internship

本文目的

BLIP-2的思想：对于多模态大模型，同时训练Image Encoder和LLM的计算成本实在是太大。该文提出了一种新的思路，使用Q-former这个模块来对齐Image Encoder和LLM之间的gap。
STA的思想：尽管现有的ANN2SNN转换方法适用于卷积网络，新兴的Transformer模型引入了自注意力和测试时标准化等独特机制，导致当前SNNs难以实现的非因果非线性交互。文章提出了一种无需训练的ANN到SNN的转换方法，该方法通过时空近似（STA）将ANN激活转换为时间尖峰序列，几乎保留了源模型的所有功能。
本文思想：BLIP-2多模态对齐非常好用但是它是是现在ANN上面的，我们现在要使用STA的思想将BLIP-2的Image Encoder和LLM部分有ANN表示转换为SNN表示的

BLIP-2


1
2
3
4
5
6
7
8
9
10
11
12
13
14
.
├── Qformer.py
├── __init__.py
├── __pycache__
├── blip2.py
├── blip2_image_text_matching.py
├── blip2_opt.py
├── blip2_qformer.py
├── blip2_t5.py
├── blip2_t5_instruct.py
├── blip2_vicuna_instruct.py
├── modeling_llama.py
├── modeling_opt.py
└── modeling_t5.py

由于我只需要改LLM部分，那么涉及到的应该是opt和t5

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
OPTForCausalLM(
  (model): OPTModel(
    (decoder): OPTDecoder(
      (embed_tokens): Embedding(50272, 2560, padding_idx=1)
      (embed_positions): OPTLearnedPositionalEmbedding(2050, 2560)
      (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
      (layers): ModuleList(
        (0): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (1): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (2): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (3): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (4): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (5): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (6): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (7): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (8): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (9): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (10): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (11): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (12): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (13): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (14): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (15): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (16): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (17): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (18): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (19): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (20): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (21): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (22): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (23): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (24): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (25): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (26): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (27): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (28): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (29): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (30): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
        (31): OPTDecoderLayer(
          (self_attn): OPTAttention(
            (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
            (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
          )
          (activation_fn): ReLU()
          (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
          (fc1): Linear(in_features=2560, out_features=10240, bias=True)
          (fc2): Linear(in_features=10240, out_features=2560, bias=True)
          (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        )
      )
    )
  )
  (lm_head): Linear(in_features=2560, out_features=50272, bias=False)
)

