-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathfair_wd_weight_to_weight.py
645 lines (605 loc) · 50.7 KB
/
fair_wd_weight_to_weight.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
import torch
import math
import time
import struct
import argparse
import numpy as np
from collections import OrderedDict
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-model', required=True, help="trained model prefix, also include dir, e.g. ../data/model-100")
args = parser.parse_args()
model_path = args.model
checkpoint = torch.load(model_path, map_location='cpu')
assert 'args' in checkpoint
assert 'model' in checkpoint
args = checkpoint['args']
model = checkpoint['model']
checkpoint_new = {}
model_new = {}
for key in model.keys():
if not key.startswith('decoder.layers'):
if not key.startswith('encoder.layers'):
if key.startswith('encoder.embed_tokens'):
embed_tokens_weight_en = model['encoder.embed_tokens.weight'].float()
embed_tokens_weight_weight_en = model['encoder.embed_tokens_weight.weight'].float()
embed_tokens_weight_bias_en = model['encoder.embed_tokens_bias.weight'].float()
embed_tokens_weight_en = embed_tokens_weight_weight_en * torch.tanh(embed_tokens_weight_en) \
+ embed_tokens_weight_bias_en
model_new['encoder.embed_tokens.weight'] = embed_tokens_weight_en
elif key.startswith('decoder.embed_tokens'):
if 'decoder.wd_embed_tokens_weight' in model.keys():
wd_embed_tokens_weight_de = model['decoder.wd_embed_tokens_weight'].float()
embed_tokens_weight_de = model['decoder.embed_tokens.weight'].float()
embed_tokens_weight_weight_de = model['decoder.embed_tokens_weight.weight'].float()
embed_tokens_weight_bias_de = model['decoder.embed_tokens_bias.weight'].float()
embed_tokens_weight_de = torch.matmul(embed_tokens_weight_de, wd_embed_tokens_weight_de)
embed_tokens_weight_de = embed_tokens_weight_weight_de * torch.tanh(embed_tokens_weight_de) \
+ embed_tokens_weight_bias_de
model_new['decoder.embed_tokens.weight'] = embed_tokens_weight_de
else:
embed_tokens_weight_de = model['decoder.embed_tokens.weight'].float()
embed_tokens_weight_weight_de = model['decoder.embed_tokens_weight.weight'].float()
embed_tokens_weight_bias_de = model['decoder.embed_tokens_bias.weight'].float()
embed_tokens_weight_de = embed_tokens_weight_weight_de * torch.tanh(embed_tokens_weight_de) \
+ embed_tokens_weight_bias_de
model_new['decoder.embed_tokens.weight'] = embed_tokens_weight_de
elif key.startswith('decoder.embed_out'):
if 'decoder.wd_embed_out_weight' in model.keys():
wd_embed_out_weight_de = model['decoder.wd_embed_out_weight'].float()
embed_out_de = model['decoder.embed_out'].float()
embed_out_de = torch.matmul(embed_out_de, wd_embed_out_weight_de)
model_new['decoder.embed_out'] = embed_out_de
else:
model_new['decoder.embed_out'] = model['decoder.embed_out']
else:
if key.startswith('decoder.wd'):
continue
elif key.startswith('encoder.wd'):
continue
else:
model_new[key] = model[key]
for i in range(args.encoder_layers):
# encoder
selfattn_q_weight_en = model['encoder.layers.' + str(i) + '.self_attn.q_weight'].float()
selfattn_k_weight_en = model['encoder.layers.' + str(i) + '.self_attn.k_weight'].float()
selfattn_v_weight_en = model['encoder.layers.' + str(i) + '.self_attn.v_weight'].float()
selfattn_tanh_weight_q_weight_en = model['encoder.layers.' + str(i) + '.self_attn.tanh_weight_q_weight'].float()
selfattn_tanh_weight_k_weight_en = model['encoder.layers.' + str(i) + '.self_attn.tanh_weight_k_weight'].float()
selfattn_tanh_weight_v_weight_en = model['encoder.layers.' + str(i) + '.self_attn.tanh_weight_v_weight'].float()
selfattn_tanh_bias_q_weight_en = model['encoder.layers.' + str(i) + '.self_attn.tanh_bias_q_weight'].float()
selfattn_tanh_bias_k_weight_en = model['encoder.layers.' + str(i) + '.self_attn.tanh_bias_k_weight'].float()
selfattn_tanh_bias_v_weight_en = model['encoder.layers.' + str(i) + '.self_attn.tanh_bias_v_weight'].float()
selfattn_q_weight_en = selfattn_tanh_weight_q_weight_en * torch.tanh(selfattn_q_weight_en) + selfattn_tanh_bias_q_weight_en
selfattn_k_weight_en = selfattn_tanh_weight_k_weight_en * torch.tanh(selfattn_k_weight_en) + selfattn_tanh_bias_k_weight_en
selfattn_v_weight_en = selfattn_tanh_weight_v_weight_en * torch.tanh(selfattn_v_weight_en) + selfattn_tanh_bias_v_weight_en
selfattn_inproj_weight_en = torch.cat((selfattn_q_weight_en, selfattn_k_weight_en, selfattn_v_weight_en), dim=0)
model_new['encoder.layers.' + str(i) + '.self_attn.in_proj_weight'] = selfattn_inproj_weight_en
selfattn_q_bias_en = model['encoder.layers.' + str(i) + '.self_attn.q_bias'].float()
selfattn_k_bias_en = model['encoder.layers.' + str(i) + '.self_attn.k_bias'].float()
selfattn_v_bias_en = model['encoder.layers.' + str(i) + '.self_attn.v_bias'].float()
selfattn_tanh_weight_q_bias_en = model['encoder.layers.' + str(i) + '.self_attn.tanh_weight_q_bias'].float()
selfattn_tanh_weight_k_bias_en = model['encoder.layers.' + str(i) + '.self_attn.tanh_weight_k_bias'].float()
selfattn_tanh_weight_v_bias_en = model['encoder.layers.' + str(i) + '.self_attn.tanh_weight_v_bias'].float()
selfattn_tanh_bias_q_bias_en = model['encoder.layers.' + str(i) + '.self_attn.tanh_bias_q_bias'].float()
selfattn_tanh_bias_k_bias_en = model['encoder.layers.' + str(i) + '.self_attn.tanh_bias_k_bias'].float()
selfattn_tanh_bias_v_bias_en = model['encoder.layers.' + str(i) + '.self_attn.tanh_bias_v_bias'].float()
selfattn_q_bias_en = selfattn_tanh_weight_q_bias_en * torch.tanh(selfattn_q_bias_en) + selfattn_tanh_bias_q_bias_en
selfattn_k_bias_en = selfattn_tanh_weight_k_bias_en * torch.tanh(selfattn_k_bias_en) + selfattn_tanh_bias_k_bias_en
selfattn_v_bias_en = selfattn_tanh_weight_v_bias_en * torch.tanh(selfattn_v_bias_en) + selfattn_tanh_bias_v_bias_en
selfattn_inproj_bias_en = torch.cat((selfattn_q_bias_en, selfattn_k_bias_en, selfattn_v_bias_en), dim=0)
