/
BERT_multi_GPU.py
472 lines (445 loc) · 22.8 KB
/
BERT_multi_GPU.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
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
import time
from load_data import *
import class_network1 as lib
from class_network1 import *
import argparse
import os
from tqdm import tqdm, trange#, tqdm_notebook, tnrange
import torch
import torch.nn as nn
import torch.optim as optim
import pandas as pd
from tqdm import tqdm
import math
import re
from random import *
import numpy as np
import sys
from datetime import timedelta
from datetime import datetime
import torch.multiprocessing as mp
from torch.utils.data import Dataset, DataLoader
from torch.utils.data.distributed import DistributedSampler
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.distributed import init_process_group, destroy_process_group
def zero_loss():
return [0]*10
def get_attn_pad_mask(seq_q):
batch_size, len_q = seq_q.size()
#batch_size, len_k = seq_k.size()
# eq(zero) is PAD token
pad_attn_mask = seq_q.data.eq(0).unsqueeze(1) # batch_size x 1 x len_k(=len_q), one is masking
return pad_attn_mask.expand(batch_size, len_q, len_q) # batch_size x len_q x len_k
class ScaledDotProductAttention(nn.Module):
def __init__(self):
super(ScaledDotProductAttention, self).__init__()
def forward(self, Q, K, V, attn_mask):
scores = torch.matmul(Q, K.transpose(-1, -2)) / np.sqrt(64) # scores : [batch_size x 12 x len_q(=len_k) x len_k(=len_q)]
scores.masked_fill_(attn_mask, -1e9) # Fills elements of self tensor with value where mask is one.
attn = nn.Softmax(dim=-1)(scores)
context = torch.matmul(attn, V)
return context, attn
class MultiHeadAttention(nn.Module):
def __init__(self):
super(MultiHeadAttention, self).__init__()
self.W_Q = nn.Linear(50, 64 * 12)
self.W_K = nn.Linear(50, 64 * 12)
self.W_V = nn.Linear(50, 64 * 12)
self.fc = nn.Linear(12 * 64, 50)
self.norm = nn.LayerNorm(50)
def forward(self, Q, K, V, attn_mask):
"""if schema=='joint':
self.W_Q=self.W_Q.cuda()
self.W_V=self.W_V.cuda()
self.W_K=self.W_K.cuda()
self.fc=self.fc.cuda()
self.norm=self.norm.cuda()"""
# q: [batch_size x len_q x 50], k: [batch_size x len_k x 50], v: [batch_size x len_k x 50]
residual, batch_size = Q, Q.size(0)
# (B, S, D) -proj-> (B, S, D) -split-> (B, S, H, W) -trans-> (B, H, S, W)
q_s = self.W_Q(Q).view(batch_size, -1, 12, 64).transpose(1,2) # q_s: [batch_size x 12 x len_q x 64]
k_s = self.W_K(K).view(batch_size, -1, 12, 64).transpose(1,2) # k_s: [batch_size x 12 x len_k x 64]
v_s = self.W_V(V).view(batch_size, -1, 12, 64).transpose(1,2) # v_s: [batch_size x 12 x len_k x 64]
attn_mask = attn_mask.unsqueeze(1).repeat(1, 12, 1, 1) # attn_mask : [batch_size x 12 x len_q x len_k]
# context: [batch_size x 12 x len_q x 64], attn: [batch_size x 12 x len_q(=len_k) x len_k(=len_q)]
context, attn = ScaledDotProductAttention()(q_s, k_s, v_s, attn_mask)
context = context.transpose(1, 2).contiguous().view(batch_size, -1, 12 * 64) # context: [batch_size x len_q x 12 * 64]
output = self.fc(context)
output = self.norm(output + residual) # output: [batch_size x len_q x 50]
return output, attn # output: [batch_size x len_q x 50]
def gelu(x):
