-
Notifications
You must be signed in to change notification settings - Fork 1
/
models.py
773 lines (638 loc) · 33.3 KB
/
models.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
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torch.utils.data
from torch.optim import Optimizer
from torch.optim.lr_scheduler import LambdaLR
from IPython import display
import pickle
import os
import time
from datetime import datetime
import matplotlib
matplotlib.use('AGG')
import matplotlib.pyplot as plt
from utils import snr_db2sigma, errors_ber, errors_bitwise_ber, errors_bler, min_sum_log_sum_exp, moving_average, extract_block_errors, extract_block_nonerrors
from polar import *
from pac_code import *
import math
import random
import numpy as np
from tqdm import tqdm
from collections import namedtuple
import sys
import csv
MODEL = 'gpt'
def get_pad_mask(seq, pad_idx):
return (seq != pad_idx).unsqueeze(-2)
def get_subsequent_mask(seq):
''' For masking out the subsequent info. '''
sz_b, len_s = seq.size()
subsequent_mask = (1 - torch.triu(
torch.ones((1, len_s, len_s), device=seq.device), diagonal=1)).bool()
return subsequent_mask
class ScaledDotProductAttention(nn.Module):
''' Scaled Dot-Product Attention '''
def __init__(self, temperature, attn_dropout=0.1):
super().__init__()
self.temperature = temperature
self.dropout = nn.Dropout(attn_dropout)
def forward(self, q, k, v, mask=None,causal=False):
attn = torch.matmul(q / self.temperature, k.transpose(2, 3))
if mask is not None:
if MODEL == 'gpt':
attn = attn.masked_fill(mask == 0, -1e9)
else:
mask=mask.unsqueeze(1)
attn = attn.masked_fill(mask == 0, -1e9)
attn = self.dropout(F.softmax(attn, dim=-1))
output = torch.matmul(attn, v)
return output, attn
class ScalarMult(nn.Module):
'''scalar multiplication layer'''
def __init__(self):
super().__init__()
self.alpha = nn.Parameter(1e-10*torch.ones(1))
def forward(self, x):
out = self.alpha*x
return out
class MultiHeadAttention(nn.Module):
''' Multi-Head Attention module '''
def __init__(self, n_head, d_model, d_k, d_v, dropout=0.1):
super().__init__()
self.n_head = n_head
self.d_k = d_k
self.d_v = d_v
self.w_qs = nn.Linear(d_model, n_head * d_k, bias=False)
self.w_ks = nn.Linear(d_model, n_head * d_k, bias=False)
self.w_vs = nn.Linear(d_model, n_head * d_v, bias=False)
self.fc = nn.Linear(n_head * d_v, d_model, bias=False)
self.attention = ScaledDotProductAttention(temperature=d_k ** 0.5)
self.dropout = nn.Dropout(dropout)
self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)
self.scalar = ScalarMult()
def forward(self, q, k, v, mask=None,causal=False):
d_k, d_v, n_head = self.d_k, self.d_v, self.n_head
sz_b, len_q, len_k, len_v = q.size(0), q.size(1), k.size(1), v.size(1)
residual = q
# Pass through the pre-attention projection: b x lq x (n*dv)
# Separate different heads: b x lq x n x dv
q = self.w_qs(q).view(sz_b, len_q, n_head, d_k)
k = self.w_ks(k).view(sz_b, len_k, n_head, d_k)
v = self.w_vs(v).view(sz_b, len_v, n_head, d_v)
# Transpose for attention dot product: b x n x lq x dv
q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
if mask is not None:
mask = mask.unsqueeze(1) # For head axis broadcasting.