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
VisionTransformer(
  (conv1): Conv2d(3, 768, kernel_size=(32, 32), stride=(32, 32), bias=False)
  (ln_pre): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
  (transformer): Transformer(
    (resblocks): Sequential(
      (0): ResidualAttentionBlock(
        (attn): MultiheadAttention(
          (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
        )
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): Sequential(
          (c_fc): Linear(in_features=768, out_features=3072, bias=True)
          (gelu): QuickGELU()
          (c_proj): Linear(in_features=3072, out_features=768, bias=True)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      )
      (1): ResidualAttentionBlock(
        (attn): MultiheadAttention(
          (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
        )
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): Sequential(
          (c_fc): Linear(in_features=768, out_features=3072, bias=True)
          (gelu): QuickGELU()
          (c_proj): Linear(in_features=3072, out_features=768, bias=True)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      )
      (2): ResidualAttentionBlock(
        (attn): MultiheadAttention(
          (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
        )
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): Sequential(
          (c_fc): Linear(in_features=768, out_features=3072, bias=True)
          (gelu): QuickGELU()
          (c_proj): Linear(in_features=3072, out_features=768, bias=True)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      )
      (3): ResidualAttentionBlock(
        (attn): MultiheadAttention(
          (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
        )
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): Sequential(
          (c_fc): Linear(in_features=768, out_features=3072, bias=True)
          (gelu): QuickGELU()
          (c_proj): Linear(in_features=3072, out_features=768, bias=True)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      )
      (4): ResidualAttentionBlock(
        (attn): MultiheadAttention(
          (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
        )
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): Sequential(
          (c_fc): Linear(in_features=768, out_features=3072, bias=True)
          (gelu): QuickGELU()
          (c_proj): Linear(in_features=3072, out_features=768, bias=True)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      )
      (5): ResidualAttentionBlock(
        (attn): MultiheadAttention(
          (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
        )
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): Sequential(
          (c_fc): Linear(in_features=768, out_features=3072, bias=True)
          (gelu): QuickGELU()
          (c_proj): Linear(in_features=3072, out_features=768, bias=True)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      )
      (6): ResidualAttentionBlock(
        (attn): MultiheadAttention(
          (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
        )
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): Sequential(
          (c_fc): Linear(in_features=768, out_features=3072, bias=True)
          (gelu): QuickGELU()
          (c_proj): Linear(in_features=3072, out_features=768, bias=True)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      )
      (7): ResidualAttentionBlock(
        (attn): MultiheadAttention(
          (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
        )
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): Sequential(
          (c_fc): Linear(in_features=768, out_features=3072, bias=True)
          (gelu): QuickGELU()
          (c_proj): Linear(in_features=3072, out_features=768, bias=True)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      )
      (8): ResidualAttentionBlock(
        (attn): MultiheadAttention(
          (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
        )
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): Sequential(
          (c_fc): Linear(in_features=768, out_features=3072, bias=True)
          (gelu): QuickGELU()
          (c_proj): Linear(in_features=3072, out_features=768, bias=True)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      )
      (9): ResidualAttentionBlock(
        (attn): MultiheadAttention(
          (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
        )
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): Sequential(
          (c_fc): Linear(in_features=768, out_features=3072, bias=True)
          (gelu): QuickGELU()
          (c_proj): Linear(in_features=3072, out_features=768, bias=True)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      )
      (10): ResidualAttentionBlock(
        (attn): MultiheadAttention(
          (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
        )
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): Sequential(
          (c_fc): Linear(in_features=768, out_features=3072, bias=True)
          (gelu): QuickGELU()
          (c_proj): Linear(in_features=3072, out_features=768, bias=True)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      )
      (11): ResidualAttentionBlock(
        (attn): MultiheadAttention(
          (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
        )
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): Sequential(
          (c_fc): Linear(in_features=768, out_features=3072, bias=True)
          (gelu): QuickGELU()
          (c_proj): Linear(in_features=3072, out_features=768, bias=True)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      )
    )
  )
  (ln_post): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
)


OPTDecoder:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
(model): OPTModel(
  (decoder): OPTDecoder(
    (embed_tokens): Embedding(50272, 2560, padding_idx=1)
    (embed_positions): OPTLearnedPositionalEmbedding(2050, 2560)
    (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
    (layers): ModuleList(
      (0): OPTDecoderLayer(
        (self_attn): OPTAttention(
          (k_proj): Linear(in_features=2560, out_features=2560, bias=True)
          (v_proj): Linear(in_features=2560, out_features=2560, bias=True)
          (q_proj): Linear(in_features=2560, out_features=2560, bias=True)
          (out_proj): Linear(in_features=2560, out_features=2560, bias=True)
        )
        (activation_fn): ReLU()
        (self_attn_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        (fc1): Linear(in_features=2560, out_features=10240, bias=True)
        (fc2): Linear(in_features=10240, out_features=2560, bias=True)
        (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
      )
  • linear
  • ReLU

vit:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
(transformer): Transformer(
  (resblocks): Sequential(
    (0): ResidualAttentionBlock(
      (attn): MultiheadAttention(
        (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
      )
      (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      (mlp): Sequential(
        (c_fc): Linear(in_features=768, out_features=3072, bias=True)
        (gelu): QuickGELU()
        (c_proj): Linear(in_features=3072, out_features=768, bias=True)
      )
      (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
    )
  • Linear
  • LN
  • GELU

gnn:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
(transformer): Transformer(
  (resblocks): Sequential(
    (0): ResidualAttentionBlock(
      (attn): MultiheadAttention(
        (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)
      )
      (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      (mlp): Sequential(
        (c_fc): Linear(in_features=768, out_features=3072, bias=True)
        (gelu): Distilled_GELU(
          (approximator): Sequential(
            (0): Linear(in_features=1, out_features=64, bias=True)
            (1): ReLU()
            (2): Linear(in_features=64, out_features=1, bias=True)
          )
        )
        (c_proj): Linear(in_features=3072, out_features=768, bias=True)
      )
      (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
    )

snn:

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
(2): ResidualAttentionBlock(
  (attn): SpikeAttention(
    (product): SpikeProduct()
    (spike_x2x): X2X(
      (approximator): Sequential(
        (0): SpikeLinear_ReLU(
          (relu): ReLU()
        )
        (1): StraightThrough()
        (2): SpikeLinear_ReLU(
          (relu): StraightThrough()
        )
      )
    )
    (spike_x2x_pos): X2X_POS(
      (approximator): Sequential(
        (0): SpikeLinear_ReLU(
          (relu): ReLU()
        )
        (1): StraightThrough()
      )
    )
  )
  (ln_1): SpikeLN(
    (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
    (spike_sqrtinv): Distilled_SQRTINV(
      (approximator): Sequential(
        (0): SpikeLinear_ReLU(
          (relu): ReLU()
        )
        (1): StraightThrough()
        (2): SpikeLinear_ReLU(
          (relu): StraightThrough()
        )
      )
    )
    (spike_x2x): X2X(
      (approximator): Sequential(
        (0): SpikeLinear_ReLU(
          (relu): ReLU()
        )
        (1): StraightThrough()
        (2): SpikeLinear_ReLU(
          (relu): StraightThrough()
        )
      )
    )
  )
  (mlp): Sequential(
    (c_fc): SpikeLinear_ReLU(
      (relu): StraightThrough()
    )
    (gelu): Distilled_GELU(
      (approximator): Sequential(
        (0): SpikeLinear_ReLU(
          (relu): ReLU()
        )
        (1): StraightThrough()
        (2): SpikeLinear_ReLU(
          (relu): StraightThrough()
        )
      )
    )
    (c_proj): SpikeLinear_ReLU(
      (relu): StraightThrough()
    )
  )
  (ln_2): SpikeLN(
    (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
    (spike_sqrtinv): Distilled_SQRTINV(
      (approximator): Sequential(
        (0): SpikeLinear_ReLU(
          (relu): ReLU()
        )
        (1): StraightThrough()
        (2): SpikeLinear_ReLU(
          (relu): StraightThrough()
        )
      )
    )
    (spike_x2x): X2X(
      (approximator): Sequential(
        (0): SpikeLinear_ReLU(
          (relu): ReLU()
        )
        (1): StraightThrough()
        (2): SpikeLinear_ReLU(
          (relu): StraightThrough()
        )
      )
    )
  )
)

下面是代码是经过如下操作后的代码：

1
2
3
4
5
6
7
8
mse = False if args.method =='normal' else True
    get_maximum_activation(train_loader, model=snn, momentum=0.9, iters=args.iters, mse=mse, percentile=args.percentile, T=args.T, neuron_wise=args.neuron_wise)

    torch.set_num_threads(10)


    
    snn.set_spike_state(use_spike=True)
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
(2): ResidualAttentionBlock(
          (attn): SpikeAttention(
            (product): SpikeProduct()
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x_pos): X2X_POS(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
              )
            )
          )
          (ln_1): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
          (mlp): Sequential(
            (c_fc): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (gelu): Distilled_GELU(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (c_proj): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
          )
          (ln_2): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
        )

实验部分

BLIP-2 在各种零样本视觉语言任务上的结果概述。与以前最先进的模型相比。 BLIP-2 实现了最高的零样本性能，同时在视觉语言预训练期间需要最少数量的可训练参数。

零样本视觉问答与最先进方法的比较。

零样本 VQA。我们对零样本视觉问答任务进行定量评估。对于OPT模型，我们使用提示“问题：{}答案：”。对于 FlanT5 型号，我们使用提示“问题：{}简答：”。在生成过程中，我们使用波束宽度为 5 的波束搜索。我们还将长度惩罚设置为 -1，这鼓励更短的答案，更好地与人工注释保持一致。

如表 2 所示。BLIP-2 在 VQAv2 (Goyal et al., 2017) 和 GQA (Hudson & Manning, 2019) 数据集上取得了最先进的结果。 尽管可训练参数少了 54 倍，但它在 VQAv2 上的性能比 Flamingo80B 高出 8.7%。在 OK-VQA（Marino 等人，2019）数据集上，BLIP-2 次于 Flamingo80B。我们假设这是因为 OK-VQA 更注重开放世界知识而不是视觉理解，而来自 Flamingo80B 的 70B Chinchilla (Hoffmann et al., 2022) 语言模型比 11B FlanT5XXL 拥有更多的知识。