return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))
class PoswiseFeedForwardNet(nn.Module):
def __init__(self):
super(PoswiseFeedForwardNet, self).__init__()
self.fc1 = nn.Linear(50, 200)
self.fc2 = nn.Linear(200, 50)
def forward(self, x):
# (batch_size, len_seq, 50) -> (batch_size, len_seq, 200) -> (batch_size, len_seq, 50)
return self.fc2(gelu(self.fc1(x)))
class EncoderLayer(nn.Module):
def __init__(self):
super(EncoderLayer, self).__init__()
self.enc_self_attn = MultiHeadAttention()
self.pos_ffn = PoswiseFeedForwardNet()
def forward(self, enc_inputs, enc_self_attn_mask):
"""if schema=='joint':
self.enc_self_attn=self.enc_self_attn.cuda()
self.pos_ffn=self.pos_ffn.cuda()"""
enc_outputs, attn = self.enc_self_attn(enc_inputs, enc_inputs, enc_inputs, enc_self_attn_mask) # enc_inputs to same Q,K,V
enc_outputs = self.pos_ffn(enc_outputs) # enc_outputs: [batch_size x len_q x 50]
return enc_outputs, attn
class BERT(nn.Module):
def __init__(self):
super(BERT, self).__init__()
#self.embedding = Embedding()
self.layer1 = EncoderLayer()
self.layer2 = EncoderLayer()
self.layer3 = EncoderLayer()
self.layer4 = EncoderLayer()
self.layer5 = EncoderLayer()
self.layer6 = EncoderLayer()
#self.fc = nn.Linear(50, 50)
#self.activ1 = nn.Tanh()
self.linear = nn.Linear(50, 50)
self.activ2 = gelu
self.norm = nn.LayerNorm(50)
#self.classifier = nn.Linear(50, 2)
# decoder is shared with embedding layer
#embed_weight = self.embedding.tok_embed.weight
#n_vocab, n_dim = embed_weight.size()
self.decoder = nn.Linear(50,191929, bias=False)
#self.decoder.weight = embed_weight
self.decoder_bias = nn.Parameter(torch.zeros(191929))
self.l= nn.Linear(50,128).cuda()
self.activ=nn.Tanh()
def forward(self, output, masked_pos,input_ids,schema):
"""if schema=='joint':
self.layer1=self.layer1.cuda()
self.layer1.enc_self_attn=self.layer1.enc_self_attn.cuda()
self.layer1.pos_ffn=self.layer1.pos_ffn.cuda()
self.layer1.enc_self_attn.W_Q=self.layer1.enc_self_attn.W_Q.cuda()
self.layer1.enc_self_attn.W_K=self.layer1.enc_self_attn.W_K.cuda()
self.layer1.enc_self_attn.W_V=self.layer1.enc_self_attn.W_V.cuda()
self.layer1.enc_self_attn.fc=self.layer1.enc_self_attn.fc.cuda()
self.layer1.enc_self_attn.norm=self.layer1.enc_self_attn.norm.cuda()
self.layer1.pos_ffn.fc1=self.layer1.pos_ffn.fc1.cuda()
self.layer1.pos_ffn.fc2=self.layer1.pos_ffn.fc2.cuda()
self.layer2=self.layer2.cuda()
self.layer2.enc_self_attn=self.layer2.enc_self_attn.cuda()
self.layer2.pos_ffn=self.layer2.pos_ffn.cuda()
self.layer2.enc_self_attn.W_Q=self.layer2.enc_self_attn.W_Q.cuda()
self.layer2.enc_self_attn.W_K=self.layer2.enc_self_attn.W_K.cuda()
self.layer2.enc_self_attn.W_V=self.layer2.enc_self_attn.W_V.cuda()
self.layer2.enc_self_attn.fc=self.layer2.enc_self_attn.fc.cuda()
self.layer2.enc_self_attn.norm=self.layer2.enc_self_attn.norm.cuda()
self.layer2.pos_ffn.fc1=self.layer2.pos_ffn.fc1.cuda()
self.layer2.pos_ffn.fc2=self.layer2.pos_ffn.fc2.cuda()
self.layer3=self.layer3.cuda()
self.layer3.enc_self_attn=self.layer3.enc_self_attn.cuda()
self.layer3.pos_ffn=self.layer3.pos_ffn.cuda()
self.layer3.enc_self_attn.W_Q=self.layer3.enc_self_attn.W_Q.cuda()