q, attn = self.attention(q, k, v, mask=mask)
# if len(list(q.size()))==4:
# q = q.view(q.size(0)*sz_b,q.size(2),q.size(3),q.size(4)).transpose(1, 2).contiguous().view(sz_b, len_q, -1)
# else:
# Transpose to move the head dimension back: b x lq x n x dv
# Combine the last two dimensions to concatenate all the heads together: b x lq x (n*dv)
q = q.transpose(1, 2).contiguous().view(sz_b, len_q, -1)
q = self.dropout(self.fc(q))
#q = self.scalar(q)
q += residual
q = self.layer_norm(q)
return q, attn
class PositionwiseFeedForward(nn.Module):
''' A two-feed-forward-layer module '''
def __init__(self, d_in, d_hid, dropout=0.1):
super().__init__()
self.w_1 = nn.Linear(d_in, d_hid) # position-wise
self.w_2 = nn.Linear(d_hid, d_in) # position-wise
self.layer_norm = nn.LayerNorm(d_in, eps=1e-6)
self.dropout = nn.Dropout(dropout)
self.scalar = ScalarMult()
def forward(self, x):
residual = x
x = self.w_2(F.gelu(self.w_1(x))) #F.gelu
x = self.dropout(x)
#x = self.scalar(x)
x += residual
x = self.layer_norm(x)
return x
class EncoderLayer(nn.Module):
''' Compose with two layers '''
def __init__(self, d_model, d_inner, n_head, d_k, d_v, dropout=0.1):
super(EncoderLayer, self).__init__()
self.slf_attn = MultiHeadAttention(n_head, d_model, d_k, d_v, dropout=dropout)
self.pos_ffn = PositionwiseFeedForward(d_model, d_inner, dropout=dropout)
def forward(self, enc_input, slf_attn_mask=None):
enc_output, enc_slf_attn = self.slf_attn(
enc_input, enc_input, enc_input, mask=slf_attn_mask)
enc_output = self.pos_ffn(enc_output)
return enc_output, enc_slf_attn
class DecoderLayer(nn.Module):
''' Compose with three layers '''
def __init__(self, d_model, d_inner, n_head, d_k, d_v, dropout=0.1):
super(DecoderLayer, self).__init__()
self.slf_attn = MultiHeadAttention(n_head, d_model, d_k, d_v, dropout=dropout)
self.enc_attn = MultiHeadAttention(n_head, d_model, d_k, d_v, dropout=dropout)
self.pos_ffn = PositionwiseFeedForward(d_model, d_inner, dropout=dropout)
def forward(
self, dec_input, enc_output,
slf_attn_mask=None, dec_enc_attn_mask=None, cross_attend=True):
dec_enc_attn=[]
dec_output, dec_slf_attn = self.slf_attn(
dec_input, dec_input, dec_input, mask=slf_attn_mask)
if cross_attend:
dec_output, dec_enc_attn = self.enc_attn(
dec_output, enc_output, enc_output, mask=dec_enc_attn_mask)
dec_output = self.pos_ffn(dec_output)
return dec_output, dec_slf_attn, dec_enc_attn
class PositionalEncoding(nn.Module):
def __init__(self, d_hid, n_position=200,num=10000):
super(PositionalEncoding, self).__init__()
# Not a parameter
self.register_buffer('pos_table', self._get_sinusoid_encoding_table(n_position, d_hid,num))
def _get_sinusoid_encoding_table(self, n_position, d_hid,num):
''' Sinusoid position encoding table '''
# TODO: make it with torch instead of numpy
def get_position_angle_vec(position,num):
return [position / np.power(num, 2 * (hid_j // 2) / d_hid) for hid_j in range(d_hid)]
sinusoid_table = np.array([get_position_angle_vec(pos_i,num) for pos_i in range(n_position)])
sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2]) # dim 2i
sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2]) # dim 2i+1
return torch.FloatTensor(sinusoid_table).unsqueeze(0)
def forward(self, x):
return x + self.pos_table[:, :x.size(1)].clone().detach()
class XFormerEncoder(nn.Module):
def __init__(self, config, layer_idx=None):
super(XFormerEncoder,self).__init__()
self.embed_dim = config.embed_dim
self.block_len = config.max_len
self.pos_emb = nn.Embedding(config.N+1, config.embed_dim,padding_idx=0)
self.position_enc = PositionalEncoding(self.embed_dim, n_position=self.block_len)
self.dropout = nn.Dropout(p=config.dropout)
self.layer_stack = nn.ModuleList([
EncoderLayer(config.embed_dim, config.embed_dim*4, config.n_head, config.embed_dim//config.n_head, config.embed_dim//config.n_head, dropout=config.dropout)
for _ in range(config.n_layers)])
self.layer_norm = nn.LayerNorm(config.embed_dim, eps=1e-6)
def forward(self,noisy_enc,src_mask,device,return_attns=False):
position_indices = torch.arange(1,self.block_len+1, device=device)
pos_enc = self.pos_emb(position_indices)
enc_output = noisy_enc*pos_enc #<---- addition instead of multiplication?
