch.clamp_min(
plm_gm_tokens.unsqueeze(-1).expand(-1, -1, self.vocab_size) - 1, min=0))
p = p.view(-1, self.vocab_size)[gm_mask]
p_not_nan = not torch.any(p.isnan())
q_not_nan = not torch.any(q.isnan())
asserts.append(p_not_nan)
asserts.append(q_not_nan)
assert_prints.append(f'p is {"not " if p_not_nan else ""}nan')
assert_prints.append(f'q is {"not " if q_not_nan else ""}nan')
pr_loss = torch.mean(bhattacharyya_d(p, q, is_log=True), dim=0)
# # Trick 1: no regularization for ground truth answer
# all_but_labels = torch.ones_like(p).scatter_(1, gm_labels.unsqueeze(1), 0.).bool()
# _p = p[all_but_labels]
# _q = q[all_but_labels]
# pr_loss = torch.mean(bhattacharyya_d(_p, _q, is_log=True), dim=0)
# # Trick 2: control regularization strength / direction
# with torch.no_grad():
# #TODO: do this with log probs or exp probs?
# _p_q = _p + eps * _q
# pr_loss = torch.mean(bhattacharyya_d(_p, _p_q, is_log=True), dim=0)
# Trick 3: make regularization contingent on expected improvement
# with torch.no_grad():
# per_word_ppl_loss = torch.nn.NLLLoss(reduction='none')
# p_ppls = per_word_ppl_loss(
# lm_logits[:, :-1].reshape(-1, self.vocab_size),
# plm_labels[:, 1:].reshape(-1))
# p_ppls = torch.gather(torch.cat(
# [torch.zeros(plm_gm_tokens.size(0), 1).to(p_ppls), p_ppls.view(plm_gm_tokens.size(0), -1)],
# dim=1), 1, plm_gm_tokens).view(-1)[gm_mask]
# q_ppls = per_word_ppl_loss(q, gm_labels)
# p_q_ppl_diff = p_ppls - q_ppls
# p_worse_mask = p_q_ppl_diff > 0
# q_worse_mask = p_q_ppl_diff < 0
# if verbose:
# p_better_ratio = torch.sum(q_worse_mask) / torch.sum(gm_mask)
# k = verbose
# top_k_p_better = [(self.pretrained_lm.tokenizer.decode(gm_labels[i]), d.item()) for d, i in
# zip(*torch.topk(p_q_ppl_diff, k, largest=False))]
# top_k_q_better = [(self.pretrained_lm.tokenizer.decode(gm_labels[i]), d.item()) for d, i in
# zip(*torch.topk(p_q_ppl_diff, k, largest=True))]
# print('LM ppl', torch.mean(p_ppls), ', graph ppl', torch.mean(q_ppls))
# print('LM better ratio', p_better_ratio.item(), torch.sum(q_worse_mask).item(),
# torch.sum(p_worse_mask).item())
# print('LM ppl - graph ppl')
# for x in top_k_p_better:
# print(x)
# print('...')
# for x in top_k_q_better[::-1]:
# print(x)
# print()
#
# if p_worse_mask.sum() == 0:
# pr_loss = torch.tensor(0.).to(p)
#
# else:
#
# with torch.no_grad():
# __q = q[p_worse_mask]
# # pr_p_loss = bhattacharyya_d(p[p_worse_mask], __q, is_log=True)
#
# # with torch.no_grad():
# # __p = p[q_worse_mask]
# # pr_q_loss = bhattacharyya_d(__p, q[q_worse_mask], is_log=True)
#
# # pr_loss = torch.mean(torch.cat([pr_p_loss, pr_q_loss], dim=0), dim=0)
#
# # Trick 4: use simple difference instead of Bhattacharrya distance
# # pr_loss = torch.mean(diff(p, q, is_log=True), dim=0)
# pr_loss = torch.mean(diff(p[p_worse_mask], __q, is_log=True), dim=0)
#
# # Trick 5: use mean-squared error of logits (see https://arxiv.org/pdf/2105.08919.pdf)
# # lm_logits = plm_outputs.logits
# # p = torch.gather(lm_logits, 1, plm_gm_tokens.unsqueeze(-1).expand(-1, -1, self.vocab_size))
# # p = p.view(-1, self.vocab_size)[gm_mask]
# # q = gm_outputs.logits
# # pr_loss = torch.nn.functional.mse_loss(p, q)
asserts.append(not pr_loss.isnan())
asserts.append(not (lbda * pr_loss).isnan())
asserts.append(pr_loss >= 0)
assert_prints.append(f'pr_loss == {pr_loss}')
assert_prints.append(f'lbda * pr_loss == {lbda} * {pr_loss} == {lbda * pr_loss}')