self.layer3.enc_self_attn.W_K=self.layer3.enc_self_attn.W_K.cuda()
self.layer3.enc_self_attn.W_V=self.layer3.enc_self_attn.W_V.cuda()
self.layer3.enc_self_attn.fc=self.layer3.enc_self_attn.fc.cuda()
self.layer3.enc_self_attn.norm=self.layer3.enc_self_attn.norm.cuda()
self.layer3.pos_ffn.fc1=self.layer3.pos_ffn.fc1.cuda()
self.layer3.pos_ffn.fc2=self.layer3.pos_ffn.fc2.cuda()
self.layer4=self.layer4.cuda()
self.layer4.enc_self_attn=self.layer4.enc_self_attn.cuda()
self.layer4.pos_ffn=self.layer4.pos_ffn.cuda()
self.layer4.enc_self_attn.W_Q=self.layer4.enc_self_attn.W_Q.cuda()
self.layer4.enc_self_attn.W_K=self.layer4.enc_self_attn.W_K.cuda()
self.layer4.enc_self_attn.W_V=self.layer4.enc_self_attn.W_V.cuda()
self.layer4.enc_self_attn.fc=self.layer4.enc_self_attn.fc.cuda()
self.layer4.enc_self_attn.norm=self.layer4.enc_self_attn.norm.cuda()
self.layer4.pos_ffn.fc1=self.layer4.pos_ffn.fc1.cuda()
self.layer4.pos_ffn.fc2=self.layer4.pos_ffn.fc2.cuda()
self.layer5=self.layer5.cuda()
self.layer5.enc_self_attn=self.layer5.enc_self_attn.cuda()
self.layer5.pos_ffn=self.layer5.pos_ffn.cuda()
self.layer5.enc_self_attn.W_Q=self.layer5.enc_self_attn.W_Q.cuda()
self.layer5.enc_self_attn.W_K=self.layer5.enc_self_attn.W_K.cuda()
self.layer5.enc_self_attn.W_V=self.layer5.enc_self_attn.W_V.cuda()
self.layer5.enc_self_attn.fc=self.layer5.enc_self_attn.fc.cuda()
self.layer5.enc_self_attn.norm=self.layer5.enc_self_attn.norm.cuda()
self.layer5.pos_ffn.fc1=self.layer5.pos_ffn.fc1.cuda()
self.layer5.pos_ffn.fc2=self.layer5.pos_ffn.fc2.cuda()
self.layer6=self.layer6.cuda()
self.layer6.enc_self_attn=self.layer6.enc_self_attn.cuda()
self.layer6.pos_ffn=self.layer6.pos_ffn.cuda()
self.layer6.enc_self_attn.W_Q=self.layer6.enc_self_attn.W_Q.cuda()
self.layer6.enc_self_attn.W_K=self.layer6.enc_self_attn.W_K.cuda()
self.layer6.enc_self_attn.W_V=self.layer6.enc_self_attn.W_V.cuda()
self.layer6.enc_self_attn.fc=self.layer6.enc_self_attn.fc.cuda()
self.layer6.enc_self_attn.norm=self.layer6.enc_self_attn.norm.cuda()
self.layer6.pos_ffn.fc1=self.layer6.pos_ffn.fc1.cuda()
self.layer6.pos_ffn.fc2=self.layer6.pos_ffn.fc2.cuda()
self.linear=self.linear.cuda()
self.norm=self.norm.cuda()
self.decoder=self.decoder.cuda()
self.decoder_bias=self.decoder_bias.cuda()
output=output.cuda()
segment_ids=segment_ids.cuda()
masked_pos=masked_pos.cuda()
"""
#output = ws
enc_self_attn_mask = get_attn_pad_mask(input_ids)
if schema=='joint':
enc_self_attn_mask=enc_self_attn_mask.cuda()
output, enc_self_attn = self.layer1(output, enc_self_attn_mask)
output, enc_self_attn = self.layer2(output, enc_self_attn_mask)
output, enc_self_attn = self.layer3(output, enc_self_attn_mask)
output, enc_self_attn = self.layer4(output, enc_self_attn_mask)
output, enc_self_attn = self.layer5(output, enc_self_attn_mask)
output, enc_self_attn = self.layer6(output, enc_self_attn_mask)
# output : [batch_size, len, 50], attn : [batch_size, 12, d_mode, 50]
# it will be decided by first token(CLS)
#h_pooled = self.activ1(self.fc(output[:, 0])) # [batch_size, 50]
#logits_clsf = self.classifier(h_pooled) # [batch_size, 2]
masked_pos = masked_pos[:, :, None].expand(-1, -1, output.size(-1)) # [batch_size, max_pred, 50]