enc_output = self.position_enc(enc_output)
enc_output = self.dropout(enc_output)
enc_output = self.layer_norm(enc_output)
enc_slf_attn_list = []
for enc_layer in self.layer_stack:
enc_output, enc_slf_attn = enc_layer(enc_output, slf_attn_mask=src_mask)
enc_slf_attn_list += [enc_slf_attn] if return_attns else []
if return_attns:
return enc_output, enc_slf_attn_list
return enc_output # [b_size,block_len,embed_dim]
class XFormerDecoder(nn.Module):
def __init__(self, config, layer_idx=None):
super(XFormerDecoder,self).__init__()
self.embed_dim = config.embed_dim
self.block_len = config.max_len
self.emb_auto = nn.Embedding(config.N+1, config.embed_dim,padding_idx=0)
self.emb_cross = nn.Embedding(config.N+1, config.embed_dim,padding_idx=0)
self.emb_inputs = nn.Embedding(4, config.embed_dim,padding_idx=3)
self.position_enc_auto = PositionalEncoding(self.embed_dim, n_position=self.block_len)
self.position_enc_cross = PositionalEncoding(self.embed_dim, n_position=self.block_len,num=5000)
self.dropout = nn.Dropout(p=config.dropout)
self.dropout_cross = nn.Dropout(p=config.dropout)
self.layer_stack = nn.ModuleList([
DecoderLayer(config.embed_dim, config.embed_dim*4, config.n_head, config.embed_dim//config.n_head, config.embed_dim//config.n_head, dropout=config.dropout)
for _ in range(config.n_layers)])
self.layer_norm = nn.LayerNorm(config.embed_dim, eps=1e-6)
self.layer_norm_cross = nn.LayerNorm(config.embed_dim, eps=1e-6)
def forward(self,noisy_enc,src_mask,trg_seq,trg_mask,device,return_attns=False):
dec_slf_attn_list, dec_enc_attn_list = [], []
position_indices = torch.arange(1,self.block_len+1, device=device)
emb_self = self.emb_auto(position_indices)
emb_cross = self.emb_cross(position_indices)
enc_output = noisy_enc*emb_cross #<---- addition instead of multiplication?
dec_output = self.emb_inputs(trg_seq)
enc_output = self.position_enc_cross(enc_output)
dec_output = self.position_enc_auto(dec_output)
dec_output = self.dropout(dec_output)
dec_output = self.layer_norm(dec_output)
enc_output = self.dropout_cross(enc_output)
enc_output = self.layer_norm_cross(enc_output)
cross_attend = [False for _ in self.layer_stack]
cross_attend[0] = True
for dec_layer in self.layer_stack:
dec_output, dec_slf_attn, dec_enc_attn = dec_layer(
dec_output, enc_output, slf_attn_mask=trg_mask, dec_enc_attn_mask=src_mask)
dec_slf_attn_list += [dec_slf_attn] if return_attns else []
dec_enc_attn_list += [dec_enc_attn] if return_attns else []
if return_attns:
return dec_output, dec_slf_attn_list
return dec_output # [b_size,block_len,embed_dim]
class XFormerGPT(nn.Module):
def __init__(self, config, layer_idx=None):
super(XFormerGPT,self).__init__()
self.embed_dim = config.embed_dim
self.block_len = config.max_len
self.position_enc_auto = PositionalEncoding(self.embed_dim, n_position=self.block_len)
self.dropout = nn.Dropout(p=config.dropout)
#self.pos_emb = nn.Embedding(config.N, config.embed_dim)
#self.dropout_cross = nn.Dropout(p=config.dropout)
self.layer_stack = nn.ModuleList([
EncoderLayer(config.embed_dim, config.embed_dim*4, config.n_head, config.embed_dim//config.n_head, config.embed_dim//config.n_head, dropout=config.dropout)
for _ in range(config.n_layers)])
self.layer_norm = nn.LayerNorm(config.embed_dim, eps=1e-6)
self.layer_norm_cross = nn.LayerNorm(config.embed_dim, eps=1e-6)