loss += gm_loss + lbda * pr_loss
asserts.append(not loss.isnan())
asserts.append(loss >= 0)
assert_prints.append(f'loss == {loss}')
assert all(asserts), (asserts, assert_prints)
return reg_result(logits=torch.gather(plm_outputs.logits, 1, torch.clamp_min(
plm_gm_tokens.unsqueeze(-1).expand(-1, -1, self.vocab_size) - 1, min=0)),
lm_result=plm_outputs,
graph_result=(gm_outputs if gm_inputs is not None \
else graph_result(loss=torch.zeros(1),
logits=torch.zeros(1, self.vocab_size),
last_hidden_state=torch.zeros(1),
cluster_size=torch.zeros(1),
cluster_ratio=torch.zeros(1))),
aux_loss=pr_loss if gm_inputs is not None else torch.zeros(1),
aux_losses=(pr_loss,) if gm_inputs is not None else (),
loss=loss)
class CombinedLM(torch.nn.Module):
def __init__(self, pretrained_lm: transformers.GPT2LMHeadModel, graph_model: GraphMLP, h_dim=768, dropout=0.2,
use_lm=True, use_graph=True):
super(CombinedLM, self).__init__()
self.pretrained_lm = pretrained_lm if use_lm else None
self.graph_model = graph_model if use_graph else None
self.tokenizer = graph_model.tokenizer
self.vocab_size = pretrained_lm.config.vocab_size
self.use_lm = use_lm
self.use_graph = use_graph
def forward(self, plm_inputs, gm_inputs, plm_gm_tokens, loss_fxn=positive_reinforcement_nllloss(),
softmax=True, lm_weight=0., gm_weight=0., aux_weight=0., **kwargs):
del kwargs
plm_inputs = plm_inputs.clone()
plm_labels = plm_inputs.clone()
plm_inputs[plm_inputs == -100] = 0
asserts = []
assert_prints = []
gm_labels = torch.gather(plm_labels, 1, plm_gm_tokens)
gm_labels[gm_labels == self.tokenizer.eos_token_id] = -100
gm_labels = gm_labels.view(-1)
gm_mask = gm_labels != -100
gm_inputs = gm_inputs.view(-1, gm_inputs.size(-1))[gm_mask]
gm_labels = gm_labels[gm_mask]
all_logits = []
sub_losses = 0
gm_outputs = None
plm_outputs = None
if self.use_graph:
gm_outputs = self.graph_model(gm_inputs, labels=gm_labels, softmax=False, loss_fxn=loss_fxn)
gm_loss = gm_outputs.loss
sub_losses += gm_weight * gm_loss.clone()
asserts.append(not gm_loss.isnan())
asserts.append(gm_loss >= 0)
assert_prints.append(f'gm_loss == {gm_loss}')
assert gm_outputs.logits.size(-1) == self.vocab_size
gm_logits = gm_outputs.logits
all_logits.append(gm_logits)
if self.use_lm:
plm_outputs = self.pretrained_lm(plm_inputs, labels=plm_labels)
lm_logits = torch.gather(plm_outputs.logits, 1, torch.clamp_min(
plm_gm_tokens.unsqueeze(-1).expand(-1, -1, self.vocab_size) - 1, min=0))
lm_logits = lm_logits.view(-1, self.vocab_size)[gm_mask]
all_logits.append(lm_logits)
lm_loss = plm_outputs.loss
sub_losses += lm_weight * lm_loss.clone()
asserts.append(not lm_loss.isnan())
asserts.append(lm_loss >= 0)
assert_prints.append(f'lm_loss == {lm_loss}')
assert plm_outputs.logits.size(-1) == self.vocab_size
out = sum(all_logits)
out_softmax = log_softmax(out, dim=-1)
pr_loss = torch.mean(bhattacharyya_d(log_softmax(lm_logits, dim=-1), out_softmax, is_log=True), dim=0)
loss1, loss2, cluster_sizes, cluster_ratios = loss_fxn(out_softmax, gm_labels)
loss = loss1 + loss2
cluster_size = cluster_sizes.sum().item()
cluster_ratio = cluster_ratios.sum().item()
total_loss = loss + sub_losses + aux_weight * pr_loss
return combined_result(logits=out_softmax if softmax else out,
loss=total_loss,
aux_loss=pr_loss,
aux_losses=(pr_loss,),
last_hidden_state=out,
lm_result=(plm_outputs if self.use_lm \
else graph_result(loss=torch.zeros(1),