我们从表 2 中得出了一个有希望的观察结果：更强的图像编码器或更强的 LLM 都会带来更好的性能。这一观察结果得到了几个事实的支持：(1) ViT-g 在 OPT 和 FlanT5 方面均优于 ViT-L。 (2) 在同一LLM系列中，较大的模型优于较小的模型。 (3) FlanT5，一种指令调整的 LLM，在 VQA 上优于无监督训练的 OPT。这一观察结果验证了 BLIP-2 作为一种通用的视觉语言预训练方法，可以有效地收获视觉和自然语言社区的快速进步。

表 3：NoCaps 和 COCO Caption 上最先进的图像字幕方法的比较。所有方法都优化微调过程中的交叉熵损失。 C: CIDEr, S: SPICE, B@4: BLEU@4

我们针对图像字幕任务对 BLIP-2 模型进行了微调，该任务要求模型为图像的视觉内容生成文本描述。我们使用提示“a photo of”作为 LLM 的初始输入，并训练模型生成具有语言建模损失的标题。我们在微调期间保持 LLM 冻结，并与图像编码器一起更新 Q-Former 的参数。我们用 ViT-g 和各种 LLMs 进行实验。详细的超参数可以在附录中找到。我们对 COCO 进行微调，并对 COCO 测试集和零样本转移到 NoCaps（Agrawal 等人，2019）验证集进行评估。

结果如表 3 所示。BLIP-2 实现了最先进的性能，与现有方法相比，NoCap 有了显着改进，展示了对外域图像的强大泛化能力。

表 4：与针对视觉问答进行微调的最先进模型的比较。

给定带注释的 VQA 数据，我们微调 Q-Former 和图像编码器的参数，同时保持 LLM 冻结。我们对开放式答案生成损失进行微调，其中 LLM 接收 Q-Former 的输出和问题作为输入，并被要求生成答案。为了提取与问题更相关的图像特征，我们还针对问题设置了 Q-Former 条件。具体来说，问题标记作为 Q-Former 的输入给出，并通过自注意力层与查询进行交互，这可以引导 Q-Former 的交叉注意力层关注信息更丰富的图像区域。

遵循 BLIP，我们的 VQA 数据包括来自 VQAv2 的训练和验证分割，以及来自 Visual Genome 的训练样本。表 4 显示了 BLIP-2 在开放式生成模型中的最新结果。

表 5：与最先进的图像文本检索方法的比较，在 COCO 上进行微调，并零样本传输到 Flickr30K。

由于图像文本检索不涉及语言生成，因此我们直接对第一阶段预训练模型进行微调，无需LLM。具体来说，我们使用与预训练相同的目标（即 ITC、ITM 和 ITG）在 COCO 上与 Q-Former 一起微调图像编码器。然后，我们在 CO 和 Flickr30K 数据集上评估图像到文本检索和文本到图像检索的模型。在推理过程中，我们遵循 Li 等人的观点。 (2021, 2022) 首先根据图像文本特征相似性选择 k=128 候选者，然后根据成对的 ITM 分数重新排名。我们尝试使用 ViT-L 和 ViT-g 作为图像编码器。详细的超参数可以在附录中找到。

表 6：基于图像的文本生成 (ITG) 损失通过强制查询提取与语言相关的视觉特征来提高图像文本检索性能。

TC 和 ITM 损失对于图像文本检索至关重要，因为它们直接学习图像文本相似性。在表 6 中，我们表明 ITG（基于图像的文本生成）损失也有利于图像文本检索。这一结果支持了我们设计表示学习目标的直觉：ITG 损失强制查询提取与文本最相关的视觉特征，从而改善视觉语言对齐。