def forward(self,trg_seq,trg_mask,device,return_attns=False,return_layer=None):
#position_indices = torch.arange(1,self.block_len+1, device=device)
#pos_enc = self.pos_emb(position_indices)
dec_slf_attn_list, dec_enc_attn_list = [], []
dec_output = self.position_enc_auto(trg_seq)
dec_output = self.dropout(dec_output)
#dec_output = self.layer_norm(dec_output)
layer=1
intermediate_layer_out = None
for dec_layer in self.layer_stack:
dec_output, dec_slf_attn = dec_layer(
dec_output, slf_attn_mask=trg_mask)
dec_slf_attn_list += [dec_slf_attn] if return_attns else []
if return_layer is not None:
if layer == return_layer:
intermediate_layer_out = dec_output
layer += 1
if return_attns:
return dec_output, dec_slf_attn_list
if return_layer is not None:
return dec_output, intermediate_layer_out
return dec_output # [b_size,block_len,embed_dim]
class XFormerEndToEndGPT(nn.Module):
def __init__(self,config):
super(XFormerEndToEndGPT,self).__init__()
self.embed_dim = config.embed_dim
self.block_len = config.max_len
self.trg_pad_idx = 2
MODEL = config.model
self.start_embed_layer = nn.Sequential(
nn.Linear(config.N,self.embed_dim),
nn.GELU(),
nn.Linear(self.embed_dim,self.embed_dim),
nn.GELU(),
nn.Linear(self.embed_dim,self.embed_dim),
)
self.learnt_pos = True
if not self.learnt_pos:
self.emb_inputs = nn.Embedding(2, self.embed_dim)
#self.emb_inputs = nn.Embedding(4, self.embed_dim,padding_idx=3)
else:
self.pos_emb = nn.Embedding(self.block_len, config.embed_dim)
self.layer_norm_inp = nn.LayerNorm(self.embed_dim, eps=1e-6)
self.layer_norm_out = nn.LayerNorm(self.embed_dim, eps=1e-6)
self.Decoder = XFormerGPT(config)
self.Lin_Decoder = nn.Linear(config.embed_dim,1)
def forward(self,noisy_enc,mask,trg_seq,device,return_layer = None):
src_mask = mask
trg_seq = trg_seq[:,:-1]
if not self.learnt_pos:
trg_seq = torch.cat((torch.ones((trg_seq.size(0),1),device=device).long(),(trg_seq==-1).long()),-1) # shift inputs forward by one token
trg_mask = get_pad_mask(trg_seq, self.trg_pad_idx) & get_subsequent_mask(trg_seq) # batch_size x max_len x max_len
trg_seq = self.emb_inputs(trg_seq)
else:
trg_seq = torch.cat((torch.ones((trg_seq.size(0),1),device=device),trg_seq),-1) # shift inputs forward by one token
trg_mask = get_pad_mask(trg_seq, self.trg_pad_idx) & get_subsequent_mask(trg_seq) # batch_size x max_len x max_len
trg_seq = torch.ones(self.embed_dim,device=device)*trg_seq.unsqueeze(-1)
position_indices = torch.arange(self.block_len, device=device)
pos_enc = self.pos_emb(position_indices)
trg_seq = trg_seq*pos_enc
start_emb = self.start_embed_layer(noisy_enc)
trg_seq[:,0] = start_emb
if return_layer is not None:
output,intermediate_layer_out = self.Decoder(trg_seq,trg_mask,device,return_layer=return_layer)
else:
output = self.Decoder(trg_seq,trg_mask,device)
logits = self.Lin_Decoder(output)
decoded_msg_bits = logits.sign()
output = torch.sigmoid(logits)
output = torch.cat((1-output,output),-1)
out_mask = mask
if return_layer is not None:
return output,decoded_msg_bits,out_mask,logits,intermediate_layer_out
return output,decoded_msg_bits,out_mask,logits # [b_size,block_len,2]
def decode(self,noisy_enc,info_positions,mask,device):
start_emb = self.start_embed_layer(noisy_enc)
inp_seq = torch.ones((noisy_enc.size(0),self.block_len,self.embed_dim),device=device)