logits=torch.zeros(gm_inputs.size(0), self.vocab_size).to(gm_inputs),
last_hidden_state=torch.zeros(1),
cluster_size=torch.zeros(1),
cluster_ratio=torch.zeros(1))),
graph_result=(gm_outputs if self.use_graph \
else graph_result(loss=torch.zeros(1),
logits=torch.zeros(gm_inputs.size(0), self.vocab_size).to(gm_inputs),
last_hidden_state=torch.zeros(1),
cluster_size=torch.zeros(1),
cluster_ratio=torch.zeros(1)))
)
def train(model, data, dev_data=None, n_data=None, randomize=True, train_mode=3, checkpoint_name='checkpoint.pt',
seed=42, epochs=50, plm_lr=1e-5, gm_lr=1e-4, lr=1e-4,
loss_fxn=positive_reinforcement_nllloss(), lbda=.5, lm_weight=0., gm_weight=0., aux_weight=1.):
param_groups = []
for module in model.children():
params = list(module.parameters())
if len(params) > 0 and any(p.requires_grad for p in params):
if module == model.pretrained_lm:
_lr = plm_lr
elif module == model.graph_model:
_lr = gm_lr
else:
_lr = lr
param_groups.append({'params': params, 'lr': _lr})
optim = torch.optim.AdamW(params=param_groups, weight_decay=.05)
model.train()
random.seed(seed)
best_dev_ppl = float('inf')
with tqdm.tqdm(None, total=epochs, desc=f'Total', unit_scale=True) as total_pbar:
for i in range(epochs):
with tqdm.tqdm(None, desc=f'Total - Epoch {i + 1}', total=epochs) as pbar:
pbar.update(i + 1)
total_pbar.set_description(f'Total - Epoch {i + 1}')
loss = 0
n = 0
if randomize:
random.shuffle(data)
with tqdm.tqdm(data, total=n_data, desc=f'Epoch {i + 1}') \
as pbar_batch:
for _, x_batch, l_batch, token_batch in pbar_batch:
optim.zero_grad()
model_outputs = model(l_batch, gm_inputs=x_batch, plm_gm_tokens=token_batch, loss_fxn=loss_fxn,
lbda=lbda, lm_weight=lm_weight, gm_weight=gm_weight, aux_weight=aux_weight,
c=0)
# c=1 -> gold graphs only (teacher forcing)
# c=0 -> auto graphs only
del x_batch, l_batch, token_batch
n += 1
loss += model_outputs.loss
batch_loss = model_outputs.loss # / batch_n
batch_loss.backward()
optim.step()
pbar_batch.set_postfix(batch_loss=batch_loss.item(),
lm_loss=model_outputs.lm_result.loss.item(),
g_loss=model_outputs.graph_result.loss.item(),
pr_loss=model_outputs.aux_loss.item())
pbar.set_postfix(total_loss=loss.item() / n, mem='{:.1f} MiB'.format(torch.cuda.max_memory_allocated() / 1000000))
if n_data is not None:
total_pbar.update(1 / n_data)
del model_outputs
torch.cuda.empty_cache()
if dev_data is not None:
model.eval()
domains = defaultdict(str)
dev_eval, _ = eval_combined(model, model.tokenizer, dev_data, domains, interesting_n=1,
device=next(model.parameters()).device)
torch.cuda.empty_cache()
print()
for m in ('gold2', 'auto2', 'lm2'):
for p in ('ppl', 'acc'):
print(i, m, p, dev_eval['all'][m][p])
dev_ppl = dev_eval['all']['auto2' if train_mode >= 6 else 'gold2']['ppl']
if dev_ppl < best_dev_ppl:
best_dev_ppl = dev_ppl
print('saving checkpoint at epoch', i, 'with best ppl', dev_ppl)
if train_mode in (0, 2):
with open(checkpoint_name.replace('_model', '_graph_model'), 'wb') as f:
torch.save(model.graph_model.state_dict(), f)
if train_mode in (1, 2):
with open(checkpoint_name.replace('_model', '_lm_model'), 'wb') as f:
torch.save(model.pretrained_lm.state_dict(), f)
if train_mode in (3, 4, 5):
with open(checkpoint_name, 'wb') as f:
torch.save(model.state_dict(), f)
if train_mode in (6, 7):
with open(checkpoint_name, 'wb') as f:
torch.save(model.state_dict(), f)
model.train()