model_new['encoder.layers.' + str(i) + '.self_attn.in_proj_bias'] = selfattn_inproj_bias_en
selfattn_outproj_weight_en = model['encoder.layers.' + str(i) + '.self_attn.out_proj.weight'].float()
selfattn_outproj_tanh_weight_weight_en = model[
'encoder.layers.' + str(i) + '.self_attn.out_proj.tanh_weight_weight'].float()
selfattn_outproj_tanh_bias_weight_en = model[
'encoder.layers.' + str(i) + '.self_attn.out_proj.tanh_bias_weight'].float()
selfattn_outproj_weight_en = selfattn_outproj_tanh_weight_weight_en * torch.tanh(selfattn_outproj_weight_en) + selfattn_outproj_tanh_bias_weight_en
model_new['encoder.layers.' + str(i) + '.self_attn.out_proj.weight'] = selfattn_outproj_weight_en
selfattn_outproj_bias_en = model['encoder.layers.' + str(i) + '.self_attn.out_proj.bias'].float()
selfattn_outproj_tanh_weight_bias_en = model[
'encoder.layers.' + str(i) + '.self_attn.out_proj.tanh_weight_bias'].float()
selfattn_outproj_tanh_bias_bias_en = model[
'encoder.layers.' + str(i) + '.self_attn.out_proj.tanh_bias_bias'].float()
selfattn_outproj_bias_en = selfattn_outproj_tanh_weight_bias_en * torch.tanh(selfattn_outproj_bias_en) + selfattn_outproj_tanh_bias_bias_en
model_new['encoder.layers.' + str(i) + '.self_attn.out_proj.bias'] = selfattn_outproj_bias_en
fc1_weight_en = model['encoder.layers.' + str(i) + '.fc1.weight'].float()
fc1_tanh_weight_weight_en = model['encoder.layers.' + str(i) + '.fc1.tanh_weight_weight'].float()
fc1_tanh_bias_weight_en = model['encoder.layers.' + str(i) + '.fc1.tanh_bias_weight'].float()
fc1_weight_en = fc1_tanh_weight_weight_en * torch.tanh(fc1_weight_en) + fc1_tanh_bias_weight_en
model_new['encoder.layers.' + str(i) + '.fc1.weight'] = fc1_weight_en
fc1_bias_en = model['encoder.layers.' + str(i) + '.fc1.bias'].float()
fc1_tanh_weight_bias_en = model['encoder.layers.' + str(i) + '.fc1.tanh_weight_bias'].float()
fc1_tanh_bias_bias_en = model['encoder.layers.' + str(i) + '.fc1.tanh_bias_bias'].float()
fc1_bias_en = fc1_tanh_weight_bias_en * torch.tanh(fc1_bias_en) + fc1_tanh_bias_bias_en
model_new['encoder.layers.' + str(i) + '.fc1.bias'] = fc1_bias_en
fc2_weight_en = model['encoder.layers.' + str(i) + '.fc2.weight'].float()
fc2_tanh_weight_weight_en = model['encoder.layers.' + str(i) + '.fc2.tanh_weight_weight'].float()
fc2_tanh_bias_weight_en = model['encoder.layers.' + str(i) + '.fc2.tanh_bias_weight'].float()
fc2_weight_en = fc2_tanh_weight_weight_en * torch.tanh(fc2_weight_en) + fc2_tanh_bias_weight_en
model_new['encoder.layers.' + str(i) + '.fc2.weight'] = fc2_weight_en
fc2_bias_en = model['encoder.layers.' + str(i) + '.fc2.bias'].float()
fc2_tanh_weight_bias_en = model['encoder.layers.' + str(i) + '.fc2.tanh_weight_bias'].float()
fc2_tanh_bias_bias_en = model['encoder.layers.' + str(i) + '.fc2.tanh_bias_bias'].float()
fc2_bias_en = fc2_tanh_weight_bias_en * torch.tanh(fc2_bias_en) + fc2_tanh_bias_bias_en
model_new['encoder.layers.' + str(i) + '.fc2.bias'] = fc2_bias_en
layernorm_0_weight = model['encoder.layers.' + str(i) + '.layer_norms.0.weight'].float()
layernorm_0_tanh_weight_weight = model['encoder.layers.' + str(i) + '.layer_norms.0.tanh_weight_weight'].float()
layernorm_0_tanh_bias_weight = model['encoder.layers.' + str(i) + '.layer_norms.0.tanh_bias_weight'].float()
layernorm_0_weight = layernorm_0_tanh_weight_weight * torch.tanh(layernorm_0_weight) + layernorm_0_tanh_bias_weight
model_new['encoder.layers.' + str(i) + '.layer_norms.0.weight'] = layernorm_0_weight
layernorm_0_bias = model['encoder.layers.' + str(i) + '.layer_norms.0.bias'].float()
layernorm_0_tanh_weight_bias = model['encoder.layers.' + str(i) + '.layer_norms.0.tanh_weight_bias'].float()
layernorm_0_tanh_bias_bias = model['encoder.layers.' + str(i) + '.layer_norms.0.tanh_bias_bias'].float()
layernorm_0_bias = layernorm_0_tanh_weight_bias * torch.tanh(layernorm_0_bias) + layernorm_0_tanh_bias_bias
model_new['encoder.layers.' + str(i) + '.layer_norms.0.bias'] = layernorm_0_bias
layernorm_1_weight = model['encoder.layers.' + str(i) + '.layer_norms.1.weight'].float()
layernorm_1_tanh_weight_weight = model['encoder.layers.' + str(i) + '.layer_norms.1.tanh_weight_weight'].float()
layernorm_1_tanh_bias_weight = model['encoder.layers.' + str(i) + '.layer_norms.1.tanh_bias_weight'].float()
layernorm_1_weight = layernorm_1_tanh_weight_weight * torch.tanh(layernorm_1_weight) + layernorm_1_tanh_bias_weight
model_new['encoder.layers.' + str(i) + '.layer_norms.1.weight'] = layernorm_1_weight
layernorm_1_bias = model['encoder.layers.' + str(i) + '.layer_norms.1.bias'].float()
layernorm_1_tanh_weight_bias = model['encoder.layers.' + str(i) + '.layer_norms.1.tanh_weight_bias'].float()
layernorm_1_tanh_bias_bias = model['encoder.layers.' + str(i) + '.layer_norms.1.tanh_bias_bias'].float()
layernorm_1_bias = layernorm_1_tanh_weight_bias * torch.tanh(layernorm_1_bias) + layernorm_1_tanh_bias_bias
model_new['encoder.layers.' + str(i) + '.layer_norms.1.bias'] = layernorm_1_bias
for i in range(args.decoder_layers):
# decoder
selfattn_q_weight = model['decoder.layers.' + str(i) + '.self_attn.q_weight'].float()
selfattn_k_weight = model['decoder.layers.' + str(i) + '.self_attn.k_weight'].float()
selfattn_v_weight = model['decoder.layers.' + str(i) + '.self_attn.v_weight'].float()
selfattn_wd_q_weight = model['decoder.layers.' + str(i) + '.self_attn.wd_q_weight'].float()
selfattn_wd_k_weight = model['decoder.layers.' + str(i) + '.self_attn.wd_k_weight'].float()
selfattn_wd_v_weight = model['decoder.layers.' + str(i) + '.self_attn.wd_v_weight'].float()
selfattn_out_wd_q_weight = model['decoder.layers.' + str(i) + '.self_attn.out_wd_q_weight'].float()
selfattn_out_wd_k_weight = model['decoder.layers.' + str(i) + '.self_attn.out_wd_k_weight'].float()
selfattn_out_wd_v_weight = model['decoder.layers.' + str(i) + '.self_attn.out_wd_v_weight'].float()