inp_seq[:,0] = start_emb
inp_mask = mask.unsqueeze(1) & get_subsequent_mask(noisy_enc)
output_bits = torch.ones((noisy_enc.size(0),self.block_len),device=device)
for i in range(noisy_enc.size(1)):
if i in info_positions:
mask_i = inp_mask[:,i,:].unsqueeze(1)
output = self.Decoder(inp_seq,mask_i,device)
output = self.Lin_Decoder(output)
next_bit = output[:,i].sign()
else:
next_bit = torch.ones((noisy_enc.size(0),1),device=device)
output_bits[:,i] = next_bit[:,0]
#print(next_bit)
if i < noisy_enc.size(1)-1:
if not self.learnt_pos:
embed_next_bit = self.emb_inputs((next_bit==1).long())
inp_seq[:,i+1] = embed_next_bit[:,0]
else:
embed_next_bit = next_bit*self.pos_emb(torch.tensor(i+1,device=device)).unsqueeze(0)
inp_seq[:,i+1] = embed_next_bit
out_mask = mask
return output_bits,out_mask
class StartEmbedder(nn.Module):
def __init__(self,inp_dim,hidden_dim,num_layers):
super(StartEmbedder,self).__init__()
self.inp_dim = inp_dim
self.hidden_dim = hidden_dim
self.layers = nn.ModuleList([nn.Linear(self.inp_dim,self.hidden_dim)]+[nn.Linear(hidden_dim,hidden_dim) for i in range(num_layers-1)])
def forward(self,x):
out = self.layers[0](x)
res = out
out = F.gelu(out)
for layer in self.layers[1:-1]:
out = layer(out)
out = F.gelu(out)
out = self.layers[-1](out)
out = out + res
return out
class rnnAttn(nn.Module):
def __init__(self, args):
super(rnnAttn, self).__init__()
#self.vocab_size = params['vocab_size']
self.d_emb = 1#args.embed_dim#params['d_emb']
self.d_hid = args.embed_dim#params['d_hid']
self.block_len = args.N
self.n_layer = 2
self.btz = args.batch_size
self.feature1 = multiplyFeature(args.mat)
#self.encoder = nn.Embedding(self.vocab_size, self.d_emb)
self.attn = Attention(self.d_hid)
self.rnn = nn.GRU(self.d_emb, self.d_hid, self.n_layer, batch_first=True)
self.startEmbedder1 = StartEmbedder(args.N,self.d_hid,3)
self.startEmbedder2 = StartEmbedder(args.N,self.d_hid,3)
# the combined_W maps the combined hidden states and context vectors to d_hid
self.combined_W = nn.Linear(self.d_hid * 3, self.d_hid)
self.decoder = nn.Sequential(
nn.Linear(self.d_hid,self.d_hid),
nn.GELU(),
nn.Linear(self.d_hid,self.d_hid),
nn.GELU(),
nn.Linear(self.d_hid,1),
)
def forward(self,noisy_enc,mask,trg_seq,device,return_layer = None, return_attn_weights=False):
"""
IMPLEMENT ME!
Copy your implementation of RNNLM, make sure it passes the RNNLM check
In addition to that, you need to add the following 3 things
1. pass rnn output to attention module, get context vectors and attention weights
2. concatenate the context vec and rnn output, pass the combined
vector to the layer dealing with the combined vectors (self.combined_W)
3. if return_attn_weights, instead of return the [N, L, V]
matrix, return the attention weight matrix
of dimension [N, L, L] which returned from the forrward function of Attnetion module
"""
batch_size, seq_len= noisy_enc.shape
#multFeat = self.feature1(noisy_enc,device)
trg_seq = trg_seq[:,:-1]
trg_seq = torch.cat((torch.ones((trg_seq.size(0),1),device=device).long(),trg_seq),-1)
start_hidden = self.startEmbedder1(noisy_enc)
#dumb_decode = self.startEmbedder2(multFeat)
hidden = torch.cat((start_hidden.unsqueeze(1),start_hidden.unsqueeze(1)),1)
hidden = torch.transpose(hidden,0,1)
hidden = hidden.contiguous()