# get masked position from final output of transformer.
h_masked = torch.gather(output, 1, masked_pos) # masking position [batch_size, max_pred, 50]
h_masked = self.norm(self.activ2(self.linear(h_masked)))
logits_lm = self.decoder(h_masked) + self.decoder_bias # [batch_size, max_pred, n_vocab]
if schema=='joint':
op=self.l(output[:, 0])
op=self.activ(op)
else:
op=0
#else:
#op=logits_lm
return logits_lm, op
class Train_reg:
def __init__(self,model,data,optimizer,gpu_id):
self.gpu_id=gpu_id
self.model=model.to(gpu_id)
self.train_data=data
self.optimizer=optimizer
self.model.layer1=model.layer1.to(gpu_id)
self.model.layer1.enc_self_attn=model.layer1.enc_self_attn.to(gpu_id)
self.model.layer1.pos_ffn=model.layer1.pos_ffn.to(gpu_id)
self.model.layer1.enc_self_attn.W_Q=model.layer1.enc_self_attn.W_Q.to(gpu_id)
self.model.layer1.enc_self_attn.W_K=model.layer1.enc_self_attn.W_K.to(gpu_id)
self.model.layer1.enc_self_attn.W_V=model.layer1.enc_self_attn.W_V.to(gpu_id)
self.model.layer1.enc_self_attn.fc=model.layer1.enc_self_attn.fc.to(gpu_id)
self.model.layer1.enc_self_attn.norm=model.layer1.enc_self_attn.norm.to(gpu_id)
self.model.layer1.pos_ffn.fc1=model.layer1.pos_ffn.fc1.to(gpu_id)
self.model.layer1.pos_ffn.fc2=model.layer1.pos_ffn.fc2.to(gpu_id)
self.model.layer2=model.layer2.to(gpu_id)
self.model.layer2.enc_self_attn=model.layer2.enc_self_attn.to(gpu_id)
self.model.layer2.pos_ffn=model.layer2.pos_ffn.to(gpu_id)
self.model.layer2.enc_self_attn.W_Q=model.layer2.enc_self_attn.W_Q.to(gpu_id)
self.model.layer2.enc_self_attn.W_K=model.layer2.enc_self_attn.W_K.to(gpu_id)
self.model.layer2.enc_self_attn.W_V=model.layer2.enc_self_attn.W_V.to(gpu_id)
self.model.layer2.enc_self_attn.fc=model.layer2.enc_self_attn.fc.to(gpu_id)
self.model.layer2.enc_self_attn.norm=model.layer2.enc_self_attn.norm.to(gpu_id)
self.model.layer2.pos_ffn.fc1=model.layer2.pos_ffn.fc1.to(gpu_id)
self.model.layer2.pos_ffn.fc2=model.layer2.pos_ffn.fc2.to(gpu_id)
self.model.layer3=model.layer3.to(gpu_id)
self.model.layer3.enc_self_attn=model.layer3.enc_self_attn.to(gpu_id)
self.model.layer3.pos_ffn=model.layer3.pos_ffn.to(gpu_id)
self.model.layer3.enc_self_attn.W_Q=model.layer3.enc_self_attn.W_Q.to(gpu_id)
self.model.layer3.enc_self_attn.W_K=model.layer3.enc_self_attn.W_K.to(gpu_id)
self.model.layer3.enc_self_attn.W_V=model.layer3.enc_self_attn.W_V.to(gpu_id)
self.model.layer3.enc_self_attn.fc=model.layer3.enc_self_attn.fc.to(gpu_id)
self.model.layer3.enc_self_attn.norm=model.layer3.enc_self_attn.norm.to(gpu_id)
self.model.layer3.pos_ffn.fc1=model.layer3.pos_ffn.fc1.to(gpu_id)
self.model.layer3.pos_ffn.fc2=model.layer3.pos_ffn.fc2.to(gpu_id)
self.model.layer4=model.layer4.to(gpu_id)