selfattn_in_wd_q_weight = model['decoder.layers.' + str(i) + '.self_attn.in_wd_q_weight'].float()
selfattn_in_wd_k_weight = model['decoder.layers.' + str(i) + '.self_attn.in_wd_k_weight'].float()
selfattn_in_wd_v_weight = model['decoder.layers.' + str(i) + '.self_attn.in_wd_v_weight'].float()
selfattn_ly_wd_q_weight = model['decoder.layers.' + str(i) + '.self_attn.ly_wd_q_weight'].float()
selfattn_ly_wd_k_weight = model['decoder.layers.' + str(i) + '.self_attn.ly_wd_k_weight'].float()
selfattn_ly_wd_v_weight = model['decoder.layers.' + str(i) + '.self_attn.ly_wd_v_weight'].float()
selfattn_tanh_weight_q_weight = model['decoder.layers.' + str(i) + '.self_attn.tanh_weight_q_weight'].float()
selfattn_tanh_weight_k_weight = model['decoder.layers.' + str(i) + '.self_attn.tanh_weight_k_weight'].float()
selfattn_tanh_weight_v_weight = model['decoder.layers.' + str(i) + '.self_attn.tanh_weight_v_weight'].float()
selfattn_tanh_bias_q_weight = model['decoder.layers.' + str(i) + '.self_attn.tanh_bias_q_weight'].float()
selfattn_tanh_bias_k_weight = model['decoder.layers.' + str(i) + '.self_attn.tanh_bias_k_weight'].float()
selfattn_tanh_bias_v_weight = model['decoder.layers.' + str(i) + '.self_attn.tanh_bias_v_weight'].float()
selfattn_q_weight = torch.transpose(selfattn_q_weight, 0, 2)
selfattn_k_weight = torch.transpose(selfattn_k_weight, 0, 2)
selfattn_v_weight = torch.transpose(selfattn_v_weight, 0, 2)
selfattn_q_weight = torch.matmul(selfattn_q_weight, selfattn_out_wd_q_weight)
selfattn_k_weight = torch.matmul(selfattn_k_weight, selfattn_out_wd_k_weight)
selfattn_v_weight = torch.matmul(selfattn_v_weight, selfattn_out_wd_v_weight)
selfattn_q_weight = torch.transpose(selfattn_q_weight, 0, 2)
selfattn_k_weight = torch.transpose(selfattn_k_weight, 0, 2)
selfattn_v_weight = torch.transpose(selfattn_v_weight, 0, 2)
selfattn_q_weight = torch.matmul(selfattn_q_weight, selfattn_ly_wd_q_weight)
selfattn_k_weight = torch.matmul(selfattn_k_weight, selfattn_ly_wd_k_weight)
selfattn_v_weight = torch.matmul(selfattn_v_weight, selfattn_ly_wd_v_weight)
selfattn_q_weight = torch.transpose(selfattn_q_weight, 1, 2)
selfattn_k_weight = torch.transpose(selfattn_k_weight, 1, 2)
selfattn_v_weight = torch.transpose(selfattn_v_weight, 1, 2)
selfattn_q_weight = torch.matmul(selfattn_q_weight, selfattn_in_wd_q_weight)
selfattn_k_weight = torch.matmul(selfattn_k_weight, selfattn_in_wd_k_weight)
selfattn_v_weight = torch.matmul(selfattn_v_weight, selfattn_in_wd_v_weight)
selfattn_q_weight = torch.transpose(selfattn_q_weight, 1, 2)
selfattn_k_weight = torch.transpose(selfattn_k_weight, 1, 2)
selfattn_v_weight = torch.transpose(selfattn_v_weight, 1, 2)
selfattn_q_weight = torch.matmul(selfattn_q_weight, selfattn_wd_q_weight)
selfattn_k_weight = torch.matmul(selfattn_k_weight, selfattn_wd_k_weight)
selfattn_v_weight = torch.matmul(selfattn_v_weight, selfattn_wd_v_weight)
selfattn_q_weight = selfattn_q_weight.squeeze(-1)
selfattn_k_weight = selfattn_k_weight.squeeze(-1)
selfattn_v_weight = selfattn_v_weight.squeeze(-1)
selfattn_q_weight = selfattn_tanh_weight_q_weight * torch.tanh(selfattn_q_weight) + selfattn_tanh_bias_q_weight
selfattn_k_weight = selfattn_tanh_weight_k_weight * torch.tanh(selfattn_k_weight) + selfattn_tanh_bias_k_weight
selfattn_v_weight = selfattn_tanh_weight_v_weight * torch.tanh(selfattn_v_weight) + selfattn_tanh_bias_v_weight
selfattn_inproj_weight = torch.cat((selfattn_q_weight, selfattn_k_weight, selfattn_v_weight), dim=0)
model_new['decoder.layers.' + str(i) + '.self_attn.in_proj_weight'] = selfattn_inproj_weight
selfattn_q_bias = model['decoder.layers.' + str(i) + '.self_attn.q_bias'].float()
selfattn_k_bias = model['decoder.layers.' + str(i) + '.self_attn.k_bias'].float()
selfattn_v_bias = model['decoder.layers.' + str(i) + '.self_attn.v_bias'].float()
selfattn_wd_q_bias = model['decoder.layers.' + str(i) + '.self_attn.wd_q_bias'].float()
selfattn_wd_k_bias = model['decoder.layers.' + str(i) + '.self_attn.wd_k_bias'].float()
selfattn_wd_v_bias = model['decoder.layers.' + str(i) + '.self_attn.wd_v_bias'].float()
selfattn_out_wd_q_bias = model['decoder.layers.' + str(i) + '.self_attn.out_wd_q_bias'].float()
selfattn_out_wd_k_bias = model['decoder.layers.' + str(i) + '.self_attn.out_wd_k_bias'].float()
selfattn_out_wd_v_bias = model['decoder.layers.' + str(i) + '.self_attn.out_wd_v_bias'].float()
selfattn_ly_wd_q_bias = model['decoder.layers.' + str(i) + '.self_attn.ly_wd_q_bias'].float()
selfattn_ly_wd_k_bias = model['decoder.layers.' + str(i) + '.self_attn.ly_wd_k_bias'].float()
selfattn_ly_wd_v_bias = model['decoder.layers.' + str(i) + '.self_attn.ly_wd_v_bias'].float()
selfattn_tanh_weight_q_bias = model['decoder.layers.' + str(i) + '.self_attn.tanh_weight_q_bias'].float()
selfattn_tanh_weight_k_bias = model['decoder.layers.' + str(i) + '.self_attn.tanh_weight_k_bias'].float()
selfattn_tanh_weight_v_bias = model['decoder.layers.' + str(i) + '.self_attn.tanh_weight_v_bias'].float()
selfattn_tanh_bias_q_bias = model['decoder.layers.' + str(i) + '.self_attn.tanh_bias_q_bias'].float()
selfattn_tanh_bias_k_bias = model['decoder.layers.' + str(i) + '.self_attn.tanh_bias_k_bias'].float()
selfattn_tanh_bias_v_bias = model['decoder.layers.' + str(i) + '.self_attn.tanh_bias_v_bias'].float()
selfattn_q_bias = torch.transpose(selfattn_q_bias, 0, 1)
selfattn_k_bias = torch.transpose(selfattn_k_bias, 0, 1)
selfattn_v_bias = torch.transpose(selfattn_v_bias, 0, 1)
selfattn_q_bias = torch.matmul(selfattn_q_bias, selfattn_out_wd_q_bias)
selfattn_k_bias = torch.matmul(selfattn_k_bias, selfattn_out_wd_k_bias)
selfattn_v_bias = torch.matmul(selfattn_v_bias, selfattn_out_wd_v_bias)
selfattn_q_bias = torch.transpose(selfattn_q_bias, 0, 1)
selfattn_k_bias = torch.transpose(selfattn_k_bias, 0, 1)