start_hidden = (torch.ones((batch_size,seq_len,1),device=device)*start_hidden.unsqueeze(1))
#dumb_decode = (torch.ones((batch_size,seq_len,1),device=device)*dumb_decode.unsqueeze(1))
#init=torch.zeros(self.n_layer, batch_size, self.d_hid).to(device)
#wordvecs = self.encoder(batch)
#print(hidden.size())
outs,last_hidden = self.rnn(trg_seq.unsqueeze(-1),hidden)
context_vec,attn_weights = self.attn(outs)
cat_vec = torch.cat((context_vec,outs,start_hidden),dim = -1)
dec = self.combined_W(cat_vec)
logits = self.decoder(torch.tanh(dec))
decoded_msg_bits = logits.sign()
output = torch.sigmoid(logits)
output = torch.cat((1-output,output),-1)
out_mask = mask
return output,decoded_msg_bits,out_mask,logits
def decode(self,noisy_enc,info_positions,mask,device):
batch_size, seq_len= noisy_enc.shape
inp_seq = torch.ones((noisy_enc.size(0),self.block_len),device=device)
inp_seq[:,0] = 1
output_bits = torch.ones((noisy_enc.size(0),self.block_len),device=device)
#multFeat = self.feature1(noisy_enc,device)
start_hidden = self.startEmbedder1(noisy_enc)
#dumb_decode = self.startEmbedder2(multFeat)
hidden = torch.cat((start_hidden.unsqueeze(1),start_hidden.unsqueeze(1)),1)
hidden = torch.transpose(hidden,0,1)
hidden = hidden.contiguous()
outs_arr = torch.ones((noisy_enc.size(0),self.block_len,self.d_hid),device=device)
start_hidden = (torch.ones((batch_size,seq_len,1),device=device)*start_hidden.unsqueeze(1))
#dumb_decode = (torch.ones((batch_size,seq_len,1),device=device)*dumb_decode.unsqueeze(1))
for i in range(noisy_enc.size(1)):
if i in info_positions:
outs,last_hidden = self.rnn(inp_seq[:,i].unsqueeze(-1).unsqueeze(-1),hidden)
outs_arr[:,i,:] = outs.squeeze()
context_vec,_ = self.attn(outs_arr)
cat_vec = torch.cat((context_vec,outs_arr,start_hidden),dim = -1)
dec = self.combined_W(cat_vec)
logits = self.decoder(torch.tanh(dec))
hidden = last_hidden
next_bit = logits[:,i].sign().squeeze()
else:
outs,last_hidden = self.rnn(inp_seq[:,i].unsqueeze(-1).unsqueeze(-1),hidden)
outs_arr[:,i,:] = outs.squeeze()
hidden = last_hidden
next_bit = torch.ones((noisy_enc.size(0)),device=device)
#print(output_bits[:,i].size())
#print(next_bit.size())
output_bits[:,i] = next_bit
#print(next_bit)
if i < noisy_enc.size(1)-1:
inp_seq[:,i+1] = next_bit
out_mask = mask
return output_bits,out_mask
class Attention(nn.Module):
def __init__(self, d_hidden):
super(Attention, self).__init__()
self.linear_w1 = nn.Linear(d_hidden, d_hidden)
self.linear_w2 = nn.Linear(d_hidden, 1)
def forward(self, x):
"""
IMPLEMENT ME!
For each time step t
1. Obtain attention scores for step 0 to (t-1)
This should be a dot product between current hidden state (x[:,t:t+1,:])
and all previous states x[:, :t, :]. While t=0, since there is not
previous context, the context vector and attention weights should be of zeros.
You might find torch.bmm useful for computing over the whole batch.
2. Turn the scores you get for 0 to (t-1) steps to a distribution.
You might find F.softmax to be helpful.
3. Obtain the sum of hidden states weighted by the attention distribution
Concat the context vector you get in step 3. to a matrix.
Also remember to store the attention weights, the attention matrix
for each training instance should be a lower triangular matrix. Specifically,
each row, element 0 to t-1 should sum to 1, the rest should be padded with 0.
e.g.