transformers导入与modeling_opt.py性能差别

整体来看，其实差不多

transforemrs导入：

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
2024-03-27 16:49:11,260 [INFO] Start training epoch 0, 2 iters per inner epoch.
/cyb/LAVIS/lavis/processors/randaugment.py:40: RuntimeWarning: overflow encountered in scalar negative
  offset = -low * scale
Train: data epoch: [0]  [0/2]  eta: 0:00:11  lr: 0.000001  loss: 4.7803  time: 5.9693  data: 0.0000  max mem: 8321
2024-03-27 16:49:17,694 [INFO] Reducer buckets have been rebuilt in this iteration.
Train: data epoch: [0]  [1/2]  eta: 0:00:03  lr: 0.000001  loss: 5.7962  time: 3.3131  data: 0.0000  max mem: 8714
Train: data epoch: [0] Total time: 0:00:06 (3.3137 s / it)
2024-03-27 16:49:17,888 [INFO] Averaged stats: lr: 0.0000  loss: 5.2883
2024-03-27 16:49:17,891 [INFO] No validation splits found.
2024-03-27 16:49:17,921 [INFO] Saving checkpoint at epoch 0 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327164/checkpoint_0.pth.
2024-03-27 16:49:18,958 [INFO] Start training
2024-03-27 16:49:19,008 [INFO] Start training epoch 1, 2 iters per inner epoch.
Train: data epoch: [1]  [0/2]  eta: 0:00:06  lr: 0.000098  loss: 4.2522  time: 3.2473  data: 0.0000  max mem: 8731
Train: data epoch: [1]  [1/2]  eta: 0:00:01  lr: 0.000098  loss: 6.3714  time: 1.8830  data: 0.0000  max mem: 8736
Train: data epoch: [1] Total time: 0:00:03 (1.8835 s / it)
2024-03-27 16:49:22,776 [INFO] Averaged stats: lr: 0.0001  loss: 5.3118
2024-03-27 16:49:22,778 [INFO] No validation splits found.
2024-03-27 16:49:22,808 [INFO] Saving checkpoint at epoch 1 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327164/checkpoint_1.pth.
2024-03-27 16:49:23,789 [INFO] Start training
2024-03-27 16:49:23,824 [INFO] Start training epoch 2, 2 iters per inner epoch.
Train: data epoch: [2]  [0/2]  eta: 0:00:06  lr: 0.000091  loss: 5.5883  time: 3.3091  data: 0.0000  max mem: 9036
Train: data epoch: [2]  [1/2]  eta: 0:00:01  lr: 0.000091  loss: 6.0959  time: 1.9160  data: 0.0000  max mem: 9533
Train: data epoch: [2] Total time: 0:00:03 (1.9168 s / it)
2024-03-27 16:49:27,658 [INFO] Averaged stats: lr: 0.0001  loss: 5.8421
2024-03-27 16:49:27,661 [INFO] No validation splits found.
2024-03-27 16:49:27,694 [INFO] Saving checkpoint at epoch 2 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327164/checkpoint_2.pth.
2024-03-27 16:49:29,868 [INFO] Start training
2024-03-27 16:49:29,904 [INFO] Start training epoch 3, 2 iters per inner epoch.
/cyb/LAVIS/lavis/processors/randaugment.py:40: RuntimeWarning: overflow encountered in scalar negative
  offset = -low * scale
Train: data epoch: [3]  [0/2]  eta: 0:00:06  lr: 0.000081  loss: 5.6558  time: 3.4074  data: 0.0000  max mem: 9534
Train: data epoch: [3]  [1/2]  eta: 0:00:02  lr: 0.000081  loss: 6.4650  time: 2.0847  data: 0.0000  max mem: 9555
Train: data epoch: [3] Total time: 0:00:04 (2.0856 s / it)
2024-03-27 16:49:34,076 [INFO] Averaged stats: lr: 0.0001  loss: 6.0604
2024-03-27 16:49:34,079 [INFO] No validation splits found.
2024-03-27 16:49:34,111 [INFO] Saving checkpoint at epoch 3 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327164/checkpoint_3.pth.
2024-03-27 16:49:36,305 [INFO] Start training
2024-03-27 16:49:36,340 [INFO] Start training epoch 4, 2 iters per inner epoch.
Train: data epoch: [4]  [0/2]  eta: 0:00:07  lr: 0.000069  loss: 4.9748  time: 3.5716  data: 0.0000  max mem: 9555
Train: data epoch: [4]  [1/2]  eta: 0:00:02  lr: 0.000069  loss: 6.2966  time: 2.1039  data: 0.0000  max mem: 9555
Train: data epoch: [4] Total time: 0:00:04 (2.1049 s / it)
2024-03-27 16:49:40,552 [INFO] Averaged stats: lr: 0.0001  loss: 5.6357
2024-03-27 16:49:40,562 [INFO] No validation splits found.
2024-03-27 16:49:40,647 [INFO] Saving checkpoint at epoch 4 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327164/checkpoint_4.pth.