selfattn_v_bias = torch.transpose(selfattn_v_bias, 0, 1)
selfattn_q_bias = torch.matmul(selfattn_q_bias, selfattn_ly_wd_q_bias)
selfattn_k_bias = torch.matmul(selfattn_k_bias, selfattn_ly_wd_k_bias)
selfattn_v_bias = torch.matmul(selfattn_v_bias, selfattn_ly_wd_v_bias)
selfattn_q_bias = torch.matmul(selfattn_q_bias, selfattn_wd_q_bias)
selfattn_k_bias = torch.matmul(selfattn_k_bias, selfattn_wd_k_bias)
selfattn_v_bias = torch.matmul(selfattn_v_bias, selfattn_wd_v_bias)
selfattn_q_bias = selfattn_q_bias.squeeze(-1)
selfattn_k_bias = selfattn_k_bias.squeeze(-1)
selfattn_v_bias = selfattn_v_bias.squeeze(-1)
selfattn_q_bias = selfattn_tanh_weight_q_bias * torch.tanh(selfattn_q_bias) + selfattn_tanh_bias_q_bias
selfattn_k_bias = selfattn_tanh_weight_k_bias * torch.tanh(selfattn_k_bias) + selfattn_tanh_bias_k_bias
selfattn_v_bias = selfattn_tanh_weight_v_bias * torch.tanh(selfattn_v_bias) + selfattn_tanh_bias_v_bias
selfattn_inproj_bias = torch.cat((selfattn_q_bias, selfattn_k_bias, selfattn_v_bias), dim=0)
model_new['decoder.layers.' + str(i) + '.self_attn.in_proj_bias'] = selfattn_inproj_bias
selfattn_outproj_weight = model['decoder.layers.' + str(i) + '.self_attn.out_proj.weight'].float()
selfattn_outproj_wd_weight = model['decoder.layers.' + str(i) + '.self_attn.out_proj.wd_weight'].float()
selfattn_outproj_out_wd_weight = model['decoder.layers.' + str(i) + '.self_attn.out_proj.out_wd_weight'].float()
selfattn_outproj_in_wd_weight = model['decoder.layers.' + str(i) + '.self_attn.out_proj.in_wd_weight'].float()
selfattn_outproj_ly_wd_weight = model['decoder.layers.' + str(i) + '.self_attn.out_proj.ly_wd_weight'].float()
selfattn_outproj_tanh_weight_weight = model[
'decoder.layers.' + str(i) + '.self_attn.out_proj.tanh_weight_weight'].float()
selfattn_outproj_tanh_bias_weight = model[
'decoder.layers.' + str(i) + '.self_attn.out_proj.tanh_bias_weight'].float()
selfattn_outproj_weight = torch.transpose(selfattn_outproj_weight, 0, 2)
selfattn_outproj_weight = torch.matmul(selfattn_outproj_weight, selfattn_outproj_out_wd_weight)
selfattn_outproj_weight = torch.transpose(selfattn_outproj_weight, 0, 2)
selfattn_outproj_weight = torch.matmul(selfattn_outproj_weight, selfattn_outproj_ly_wd_weight)
selfattn_outproj_weight = torch.transpose(selfattn_outproj_weight, 1, 2)
selfattn_outproj_weight = torch.matmul(selfattn_outproj_weight, selfattn_outproj_in_wd_weight)
selfattn_outproj_weight = torch.transpose(selfattn_outproj_weight, 1, 2)
selfattn_outproj_weight = torch.matmul(selfattn_outproj_weight, selfattn_outproj_wd_weight)
selfattn_outproj_weight = selfattn_outproj_weight.squeeze(-1)
selfattn_outproj_weight = selfattn_outproj_tanh_weight_weight * torch.tanh(
selfattn_outproj_weight) + selfattn_outproj_tanh_bias_weight
model_new['decoder.layers.' + str(i) + '.self_attn.out_proj.weight'] = selfattn_outproj_weight
selfattn_outproj_bias = model['decoder.layers.' + str(i) + '.self_attn.out_proj.bias'].float()
selfattn_outproj_wd_bias = model['decoder.layers.' + str(i) + '.self_attn.out_proj.wd_bias'].float()
selfattn_outproj_out_wd_bias = model['decoder.layers.' + str(i) + '.self_attn.out_proj.out_wd_bias'].float()
selfattn_outproj_ly_wd_bias = model['decoder.layers.' + str(i) + '.self_attn.out_proj.ly_wd_bias'].float()
selfattn_outproj_tanh_weight_bias = model[
'decoder.layers.' + str(i) + '.self_attn.out_proj.tanh_weight_bias'].float()
selfattn_outproj_tanh_bias_bias = model[
'decoder.layers.' + str(i) + '.self_attn.out_proj.tanh_bias_bias'].float()
selfattn_outproj_bias = torch.transpose(selfattn_outproj_bias, 0, 1)
selfattn_outproj_bias = torch.matmul(selfattn_outproj_bias, selfattn_outproj_out_wd_bias)
selfattn_outproj_bias = torch.transpose(selfattn_outproj_bias, 0, 1)
selfattn_outproj_bias = torch.matmul(selfattn_outproj_bias, selfattn_outproj_ly_wd_bias)
selfattn_outproj_bias = torch.matmul(selfattn_outproj_bias, selfattn_outproj_wd_bias)
selfattn_outproj_bias = selfattn_outproj_bias.squeeze(-1)
selfattn_outproj_bias = selfattn_outproj_tanh_weight_bias * torch.tanh(
selfattn_outproj_bias) + selfattn_outproj_tanh_bias_bias
model_new['decoder.layers.' + str(i) + '.self_attn.out_proj.bias'] = selfattn_outproj_bias
selfattn_layernorm_weight = model['decoder.layers.' + str(i) + '.self_attn_layer_norm.weight'].float()
selfattn_layernorm_wd_weight = model['decoder.layers.' + str(i) + '.self_attn_layer_norm.wd_weight'].float()
selfattn_layernorm_out_wd_weight = model[
'decoder.layers.' + str(i) + '.self_attn_layer_norm.out_wd_weight'].float()
selfattn_layernorm_ly_wd_weight = model[
'decoder.layers.' + str(i) + '.self_attn_layer_norm.ly_wd_weight'].float()
selfattn_layernorm_tanh_weight_weight = model[
'decoder.layers.' + str(i) + '.self_attn_layer_norm.tanh_weight_weight'].float()
selfattn_layernorm_tanh_bias_weight = model[
'decoder.layers.' + str(i) + '.self_attn_layer_norm.tanh_bias_weight'].float()
selfattn_layernorm_weight = torch.transpose(selfattn_layernorm_weight, 0, 1)
selfattn_layernorm_weight = torch.matmul(selfattn_layernorm_weight, selfattn_layernorm_out_wd_weight)
selfattn_layernorm_weight = torch.transpose(selfattn_layernorm_weight, 0, 1)
selfattn_layernorm_weight = torch.matmul(selfattn_layernorm_weight, selfattn_layernorm_ly_wd_weight)
selfattn_layernorm_weight = torch.matmul(selfattn_layernorm_weight, selfattn_layernorm_wd_weight)
selfattn_layernorm_weight = selfattn_layernorm_weight.squeeze(-1)
selfattn_layernorm_weight = selfattn_layernorm_tanh_weight_weight * torch.tanh(
selfattn_layernorm_weight) + selfattn_layernorm_tanh_bias_weight
model_new['decoder.layers.' + str(i) + '.self_attn_layer_norm.weight'] = selfattn_layernorm_weight
selfattn_layernorm_bias = model['decoder.layers.' + str(i) + '.self_attn_layer_norm.bias'].float()
selfattn_layernorm_wd_bias = model['decoder.layers.' + str(i) + '.self_attn_layer_norm.wd_bias'].float()