[ [0.0000, 0.0000, 0.0000, 0.0000],
[1.0000, 0.0000, 0.0000, 0.0000],
[0.4246, 0.5754, 0.0000, 0.0000],
[0.2798, 0.3792, 0.3409, 0.0000] ]
Return the context vector matrix and the attention weight matrix
"""
batch_seq_len = x.shape[1]
modif_hidden = self.linear_w1(x)
attn_logits = torch.bmm(x,modif_hidden.transpose(1,2))
mask = torch.triu(-10000000000000000.0*torch.ones((batch_seq_len,batch_seq_len),device=x.device))
attn_weights = nn.functional.softmax(attn_logits + mask,-1)
mult_mask = torch.ones(attn_weights.shape,device=x.device)
mult_mask[:,0,:]=0
attn_weights = attn_weights*mult_mask
context_vecs = torch.bmm(attn_weights,x)
return context_vecs, attn_weights
class XFormerEndToEndDecoder(nn.Module):
def __init__(self,config):
super(XFormerEndToEndDecoder,self).__init__()
self.embed_dim = config.embed_dim
self.block_len = config.max_len
self.trg_pad_idx = 3
self.start_idx = 2
self.Decoder = XFormerDecoder(config)
self.Lin_Decoder = nn.Linear(config.embed_dim,1)
def forward(self,noisy_enc,mask,trg_seq,device):
src_mask = mask
trg_seq = trg_seq[:,:-1]
trg_seq = torch.cat((2*torch.ones((trg_seq.size(0),1),device=device).long(),(trg_seq==1).long()),-1)
trg_mask = get_pad_mask(trg_seq, self.trg_pad_idx) & get_subsequent_mask(trg_seq)
batch_size = trg_mask.size(0)
max_len = trg_mask.size(1)
trg_mask = trg_mask.view(batch_size*max_len,max_len)
trg_maskh = torch.cat((trg_mask[:,1:],torch.zeros((trg_mask.size(0),1),device=device)),-1).float()
trg_seq = (trg_seq*torch.ones((max_len,batch_size,max_len),device=device)).long().permute((1,0,2)).reshape(batch_size*max_len,max_len)
noisy_enc = (noisy_enc*torch.ones((max_len,batch_size,max_len),device=device)).permute((1,0,2)).reshape(batch_size*max_len,max_len)
src_mask = (src_mask*torch.ones((max_len,batch_size,max_len),device=device)).permute((1,0,2)).reshape(batch_size*max_len,max_len)
#noisy_enc : [b_size, block_len]
output = torch.ones(self.embed_dim,device=device)*noisy_enc.unsqueeze(-1)
#print(trg_mask.size())
#noisy_enc : [b_size,block_len,embed_dim]
output = self.Decoder(output,src_mask,trg_seq,trg_mask,device)
logits = self.Lin_Decoder(output)
decoded_msg_bits = logits.sign()
output = torch.sigmoid(logits)
output = torch.cat((1-output,output),-1)
out_mask = trg_mask.float() - trg_maskh
return output,decoded_msg_bits,out_mask,logits # [b_size,block_len,2]
def decode(self,noisy_enc,info_positions, mask,device):
enc_input = torch.ones(self.embed_dim,device=device)*noisy_enc.unsqueeze(-1)
inp_seq = torch.ones((noisy_enc.size(0),self.block_len),device=device).long()
inp_seq[:,0] = 2
inp_mask = mask.unsqueeze(1) & get_subsequent_mask(noisy_enc)
output_bits = torch.ones((noisy_enc.size(0),self.block_len),device=device)
for i in range(noisy_enc.size(1)):
if i in info_positions:
mask_i = inp_mask[:,i,:]
output = self.Decoder(enc_input,mask,inp_seq,mask_i,device)
output = self.Lin_Decoder(output)
next_bit = output[:,i].sign()
else:
next_bit = torch.ones((noisy_enc.size(0),1),device=device)
output_bits[:,i] = next_bit[:,0]
embed_next_bit = ((next_bit==1).long())
if i < noisy_enc.size(1)-1:
inp_seq[:,i+1] = embed_next_bit[:,0]
out_mask = mask
return output_bits,out_mask
class XFormerEndToEndEncoder(nn.Module):
def __init__(self,config):
super(XFormerEndToEndEncoder,self).__init__()
self.embed_dim = config.embed_dim
self.block_len = config.max_len
self.Encoder = XFormerEncoder(config)
MODEL = config.model
self.Lin_Decoder = nn.Linear(config.embed_dim,1)
def forward(self,noisy_enc,mask,trg_seq,device):