self.model.layer4.enc_self_attn=model.layer4.enc_self_attn.to(gpu_id)
self.model.layer4.pos_ffn=model.layer4.pos_ffn.to(gpu_id)
self.model.layer4.enc_self_attn.W_Q=model.layer4.enc_self_attn.W_Q.to(gpu_id)
self.model.layer4.enc_self_attn.W_K=model.layer4.enc_self_attn.W_K.to(gpu_id)
self.model.layer4.enc_self_attn.W_V=model.layer4.enc_self_attn.W_V.to(gpu_id)
self.model.layer4.enc_self_attn.fc=model.layer4.enc_self_attn.fc.to(gpu_id)
self.model.layer4.enc_self_attn.norm=model.layer4.enc_self_attn.norm.to(gpu_id)
self.model.layer4.pos_ffn.fc1=model.layer4.pos_ffn.fc1.to(gpu_id)
self.model.layer4.pos_ffn.fc2=model.layer4.pos_ffn.fc2.to(gpu_id)
self.model.layer5=model.layer5.to(gpu_id)
self.model.layer5.enc_self_attn=model.layer5.enc_self_attn.to(gpu_id)
self.model.layer5.pos_ffn=model.layer5.pos_ffn.to(gpu_id)
self.model.layer5.enc_self_attn.W_Q=model.layer5.enc_self_attn.W_Q.to(gpu_id)
self.model.layer5.enc_self_attn.W_K=model.layer5.enc_self_attn.W_K.to(gpu_id)
self.model.layer5.enc_self_attn.W_V=model.layer5.enc_self_attn.W_V.to(gpu_id)
self.model.layer5.enc_self_attn.fc=model.layer5.enc_self_attn.fc.to(gpu_id)
self.model.layer5.enc_self_attn.norm=model.layer5.enc_self_attn.norm.to(gpu_id)
self.model.layer5.pos_ffn.fc1=model.layer5.pos_ffn.fc1.to(gpu_id)
self.model.layer5.pos_ffn.fc2=model.layer5.pos_ffn.fc2.to(gpu_id)
self.model.layer6=model.layer6.to(gpu_id)
self.model.layer6.enc_self_attn=model.layer6.enc_self_attn.to(gpu_id)
self.model.layer6.pos_ffn=model.layer6.pos_ffn.to(gpu_id)
self.model.layer6.enc_self_attn.W_Q=model.layer6.enc_self_attn.W_Q.to(gpu_id)
self.model.layer6.enc_self_attn.W_K=model.layer6.enc_self_attn.W_K.to(gpu_id)
self.model.layer6.enc_self_attn.W_V=model.layer6.enc_self_attn.W_V.to(gpu_id)
self.model.layer6.enc_self_attn.fc=model.layer6.enc_self_attn.fc.to(gpu_id)
self.model.layer6.enc_self_attn.norm=model.layer6.enc_self_attn.norm.to(gpu_id)
self.model.layer6.pos_ffn.fc1=model.layer6.pos_ffn.fc1.to(gpu_id)
self.model.layer6.pos_ffn.fc2=model.layer6.pos_ffn.fc2.to(gpu_id)
self.model.linear=model.linear.to(gpu_id)
self.model.norm=model.norm.to(gpu_id)
self.model=DDP(model,device_ids=[gpu_id],find_unused_parameters=True)
def _run_batch(self,embed_w,masked_pos,input_ids,masked_tokens):
criterion = nn.CrossEntropyLoss()
self.optimizer.zero_grad()
logits_lm, _ = self.model(embed_w, masked_pos,input_ids,'pretrain')
loss = criterion(logits_lm.transpose(1, 2), masked_tokens)
loss.backward()
#print(loss.item())
self.optimizer.step()
def _save_checkpoint(self, epoch):
ckp = self.model.module.state_dict()
PATH = 'models/bignotes/multiGPU/Model'+str(epoch)+'.pt'
torch.save(ckp, PATH)
print(f"Epoch {epoch} | Training checkpoint saved at {PATH}")