selfattn_layernorm_out_wd_bias = model['decoder.layers.' + str(i) + '.self_attn_layer_norm.out_wd_bias'].float()
selfattn_layernorm_ly_wd_bias = model['decoder.layers.' + str(i) + '.self_attn_layer_norm.ly_wd_bias'].float()
selfattn_layernorm_tanh_weight_bias = model[
'decoder.layers.' + str(i) + '.self_attn_layer_norm.tanh_weight_bias'].float()
selfattn_layernorm_tanh_bias_bias = model[
'decoder.layers.' + str(i) + '.self_attn_layer_norm.tanh_bias_bias'].float()
selfattn_layernorm_bias = torch.transpose(selfattn_layernorm_bias, 0, 1)
selfattn_layernorm_bias = torch.matmul(selfattn_layernorm_bias, selfattn_layernorm_out_wd_bias)
selfattn_layernorm_bias = torch.transpose(selfattn_layernorm_bias, 0, 1)
selfattn_layernorm_bias = torch.matmul(selfattn_layernorm_bias, selfattn_layernorm_ly_wd_bias)
selfattn_layernorm_bias = torch.matmul(selfattn_layernorm_bias, selfattn_layernorm_wd_bias)
selfattn_layernorm_bias = selfattn_layernorm_bias.squeeze(-1)
selfattn_layernorm_bias = selfattn_layernorm_tanh_weight_bias * torch.tanh(
selfattn_layernorm_bias) + selfattn_layernorm_tanh_bias_bias
model_new['decoder.layers.' + str(i) + '.self_attn_layer_norm.bias'] = selfattn_layernorm_bias
encoderattn_layernorm_weight = model['decoder.layers.' + str(i) + '.encoder_attn_layer_norm.weight'].float()
encoderattn_layernorm_wd_weight = model[
'decoder.layers.' + str(i) + '.encoder_attn_layer_norm.wd_weight'].float()
encoderattn_layernorm_out_wd_weight = model[
'decoder.layers.' + str(i) + '.encoder_attn_layer_norm.out_wd_weight'].float()
encoderattn_layernorm_ly_wd_weight = model[
'decoder.layers.' + str(i) + '.encoder_attn_layer_norm.ly_wd_weight'].float()
encoderattn_layernorm_tanh_weight_weight = model[
'decoder.layers.' + str(i) + '.encoder_attn_layer_norm.tanh_weight_weight'].float()
encoderattn_layernorm_tanh_bias_weight = model[
'decoder.layers.' + str(i) + '.encoder_attn_layer_norm.tanh_bias_weight'].float()
encoderattn_layernorm_weight = torch.transpose(encoderattn_layernorm_weight, 0, 1)
encoderattn_layernorm_weight = torch.matmul(encoderattn_layernorm_weight, encoderattn_layernorm_out_wd_weight)
encoderattn_layernorm_weight = torch.transpose(encoderattn_layernorm_weight, 0, 1)
encoderattn_layernorm_weight = torch.matmul(encoderattn_layernorm_weight, encoderattn_layernorm_ly_wd_weight)
encoderattn_layernorm_weight = torch.matmul(encoderattn_layernorm_weight, encoderattn_layernorm_wd_weight)
encoderattn_layernorm_weight = encoderattn_layernorm_weight.squeeze(-1)
encoderattn_layernorm_weight = encoderattn_layernorm_tanh_weight_weight * torch.tanh(
encoderattn_layernorm_weight) + encoderattn_layernorm_tanh_bias_weight
model_new['decoder.layers.' + str(i) + '.encoder_attn_layer_norm.weight'] = encoderattn_layernorm_weight
encoderattn_layernorm_bias = model['decoder.layers.' + str(i) + '.encoder_attn_layer_norm.bias'].float()
encoderattn_layernorm_wd_bias = model['decoder.layers.' + str(i) + '.encoder_attn_layer_norm.wd_bias'].float()
encoderattn_layernorm_out_wd_bias = model[
'decoder.layers.' + str(i) + '.encoder_attn_layer_norm.out_wd_bias'].float()
encoderattn_layernorm_ly_wd_bias = model[
'decoder.layers.' + str(i) + '.encoder_attn_layer_norm.ly_wd_bias'].float()
encoderattn_layernorm_tanh_weight_bias = model[
'decoder.layers.' + str(i) + '.encoder_attn_layer_norm.tanh_weight_bias'].float()
encoderattn_layernorm_tanh_bias_bias = model[
'decoder.layers.' + str(i) + '.encoder_attn_layer_norm.tanh_bias_bias'].float()
encoderattn_layernorm_bias = torch.transpose(encoderattn_layernorm_bias, 0, 1)
encoderattn_layernorm_bias = torch.matmul(encoderattn_layernorm_bias, encoderattn_layernorm_out_wd_bias)
encoderattn_layernorm_bias = torch.transpose(encoderattn_layernorm_bias, 0, 1)
encoderattn_layernorm_bias = torch.matmul(encoderattn_layernorm_bias, encoderattn_layernorm_ly_wd_bias)
encoderattn_layernorm_bias = torch.matmul(encoderattn_layernorm_bias, encoderattn_layernorm_wd_bias)
encoderattn_layernorm_bias = encoderattn_layernorm_bias.squeeze(-1)
encoderattn_layernorm_bias = encoderattn_layernorm_tanh_weight_bias * torch.tanh(
encoderattn_layernorm_bias) + encoderattn_layernorm_tanh_bias_bias
model_new['decoder.layers.' + str(i) + '.encoder_attn_layer_norm.bias'] = encoderattn_layernorm_bias
fc1_weight = model['decoder.layers.' + str(i) + '.fc1.weight'].float()
fc1_wd_weight = model['decoder.layers.' + str(i) + '.fc1.wd_weight'].float()
fc1_out_wd_weight = model['decoder.layers.' + str(i) + '.fc1.out_wd_weight'].float()
fc1_in_wd_weight = model['decoder.layers.' + str(i) + '.fc1.in_wd_weight'].float()
fc1_ly_wd_weight = model['decoder.layers.' + str(i) + '.fc1.ly_wd_weight'].float()
fc1_tanh_weight_weight = model['decoder.layers.' + str(i) + '.fc1.tanh_weight_weight'].float()
fc1_tanh_bias_weight = model['decoder.layers.' + str(i) + '.fc1.tanh_bias_weight'].float()
fc1_weight = torch.transpose(fc1_weight, 0, 2)
fc1_weight = torch.matmul(fc1_weight, fc1_out_wd_weight)
fc1_weight = torch.transpose(fc1_weight, 0, 2)
fc1_weight = torch.matmul(fc1_weight, fc1_ly_wd_weight)
fc1_weight = torch.transpose(fc1_weight, 1, 2)
fc1_weight = torch.matmul(fc1_weight, fc1_in_wd_weight)
fc1_weight = torch.transpose(fc1_weight, 1, 2)
fc1_weight = torch.matmul(fc1_weight, fc1_wd_weight)
fc1_weight = fc1_weight.squeeze(-1)
fc1_weight = fc1_tanh_weight_weight * torch.tanh(fc1_weight) + fc1_tanh_bias_weight
model_new['decoder.layers.' + str(i) + '.fc1.weight'] = fc1_weight
fc1_bias = model['decoder.layers.' + str(i) + '.fc1.bias'].float()
fc1_wd_bias = model['decoder.layers.' + str(i) + '.fc1.wd_bias'].float()
fc1_out_wd_bias = model['decoder.layers.' + str(i) + '.fc1.out_wd_bias'].float()
fc1_ly_wd_bias = model['decoder.layers.' + str(i) + '.fc1.ly_wd_bias'].float()
fc1_tanh_weight_bias = model['decoder.layers.' + str(i) + '.fc1.tanh_weight_bias'].float()