2024-03-27 16:49:43,149 [INFO] Start training
2024-03-27 16:49:43,184 [INFO] Start training epoch 5, 2 iters per inner epoch.
/cyb/LAVIS/lavis/processors/randaugment.py:40: RuntimeWarning: overflow encountered in scalar negative
  offset = -low * scale
Train: data epoch: [5]  [0/2]  eta: 0:00:06  lr: 0.000055  loss: 4.6983  time: 3.4429  data: 0.0000  max mem: 9555
Train: data epoch: [5]  [1/2]  eta: 0:00:02  lr: 0.000055  loss: 7.9794  time: 2.1261  data: 0.0000  max mem: 9555
Train: data epoch: [5] Total time: 0:00:04 (2.1273 s / it)
2024-03-27 16:49:47,440 [INFO] Averaged stats: lr: 0.0001  loss: 6.3388
2024-03-27 16:49:47,443 [INFO] No validation splits found.
2024-03-27 16:49:47,475 [INFO] Saving checkpoint at epoch 5 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327164/checkpoint_5.pth.
2024-03-27 16:49:49,912 [INFO] Start training
2024-03-27 16:49:49,948 [INFO] Start training epoch 6, 2 iters per inner epoch.
Train: data epoch: [6]  [0/2]  eta: 0:00:07  lr: 0.000041  loss: 4.7854  time: 3.5252  data: 0.0000  max mem: 9555
Train: data epoch: [6]  [1/2]  eta: 0:00:02  lr: 0.000041  loss: 5.9835  time: 2.1675  data: 0.0000  max mem: 9555
Train: data epoch: [6] Total time: 0:00:04 (2.1683 s / it)
2024-03-27 16:49:54,286 [INFO] Averaged stats: lr: 0.0000  loss: 5.3845
2024-03-27 16:49:54,289 [INFO] No validation splits found.
2024-03-27 16:49:54,332 [INFO] Saving checkpoint at epoch 6 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327164/checkpoint_6.pth.
2024-03-27 16:49:56,880 [INFO] Start training
2024-03-27 16:49:56,914 [INFO] Start training epoch 7, 2 iters per inner epoch.
Train: data epoch: [7]  [0/2]  eta: 0:00:06  lr: 0.000029  loss: 4.6165  time: 3.4250  data: 0.0000  max mem: 9555
Train: data epoch: [7]  [1/2]  eta: 0:00:01  lr: 0.000029  loss: 6.2248  time: 1.9936  data: 0.0000  max mem: 9555
Train: data epoch: [7] Total time: 0:00:03 (1.9946 s / it)
2024-03-27 16:50:00,904 [INFO] Averaged stats: lr: 0.0000  loss: 5.4207
2024-03-27 16:50:00,908 [INFO] No validation splits found.
2024-03-27 16:50:00,967 [INFO] Saving checkpoint at epoch 7 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327164/checkpoint_7.pth.
2024-03-27 16:50:03,337 [INFO] Start training
2024-03-27 16:50:03,373 [INFO] Start training epoch 8, 2 iters per inner epoch.
/cyb/LAVIS/lavis/processors/randaugment.py:40: RuntimeWarning: overflow encountered in scalar negative
  offset = -low * scale
Train: data epoch: [8]  [0/2]  eta: 0:00:06  lr: 0.000019  loss: 5.0990  time: 3.4835  data: 0.0000  max mem: 9555
Train: data epoch: [8]  [1/2]  eta: 0:00:02  lr: 0.000019  loss: 6.2343  time: 2.1475  data: 0.0000  max mem: 9555
Train: data epoch: [8] Total time: 0:00:04 (2.1485 s / it)
2024-03-27 16:50:07,670 [INFO] Averaged stats: lr: 0.0000  loss: 5.6667
2024-03-27 16:50:07,674 [INFO] No validation splits found.
2024-03-27 16:50:07,705 [INFO] Saving checkpoint at epoch 8 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327164/checkpoint_8.pth.
2024-03-27 16:50:10,070 [INFO] Start training
2024-03-27 16:50:10,104 [INFO] Start training epoch 9, 2 iters per inner epoch.
Train: data epoch: [9]  [0/2]  eta: 0:00:06  lr: 0.000012  loss: 5.9895  time: 3.4054  data: 0.0000  max mem: 9556
Train: data epoch: [9]  [1/2]  eta: 0:00:02  lr: 0.000012  loss: 5.6200  time: 2.0422  data: 0.0000  max mem: 9556
Train: data epoch: [9] Total time: 0:00:04 (2.0434 s / it)
2024-03-27 16:50:14,192 [INFO] Averaged stats: lr: 0.0000  loss: 5.8048
2024-03-27 16:50:14,197 [INFO] No validation splits found.
2024-03-27 16:50:14,233 [INFO] Saving checkpoint at epoch 9 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327164/checkpoint_9.pth.
2024-03-27 16:50:16,610 [INFO] No validation splits found.
2024-03-27 16:50:16,610 [INFO] Training time 0:01:10