def _run_epoch(self, epoch):
b_sz = len(next(iter(self.train_data))[0])
print(f"[GPU{self.gpu_id}] Epoch {epoch} | Batchsize: {b_sz} | Steps: {len(self.train_data)}")
self.train_data.sampler.set_epoch(epoch)
for embed_w,masked_pos,input_ids,masked_tokens in self.train_data:
embed_w = embed_w.to(self.gpu_id)
masked_pos = masked_pos.to(self.gpu_id)
input_ids = input_ids.to(self.gpu_id)
masked_tokens = masked_tokens.to(self.gpu_id)
self._run_batch(embed_w,masked_pos,input_ids,masked_tokens)
def train(self, max_epochs: int):
for epoch in range(max_epochs):
self._run_epoch(epoch)
if self.gpu_id == 1 and epoch % 2 == 0:
self._save_checkpoint(epoch)
def ddp_setup(rank: int, world_size: int):
"""
Args:
rank: Unique identifier of each process
world_size: Total number of processes
"""
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "12355"
init_process_group(backend="nccl", rank=rank, world_size=world_size)
def prepare_dataloader(dataset: Dataset, batch_size: int):
return DataLoader(
dataset,
batch_size=batch_size,
pin_memory=True,
shuffle=False,
sampler=DistributedSampler(dataset)
)
def load_train_objs(embed_w,masked_pos,input_ids,masked_tokens):
torch.save(embed_w,'embed_w1.pt')
torch.save(masked_pos,'masked_pos1.pt')
torch.save(input_ids,'input_ids1.pt')
torch.save(masked_tokens,'masked_tokens1.pt')
train_set = BERTDATA(embed_w,masked_pos,input_ids,masked_tokens) # load your dataset
model = BERT()
optimizer = optim.Adam(model.parameters(), lr=0.001)
return train_set, model, optimizer
def prepare_dataloader(dataset: Dataset, batch_size: int):
return DataLoader(
dataset,
batch_size=batch_size,
pin_memory=True,
shuffle=False,
sampler=DistributedSampler(dataset)
)
def main(rank: int, world_size: int, embed_w,masked_pos,input_ids,masked_tokens, total_epochs=500, batch_size=150):
ddp_setup(rank, world_size)
dataset, model, optimizer = load_train_objs(embed_w,masked_pos,input_ids,masked_tokens)
train_data = prepare_dataloader(dataset, batch_size)
trainer = Train_reg(model, train_data, optimizer, rank)
trainer.train(total_epochs)
destroy_process_group()
class BERTDATA(Dataset):
def __init__(self,embed_w,masked_pos,input_ids,masked_tokens):
self.input_ids=input_ids
self.embed_w=embed_w
self.masked_pos=masked_pos
self.masked_tokens=masked_tokens
def __len__(self):
return len(self.embed_w)
def __getitem__(self, idx):
embw=self.embed_w[idx]
mpos=self.masked_pos[idx]
ipi=self.input_ids[idx]
mtok=self.masked_tokens[idx]
return embw,mpos,ipi,mtok
#import argparse
#parser = argparse.ArgumentParser(description='simple distributed training job')
#parser.add_argument('total_epochs', type=int, default=200, help='Total epochs to train the model')