fc1_tanh_bias_bias = model['decoder.layers.' + str(i) + '.fc1.tanh_bias_bias'].float()
fc1_bias = torch.transpose(fc1_bias, 0, 1)
fc1_bias = torch.matmul(fc1_bias, fc1_out_wd_bias)
fc1_bias = torch.transpose(fc1_bias, 0, 1)
fc1_bias = torch.matmul(fc1_bias, fc1_ly_wd_bias)
fc1_bias = torch.matmul(fc1_bias, fc1_wd_bias)
fc1_bias = fc1_bias.squeeze(-1)
fc1_bias = fc1_tanh_weight_bias * torch.tanh(fc1_bias) + fc1_tanh_bias_bias
model_new['decoder.layers.' + str(i) + '.fc1.bias'] = fc1_bias
fc2_weight = model['decoder.layers.' + str(i) + '.fc2.weight'].float()
fc2_wd_weight = model['decoder.layers.' + str(i) + '.fc2.wd_weight'].float()
fc2_out_wd_weight = model['decoder.layers.' + str(i) + '.fc2.out_wd_weight'].float()
fc2_in_wd_weight = model['decoder.layers.' + str(i) + '.fc2.in_wd_weight'].float()
fc2_ly_wd_weight = model['decoder.layers.' + str(i) + '.fc2.ly_wd_weight'].float()
fc2_tanh_weight_weight = model['decoder.layers.' + str(i) + '.fc2.tanh_weight_weight'].float()
fc2_tanh_bias_weight = model['decoder.layers.' + str(i) + '.fc2.tanh_bias_weight'].float()
fc2_weight = torch.transpose(fc2_weight, 0, 2)
fc2_weight = torch.matmul(fc2_weight, fc2_out_wd_weight)
fc2_weight = torch.transpose(fc2_weight, 0, 2)
fc2_weight = torch.matmul(fc2_weight, fc2_ly_wd_weight)
fc2_weight = torch.transpose(fc2_weight, 1, 2)
fc2_weight = torch.matmul(fc2_weight, fc2_in_wd_weight)
fc2_weight = torch.transpose(fc2_weight, 1, 2)
fc2_weight = torch.matmul(fc2_weight, fc2_wd_weight)
fc2_weight = fc2_weight.squeeze(-1)
fc2_weight = fc2_tanh_weight_weight * torch.tanh(fc2_weight) + fc2_tanh_bias_weight
model_new['decoder.layers.' + str(i) + '.fc2.weight'] = fc2_weight
fc2_bias = model['decoder.layers.' + str(i) + '.fc2.bias'].float()
fc2_wd_bias = model['decoder.layers.' + str(i) + '.fc2.wd_bias'].float()
fc2_out_wd_bias = model['decoder.layers.' + str(i) + '.fc2.out_wd_bias'].float()
fc2_ly_wd_bias = model['decoder.layers.' + str(i) + '.fc2.ly_wd_bias'].float()
fc2_tanh_weight_bias = model['decoder.layers.' + str(i) + '.fc2.tanh_weight_bias'].float()
fc2_tanh_bias_bias = model['decoder.layers.' + str(i) + '.fc2.tanh_bias_bias'].float()
fc2_bias = torch.transpose(fc2_bias, 0, 1)
fc2_bias = torch.matmul(fc2_bias, fc2_out_wd_bias)
fc2_bias = torch.transpose(fc2_bias, 0, 1)
fc2_bias = torch.matmul(fc2_bias, fc2_ly_wd_bias)
fc2_bias = torch.matmul(fc2_bias, fc2_wd_bias)
fc2_bias = fc2_bias.squeeze(-1)
fc2_bias = fc2_tanh_weight_bias * torch.tanh(fc2_bias) + fc2_tanh_bias_bias
model_new['decoder.layers.' + str(i) + '.fc2.bias'] = fc2_bias
final_layernorm_weight = model['decoder.layers.' + str(i) + '.final_layer_norm.weight'].float()
final_layernorm_wd_weight = model['decoder.layers.' + str(i) + '.final_layer_norm.wd_weight'].float()
final_layernorm_out_wd_weight = model['decoder.layers.' + str(i) + '.final_layer_norm.out_wd_weight'].float()
final_layernorm_ly_wd_weight = model['decoder.layers.' + str(i) + '.final_layer_norm.ly_wd_weight'].float()
final_layernorm_tanh_weight_weight = model[
'decoder.layers.' + str(i) + '.final_layer_norm.tanh_weight_weight'].float()
final_layernorm_tanh_bias_weight = model[
'decoder.layers.' + str(i) + '.final_layer_norm.tanh_bias_weight'].float()
final_layernorm_weight = torch.transpose(final_layernorm_weight, 0, 1)
final_layernorm_weight = torch.matmul(final_layernorm_weight, final_layernorm_out_wd_weight)
final_layernorm_weight = torch.transpose(final_layernorm_weight, 0, 1)
final_layernorm_weight = torch.matmul(final_layernorm_weight, final_layernorm_ly_wd_weight)
final_layernorm_weight = torch.matmul(final_layernorm_weight, final_layernorm_wd_weight)
final_layernorm_weight = final_layernorm_weight.squeeze(-1)
final_layernorm_weight = final_layernorm_tanh_weight_weight * torch.tanh(
final_layernorm_weight) + final_layernorm_tanh_bias_weight
model_new['decoder.layers.' + str(i) + '.final_layer_norm.weight'] = final_layernorm_weight
final_layernorm_bias = model['decoder.layers.' + str(i) + '.final_layer_norm.bias'].float()
final_layernorm_wd_bias = model['decoder.layers.' + str(i) + '.final_layer_norm.wd_bias'].float()
final_layernorm_out_wd_bias = model['decoder.layers.' + str(i) + '.final_layer_norm.out_wd_bias'].float()
final_layernorm_ly_wd_bias = model['decoder.layers.' + str(i) + '.final_layer_norm.ly_wd_bias'].float()
final_layernorm_tanh_weight_bias = model[
'decoder.layers.' + str(i) + '.final_layer_norm.tanh_weight_bias'].float()
final_layernorm_tanh_bias_bias = model['decoder.layers.' + str(i) + '.final_layer_norm.tanh_bias_bias'].float()
final_layernorm_bias = torch.transpose(final_layernorm_bias, 0, 1)
final_layernorm_bias = torch.matmul(final_layernorm_bias, final_layernorm_out_wd_bias)
final_layernorm_bias = torch.transpose(final_layernorm_bias, 0, 1)
final_layernorm_bias = torch.matmul(final_layernorm_bias, final_layernorm_ly_wd_bias)
final_layernorm_bias = torch.matmul(final_layernorm_bias, final_layernorm_wd_bias)
final_layernorm_bias = final_layernorm_bias.squeeze(-1)
final_layernorm_bias = final_layernorm_tanh_weight_bias * torch.tanh(
final_layernorm_bias) + final_layernorm_tanh_bias_bias
model_new['decoder.layers.' + str(i) + '.final_layer_norm.bias'] = final_layernorm_bias
encoderattn_q_weight = model['decoder.layers.' + str(i) + '.encoder_attn.q_weight'].float()
encoderattn_k_weight = model['decoder.layers.' + str(i) + '.encoder_attn.k_weight'].float()
encoderattn_v_weight = model['decoder.layers.' + str(i) + '.encoder_attn.v_weight'].float()
encoderattn_wd_q_weight = model['decoder.layers.' + str(i) + '.encoder_attn.wd_q_weight'].float()
encoderattn_wd_k_weight = model['decoder.layers.' + str(i) + '.encoder_attn.wd_k_weight'].float()
encoderattn_wd_v_weight = model['decoder.layers.' + str(i) + '.encoder_attn.wd_v_weight'].float()