modeling_opt.py文件：

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
2024-03-27 16:58:03,450 [INFO] Start training epoch 0, 2 iters per inner epoch.
/cyb/LAVIS/lavis/processors/randaugment.py:40: RuntimeWarning: overflow encountered in scalar negative
  offset = -low * scale
Train: data epoch: [0]  [0/2]  eta: 0:00:12  lr: 0.000001  loss: 4.7803  time: 6.4565  data: 0.0000  max mem: 8321
2024-03-27 16:58:10,549 [INFO] Reducer buckets have been rebuilt in this iteration.
Train: data epoch: [0]  [1/2]  eta: 0:00:03  lr: 0.000001  loss: 5.7962  time: 3.6563  data: 0.0000  max mem: 8714
Train: data epoch: [0] Total time: 0:00:07 (3.6570 s / it)
2024-03-27 16:58:10,765 [INFO] Averaged stats: lr: 0.0000  loss: 5.2883
2024-03-27 16:58:10,770 [INFO] No validation splits found.
2024-03-27 16:58:10,815 [INFO] Saving checkpoint at epoch 0 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327165/checkpoint_0.pth.
2024-03-27 16:58:11,871 [INFO] Start training
2024-03-27 16:58:11,907 [INFO] Start training epoch 1, 2 iters per inner epoch.
Train: data epoch: [1]  [0/2]  eta: 0:00:06  lr: 0.000098  loss: 3.7817  time: 3.4140  data: 0.0000  max mem: 9036
Train: data epoch: [1]  [1/2]  eta: 0:00:02  lr: 0.000098  loss: 5.3590  time: 2.0641  data: 0.0000  max mem: 9555
Train: data epoch: [1] Total time: 0:00:04 (2.0650 s / it)
2024-03-27 16:58:16,037 [INFO] Averaged stats: lr: 0.0001  loss: 4.5704
2024-03-27 16:58:16,041 [INFO] No validation splits found.
2024-03-27 16:58:16,071 [INFO] Saving checkpoint at epoch 1 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327165/checkpoint_1.pth.
2024-03-27 16:58:18,357 [INFO] Start training
2024-03-27 16:58:18,399 [INFO] Start training epoch 2, 2 iters per inner epoch.
Train: data epoch: [2]  [0/2]  eta: 0:00:07  lr: 0.000091  loss: 4.9213  time: 3.6945  data: 0.0000  max mem: 9556
Train: data epoch: [2]  [1/2]  eta: 0:00:02  lr: 0.000091  loss: 6.5263  time: 2.1551  data: 0.0000  max mem: 9556
Train: data epoch: [2] Total time: 0:00:04 (2.1560 s / it)
2024-03-27 16:58:22,713 [INFO] Averaged stats: lr: 0.0001  loss: 5.7238
2024-03-27 16:58:22,717 [INFO] No validation splits found.
2024-03-27 16:58:22,765 [INFO] Saving checkpoint at epoch 2 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327165/checkpoint_2.pth.
2024-03-27 16:58:25,206 [INFO] Start training
2024-03-27 16:58:25,241 [INFO] Start training epoch 3, 2 iters per inner epoch.
Train: data epoch: [3]  [0/2]  eta: 0:00:06  lr: 0.000081  loss: 4.6143  time: 3.4231  data: 0.0000  max mem: 9556
Train: data epoch: [3]  [1/2]  eta: 0:00:01  lr: 0.000081  loss: 5.9560  time: 1.9856  data: 0.0000  max mem: 9556
Train: data epoch: [3] Total time: 0:00:03 (1.9866 s / it)
2024-03-27 16:58:29,215 [INFO] Averaged stats: lr: 0.0001  loss: 5.2851
2024-03-27 16:58:29,220 [INFO] No validation splits found.
2024-03-27 16:58:29,270 [INFO] Saving checkpoint at epoch 3 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327165/checkpoint_3.pth.
2024-03-27 16:58:31,729 [INFO] Start training
2024-03-27 16:58:31,772 [INFO] Start training epoch 4, 2 iters per inner epoch.
Train: data epoch: [4]  [0/2]  eta: 0:00:06  lr: 0.000069  loss: 4.4050  time: 3.3401  data: 0.0000  max mem: 9556