#parser.add_argument('save_every', type=int, help='How often to save a snapshot')
#parser.add_argument('--batch_size', default=50, type=int, help='Input batch size on each device (default: 32)')
#args = parser.parse_args(args=[])
if __name__ == '__main__':
hasmap=pd.read_csv('data/Hashmap1.csv')
hm1=dict(hasmap.values)
hm= {'[PAD]': 0, '[CLS]': 1, '[SEP]': 2, '[MASK]': 3}
number_dict={0: '[PAD]', 1: '[CLS]', 2: '[SEP]', 3: '[MASK]'}
for key,value in hm1.items():
hm[key]=(value+4)
number_dict[(value+4)]=key
#print(len(hm))
note1=pd.read_csv('data/data/chunk4_encoded.csv')
#note2=pd.read_csv('data/chunk2_encoded1.csv')
#note3=pd.read_csv('data/chunk3_encoded1.csv')
#note1=pd.concat([note,note2,note3],axis=0)
print(len(note1))
note1=note1.dropna(subset=['NOTE_TEXT'])
#note1=note1[]
note1.reset_index(drop=True)
#inter=pd.read_csv('data/interactions.csv')
#inter=inter[['user_id','timestamp']]
#str_d='2007/12/01'
#d = datetime.strptime(str_d, "%Y/%m/%d")
#note1['timestamp']=note1['NOTE_DATETIME'].apply(lambda x: (x-d).days)
note1=note1[['pid','timestamp','NOTE_TEXT']]
#inter['pid']=inter['user_id']
#inter=inter[['pid','timestamp']]
sentences=[]
for row, data in tqdm(note1.iterrows()):
sentences.append(data['NOTE_TEXT'])
#print(len(sentences))
#sentences=sentences[:9000]
token_list=list()
ml=0
for sentence in sentences:
arr=[int(s) for s in sentence.split()]
if len(arr)>ml:
ml=len(arr)
token_list.append(arr)
#print(ml)
#print(token_list[:3])
maxlen = 325 # maximum of length
batch_size = len(sentences) # Full Chunk Value For now. Need to think about this
max_pred = 10 # max tokens of prediction
n_layers = 6 # number of Encoder of Encoder Layer
n_heads = 12 # number of heads in Multi-Head Attention
d_model = 50 # Embedding Size
d_ff = 50 * 4 # 4*50, FeedForward dimension
d_k = d_v = 64 # dimension of K(=Q), V
n_segments = 2
batcher=6000
vocab_size =len(hm)
#print(vocab_size)
batch=make_batch(sentences,hm,token_list,max_pred,maxlen,vocab_size,number_dict)
input_ids, _, masked_tokens, masked_pos = map(torch.LongTensor, zip(*batch))
unpickled_df = np.load("data/data/L10T50G100A1ngU_iter99.p",allow_pickle=True)
voc1=unpickled_df[1]
embed_w = torch.rand(((len(sentences)),325,50))
va=np.random.rand(50)
for i in tqdm(range(batch_size)):
for j in range(maxlen):
pos=int(input_ids[i][j])
if pos<=3:
ve=va
else:
#if pos+4>=vocab_size:
#pos=vocab_size-5
ve=voc1[(pos-4)]
embed_w[i,j,:]=torch.tensor(ve)
input_ids=input_ids
#segment_ids=segment_ids
masked_tokens=masked_tokens
embed_w=embed_w
masked_pos=masked_pos
world_size = torch.cuda.device_count()
#args.total_epochs=200
mp.spawn(main, args=(world_size,embed_w,masked_pos,input_ids,masked_tokens,500,150,), nprocs=world_size)