encoderattn_out_wd_q_weight = model['decoder.layers.' + str(i) + '.encoder_attn.out_wd_q_weight'].float()
encoderattn_out_wd_k_weight = model['decoder.layers.' + str(i) + '.encoder_attn.out_wd_k_weight'].float()
encoderattn_out_wd_v_weight = model['decoder.layers.' + str(i) + '.encoder_attn.out_wd_v_weight'].float()
encoderattn_in_wd_q_weight = model['decoder.layers.' + str(i) + '.encoder_attn.in_wd_q_weight'].float()
encoderattn_in_wd_k_weight = model['decoder.layers.' + str(i) + '.encoder_attn.in_wd_k_weight'].float()
encoderattn_in_wd_v_weight = model['decoder.layers.' + str(i) + '.encoder_attn.in_wd_v_weight'].float()
encoderattn_ly_wd_q_weight = model['decoder.layers.' + str(i) + '.encoder_attn.ly_wd_q_weight'].float()
encoderattn_ly_wd_k_weight = model['decoder.layers.' + str(i) + '.encoder_attn.ly_wd_k_weight'].float()
encoderattn_ly_wd_v_weight = model['decoder.layers.' + str(i) + '.encoder_attn.ly_wd_v_weight'].float()
encoderattn_tanh_weight_q_weight = model[
'decoder.layers.' + str(i) + '.encoder_attn.tanh_weight_q_weight'].float()
encoderattn_tanh_weight_k_weight = model[
'decoder.layers.' + str(i) + '.encoder_attn.tanh_weight_k_weight'].float()
encoderattn_tanh_weight_v_weight = model[
'decoder.layers.' + str(i) + '.encoder_attn.tanh_weight_v_weight'].float()
encoderattn_tanh_bias_q_weight = model['decoder.layers.' + str(i) + '.encoder_attn.tanh_bias_q_weight'].float()
encoderattn_tanh_bias_k_weight = model['decoder.layers.' + str(i) + '.encoder_attn.tanh_bias_k_weight'].float()
encoderattn_tanh_bias_v_weight = model['decoder.layers.' + str(i) + '.encoder_attn.tanh_bias_v_weight'].float()
encoderattn_q_weight = torch.transpose(encoderattn_q_weight, 0, 2)
encoderattn_k_weight = torch.transpose(encoderattn_k_weight, 0, 2)
encoderattn_v_weight = torch.transpose(encoderattn_v_weight, 0, 2)
encoderattn_q_weight = torch.matmul(encoderattn_q_weight, encoderattn_out_wd_q_weight)
encoderattn_k_weight = torch.matmul(encoderattn_k_weight, encoderattn_out_wd_k_weight)
encoderattn_v_weight = torch.matmul(encoderattn_v_weight, encoderattn_out_wd_v_weight)
encoderattn_q_weight = torch.transpose(encoderattn_q_weight, 0, 2)
encoderattn_k_weight = torch.transpose(encoderattn_k_weight, 0, 2)
encoderattn_v_weight = torch.transpose(encoderattn_v_weight, 0, 2)
encoderattn_q_weight = torch.matmul(encoderattn_q_weight, encoderattn_ly_wd_q_weight)
encoderattn_k_weight = torch.matmul(encoderattn_k_weight, encoderattn_ly_wd_k_weight)
encoderattn_v_weight = torch.matmul(encoderattn_v_weight, encoderattn_ly_wd_v_weight)
encoderattn_q_weight = torch.transpose(encoderattn_q_weight, 1, 2)
encoderattn_k_weight = torch.transpose(encoderattn_k_weight, 1, 2)
encoderattn_v_weight = torch.transpose(encoderattn_v_weight, 1, 2)
encoderattn_q_weight = torch.matmul(encoderattn_q_weight, encoderattn_in_wd_q_weight)
encoderattn_k_weight = torch.matmul(encoderattn_k_weight, encoderattn_in_wd_k_weight)
encoderattn_v_weight = torch.matmul(encoderattn_v_weight, encoderattn_in_wd_v_weight)
encoderattn_q_weight = torch.transpose(encoderattn_q_weight, 1, 2)
encoderattn_k_weight = torch.transpose(encoderattn_k_weight, 1, 2)
encoderattn_v_weight = torch.transpose(encoderattn_v_weight, 1, 2)
encoderattn_q_weight = torch.matmul(encoderattn_q_weight, encoderattn_wd_q_weight)
encoderattn_k_weight = torch.matmul(encoderattn_k_weight, encoderattn_wd_k_weight)
encoderattn_v_weight = torch.matmul(encoderattn_v_weight, encoderattn_wd_v_weight)
encoderattn_q_weight = encoderattn_q_weight.squeeze(-1)
encoderattn_k_weight = encoderattn_k_weight.squeeze(-1)
encoderattn_v_weight = encoderattn_v_weight.squeeze(-1)
encoderattn_q_weight = encoderattn_tanh_weight_q_weight * torch.tanh(
encoderattn_q_weight) + encoderattn_tanh_bias_q_weight
encoderattn_k_weight = encoderattn_tanh_weight_k_weight * torch.tanh(
encoderattn_k_weight) + encoderattn_tanh_bias_k_weight
encoderattn_v_weight = encoderattn_tanh_weight_v_weight * torch.tanh(
encoderattn_v_weight) + encoderattn_tanh_bias_v_weight
# encoderattn_inproj_weight = torch.cat((encoderattn_q_weight, encoderattn_k_weight, encoderattn_v_weight), dim=0)
# model_new['decoder.layers.' + str(i) + '.encoder_attn.in_proj_weight'] = encoderattn_inproj_weight
model_new['decoder.layers.' + str(i) + '.encoder_attn.q_weight'] = encoderattn_q_weight
model_new['decoder.layers.' + str(i) + '.encoder_attn.k_weight'] = encoderattn_k_weight
model_new['decoder.layers.' + str(i) + '.encoder_attn.v_weight'] = encoderattn_v_weight
encoderattn_q_bias = model['decoder.layers.' + str(i) + '.encoder_attn.q_bias'].float()
encoderattn_k_bias = model['decoder.layers.' + str(i) + '.encoder_attn.k_bias'].float()
encoderattn_v_bias = model['decoder.layers.' + str(i) + '.encoder_attn.v_bias'].float()
encoderattn_wd_q_bias = model['decoder.layers.' + str(i) + '.encoder_attn.wd_q_bias'].float()
encoderattn_wd_k_bias = model['decoder.layers.' + str(i) + '.encoder_attn.wd_k_bias'].float()
encoderattn_wd_v_bias = model['decoder.layers.' + str(i) + '.encoder_attn.wd_v_bias'].float()
encoderattn_out_wd_q_bias = model['decoder.layers.' + str(i) + '.encoder_attn.out_wd_q_bias'].float()
encoderattn_out_wd_k_bias = model['decoder.layers.' + str(i) + '.encoder_attn.out_wd_k_bias'].float()
encoderattn_out_wd_v_bias = model['decoder.layers.' + str(i) + '.encoder_attn.out_wd_v_bias'].float()
encoderattn_ly_wd_q_bias = model['decoder.layers.' + str(i) + '.encoder_attn.ly_wd_q_bias'].float()
encoderattn_ly_wd_k_bias = model['decoder.layers.' + str(i) + '.encoder_attn.ly_wd_k_bias'].float()
encoderattn_ly_wd_v_bias = model['decoder.layers.' + str(i) + '.encoder_attn.ly_wd_v_bias'].float()
encoderattn_tanh_weight_q_bias = model['decoder.layers.' + str(i) + '.encoder_attn.tanh_weight_q_bias'].float()
encoderattn_tanh_weight_k_bias = model['decoder.layers.' + str(i) + '.encoder_attn.tanh_weight_k_bias'].float()