#noisy_enc : [b_size, block_len]
output = torch.ones(self.embed_dim,device=device)*noisy_enc.unsqueeze(-1)
#noisy_enc : [b_size,block_len,embed_dim]
output = self.Encoder(output,mask,device)
logits = self.Lin_Decoder(output)
decoded_msg_bits = logits.sign()
output = torch.sigmoid(logits)
output = torch.cat((1-output,output),-1)
out_mask = mask
return output,decoded_msg_bits,out_mask,logits # [b_size,block_len,2]
def decode(self,noisy_enc,info_positions,mask,device,trg_seq=None):
_,decoded_msg_bits,out_mask,_ = self.forward(noisy_enc,mask,trg_seq,device)
#decoded_msg_bits = (decoded_msg_bits==1).long()
return decoded_msg_bits,out_mask
class convNet(nn.Module):
def __init__(self,config):
super(convNet,self).__init__()
self.hidden_dim = config.embed_dim
self.input_len = config.max_len
self.output_len = config.N
bias = not config.dont_use_bias
self.kernel = 7
self.padding = int((self.kernel-1)/2)
self.layers1 = nn.Sequential(
nn.Conv1d(1,int(self.hidden_dim/2),self.kernel,padding=self.padding,bias=bias),
nn.GELU(),
nn.Conv1d(int(self.hidden_dim/2),int(self.hidden_dim/2),self.kernel,padding=2*self.padding,dilation=2,bias=bias),
nn.GELU(),
)
self.layers2 = nn.Sequential(
nn.Conv1d(int(self.hidden_dim/2),int(self.hidden_dim/2),self.kernel,padding=4*self.padding,dilation=4,bias=bias),
nn.GELU(),
nn.Conv1d(int(self.hidden_dim/2),int(self.hidden_dim/2),self.kernel,padding=self.padding,bias=bias),
nn.GELU(),
)
self.layers3 = nn.Sequential(
nn.Conv1d(int(self.hidden_dim/2),int(self.hidden_dim/2),self.kernel,padding=2*self.padding,dilation=2,bias=bias),
nn.GELU(),
nn.Conv1d(int(self.hidden_dim/2),int(self.hidden_dim/2),self.kernel,padding=4*self.padding,dilation=4,bias=bias),
nn.GELU(),
)
self.layers4 = nn.Sequential(
nn.Conv1d(int(self.hidden_dim/2),int(self.hidden_dim/2),self.kernel,padding=self.padding,bias=bias),
nn.GELU(),
nn.Conv1d(int(self.hidden_dim/2),int(self.hidden_dim/2),self.kernel,padding=2*self.padding,dilation=2,bias=bias),
nn.GELU(),
)
self.layers5 = nn.Sequential(
nn.Conv1d(int(self.hidden_dim/2),int(self.hidden_dim),self.kernel,padding=4*self.padding,dilation=4,bias=bias),
nn.GELU(),
nn.Conv1d(self.hidden_dim,self.hidden_dim,self.kernel,padding=self.padding,bias=bias),
nn.GELU(),
)
self.layersFin = nn.Sequential(
nn.Linear(self.hidden_dim*self.output_len , 4*self.output_len),
nn.GELU(),
nn.Linear(4*self.output_len , self.output_len),
nn.GELU(),
nn.Linear(self.output_len , self.output_len)
)
self.layer_norm = nn.LayerNorm(self.output_len, eps=1e-6)
self.dropout = nn.Dropout(config.dropout)
def forward(self,noisy_enc,mask,trg_seq,device):
input1 = noisy_enc.unsqueeze(1)
input2 = self.layers1(input1)
residual2 = input2
input3 = self.layers2(input2) + residual2
residual3 = input3
input4 = self.layers3(input3)+ residual3
residual4 = input4
input5 = self.layers4(input4) + residual4
residual5 = input5
input6 = self.layers5(input5)
output = self.layer_norm(self.dropout(self.layersFin(torch.flatten(input6,start_dim=1))))
#print(output.size())
logits = output.squeeze().unsqueeze(-1)
decoded_msg_bits = logits.sign()
output = torch.sigmoid(logits)
output = torch.cat((1-output,output),-1)
out_mask = mask
return output,decoded_msg_bits,out_mask,logits,input4 # [b_size,block_len,2]
def decode(self,noisy_enc,info_positions,mask,device,trg_seq=None):
_,decoded_msg_bits,out_mask,_,_ = self.forward(noisy_enc,mask,trg_seq,device)
#decoded_msg_bits = (decoded_msg_bits==1).long()
return decoded_msg_bits,out_mask