encoderattn_tanh_weight_v_bias = model['decoder.layers.' + str(i) + '.encoder_attn.tanh_weight_v_bias'].float()
encoderattn_tanh_bias_q_bias = model['decoder.layers.' + str(i) + '.encoder_attn.tanh_bias_q_bias'].float()
encoderattn_tanh_bias_k_bias = model['decoder.layers.' + str(i) + '.encoder_attn.tanh_bias_k_bias'].float()
encoderattn_tanh_bias_v_bias = model['decoder.layers.' + str(i) + '.encoder_attn.tanh_bias_v_bias'].float()
encoderattn_q_bias = torch.transpose(encoderattn_q_bias, 0, 1)
encoderattn_k_bias = torch.transpose(encoderattn_k_bias, 0, 1)
encoderattn_v_bias = torch.transpose(encoderattn_v_bias, 0, 1)
encoderattn_q_bias = torch.matmul(encoderattn_q_bias, encoderattn_out_wd_q_bias)
encoderattn_k_bias = torch.matmul(encoderattn_k_bias, encoderattn_out_wd_k_bias)
encoderattn_v_bias = torch.matmul(encoderattn_v_bias, encoderattn_out_wd_v_bias)
encoderattn_q_bias = torch.transpose(encoderattn_q_bias, 0, 1)
encoderattn_k_bias = torch.transpose(encoderattn_k_bias, 0, 1)
encoderattn_v_bias = torch.transpose(encoderattn_v_bias, 0, 1)
encoderattn_q_bias = torch.matmul(encoderattn_q_bias, encoderattn_ly_wd_q_bias)
encoderattn_k_bias = torch.matmul(encoderattn_k_bias, encoderattn_ly_wd_k_bias)
encoderattn_v_bias = torch.matmul(encoderattn_v_bias, encoderattn_ly_wd_v_bias)
encoderattn_q_bias = torch.matmul(encoderattn_q_bias, encoderattn_wd_q_bias)
encoderattn_k_bias = torch.matmul(encoderattn_k_bias, encoderattn_wd_k_bias)
encoderattn_v_bias = torch.matmul(encoderattn_v_bias, encoderattn_wd_v_bias)
encoderattn_q_bias = encoderattn_q_bias.squeeze(-1)
encoderattn_k_bias = encoderattn_k_bias.squeeze(-1)
encoderattn_v_bias = encoderattn_v_bias.squeeze(-1)
encoderattn_q_bias = encoderattn_tanh_weight_q_bias * torch.tanh(
encoderattn_q_bias) + encoderattn_tanh_bias_q_bias
encoderattn_k_bias = encoderattn_tanh_weight_k_bias * torch.tanh(
encoderattn_k_bias) + encoderattn_tanh_bias_k_bias
encoderattn_v_bias = encoderattn_tanh_weight_v_bias * torch.tanh(
encoderattn_v_bias) + encoderattn_tanh_bias_v_bias
model_new['decoder.layers.' + str(i) + '.encoder_attn.q_bias'] = encoderattn_q_bias
model_new['decoder.layers.' + str(i) + '.encoder_attn.k_bias'] = encoderattn_k_bias
model_new['decoder.layers.' + str(i) + '.encoder_attn.v_bias'] = encoderattn_v_bias
encoderattn_outproj_weight = model['decoder.layers.' + str(i) + '.encoder_attn.out_proj.weight'].float()
encoderattn_outproj_wd_weight = model['decoder.layers.' + str(i) + '.encoder_attn.out_proj.wd_weight'].float()
encoderattn_outproj_out_wd_weight = model[
'decoder.layers.' + str(i) + '.encoder_attn.out_proj.out_wd_weight'].float()
encoderattn_outproj_in_wd_weight = model[
'decoder.layers.' + str(i) + '.encoder_attn.out_proj.in_wd_weight'].float()
encoderattn_outproj_ly_wd_weight = model[
'decoder.layers.' + str(i) + '.encoder_attn.out_proj.ly_wd_weight'].float()
encoderattn_outproj_tanh_weight_weight = model[
'decoder.layers.' + str(i) + '.encoder_attn.out_proj.tanh_weight_weight'].float()
encoderattn_outproj_tanh_bias_weight = model[
'decoder.layers.' + str(i) + '.encoder_attn.out_proj.tanh_bias_weight'].float()
encoderattn_outproj_weight = torch.transpose(encoderattn_outproj_weight, 0, 2)
encoderattn_outproj_weight = torch.matmul(encoderattn_outproj_weight, encoderattn_outproj_out_wd_weight)
encoderattn_outproj_weight = torch.transpose(encoderattn_outproj_weight, 0, 2)
encoderattn_outproj_weight = torch.matmul(encoderattn_outproj_weight, encoderattn_outproj_ly_wd_weight)
encoderattn_outproj_weight = torch.transpose(encoderattn_outproj_weight, 1, 2)
encoderattn_outproj_weight = torch.matmul(encoderattn_outproj_weight, encoderattn_outproj_in_wd_weight)
encoderattn_outproj_weight = torch.transpose(encoderattn_outproj_weight, 1, 2)
encoderattn_outproj_weight = torch.matmul(encoderattn_outproj_weight, encoderattn_outproj_wd_weight)
encoderattn_outproj_weight = encoderattn_outproj_weight.squeeze(-1)
encoderattn_outproj_weight = encoderattn_outproj_tanh_weight_weight * torch.tanh(
encoderattn_outproj_weight) + encoderattn_outproj_tanh_bias_weight
model_new['decoder.layers.' + str(i) + '.encoder_attn.out_proj.weight'] = encoderattn_outproj_weight
encoderattn_outproj_bias = model['decoder.layers.' + str(i) + '.encoder_attn.out_proj.bias'].float()
encoderattn_outproj_wd_bias = model['decoder.layers.' + str(i) + '.encoder_attn.out_proj.wd_bias'].float()
encoderattn_outproj_out_wd_bias = model[
'decoder.layers.' + str(i) + '.encoder_attn.out_proj.out_wd_bias'].float()
encoderattn_outproj_ly_wd_bias = model['decoder.layers.' + str(i) + '.encoder_attn.out_proj.ly_wd_bias'].float()
encoderattn_outproj_tanh_weight_bias = model[
'decoder.layers.' + str(i) + '.encoder_attn.out_proj.tanh_weight_bias'].float()
encoderattn_outproj_tanh_bias_bias = model[
'decoder.layers.' + str(i) + '.encoder_attn.out_proj.tanh_bias_bias'].float()
encoderattn_outproj_bias = torch.transpose(encoderattn_outproj_bias, 0, 1)
encoderattn_outproj_bias = torch.matmul(encoderattn_outproj_bias, encoderattn_outproj_out_wd_bias)
encoderattn_outproj_bias = torch.transpose(encoderattn_outproj_bias, 0, 1)
encoderattn_outproj_bias = torch.matmul(encoderattn_outproj_bias, encoderattn_outproj_ly_wd_bias)
encoderattn_outproj_bias = torch.matmul(encoderattn_outproj_bias, encoderattn_outproj_wd_bias)
encoderattn_outproj_bias = encoderattn_outproj_bias.squeeze(-1)
encoderattn_outproj_bias = encoderattn_outproj_tanh_weight_bias * torch.tanh(
encoderattn_outproj_bias) + encoderattn_outproj_tanh_bias_bias
model_new['decoder.layers.' + str(i) + '.encoder_attn.out_proj.bias'] = encoderattn_outproj_bias
args.arch = args.arch.replace('wd_v52', 'stu')
checkpoint_new['args'] = args
checkpoint_new['model'] = model_new
torch.save(checkpoint_new, 'checkpoint_new.pt')
print("finished!")