Train: data epoch: [4]  [1/2]  eta: 0:00:02  lr: 0.000069  loss: 5.5067  time: 2.1267  data: 0.0000  max mem: 9556
Train: data epoch: [4] Total time: 0:00:04 (2.1275 s / it)
2024-03-27 16:58:36,028 [INFO] Averaged stats: lr: 0.0001  loss: 4.9559
2024-03-27 16:58:36,031 [INFO] No validation splits found.
2024-03-27 16:58:36,064 [INFO] Saving checkpoint at epoch 4 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327165/checkpoint_4.pth.
2024-03-27 16:58:38,607 [INFO] Start training
2024-03-27 16:58:38,643 [INFO] Start training epoch 5, 2 iters per inner epoch.
Train: data epoch: [5]  [0/2]  eta: 0:00:07  lr: 0.000055  loss: 4.0671  time: 3.5544  data: 0.0000  max mem: 9556
Train: data epoch: [5]  [1/2]  eta: 0:00:02  lr: 0.000055  loss: 5.3782  time: 2.2167  data: 0.0000  max mem: 9556
Train: data epoch: [5] Total time: 0:00:04 (2.2181 s / it)
2024-03-27 16:58:43,080 [INFO] Averaged stats: lr: 0.0001  loss: 4.7227
2024-03-27 16:58:43,083 [INFO] No validation splits found.
2024-03-27 16:58:43,117 [INFO] Saving checkpoint at epoch 5 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327165/checkpoint_5.pth.
2024-03-27 16:58:45,519 [INFO] Start training
2024-03-27 16:58:45,555 [INFO] Start training epoch 6, 2 iters per inner epoch.
Train: data epoch: [6]  [0/2]  eta: 0:00:07  lr: 0.000041  loss: 4.1874  time: 3.5827  data: 0.0000  max mem: 9556
Train: data epoch: [6]  [1/2]  eta: 0:00:02  lr: 0.000041  loss: 3.9796  time: 2.1472  data: 0.0000  max mem: 9556
Train: data epoch: [6] Total time: 0:00:04 (2.1486 s / it)
2024-03-27 16:58:49,853 [INFO] Averaged stats: lr: 0.0000  loss: 4.0835
2024-03-27 16:58:49,856 [INFO] No validation splits found.
2024-03-27 16:58:49,889 [INFO] Saving checkpoint at epoch 6 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327165/checkpoint_6.pth.
2024-03-27 16:58:52,262 [INFO] Start training
2024-03-27 16:58:52,298 [INFO] Start training epoch 7, 2 iters per inner epoch.
Train: data epoch: [7]  [0/2]  eta: 0:00:06  lr: 0.000029  loss: 4.0438  time: 3.3334  data: 0.0000  max mem: 9556
Train: data epoch: [7]  [1/2]  eta: 0:00:02  lr: 0.000029  loss: 4.3911  time: 2.2152  data: 0.0000  max mem: 9556
Train: data epoch: [7] Total time: 0:00:04 (2.2175 s / it)
2024-03-27 16:58:56,737 [INFO] Averaged stats: lr: 0.0000  loss: 4.2175
2024-03-27 16:58:56,761 [INFO] No validation splits found.
2024-03-27 16:58:56,876 [INFO] Saving checkpoint at epoch 7 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327165/checkpoint_7.pth.
2024-03-27 16:58:59,494 [INFO] Start training
2024-03-27 16:58:59,534 [INFO] Start training epoch 8, 2 iters per inner epoch.
Train: data epoch: [8]  [0/2]  eta: 0:00:06  lr: 0.000019  loss: 3.9183  time: 3.3688  data: 0.0001  max mem: 9556
Train: data epoch: [8]  [1/2]  eta: 0:00:02  lr: 0.000019  loss: 4.3744  time: 2.1282  data: 0.0000  max mem: 9556
Train: data epoch: [8] Total time: 0:00:04 (2.1292 s / it)
2024-03-27 16:59:03,794 [INFO] Averaged stats: lr: 0.0000  loss: 4.1464
2024-03-27 16:59:03,797 [INFO] No validation splits found.
2024-03-27 16:59:03,830 [INFO] Saving checkpoint at epoch 8 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327165/checkpoint_8.pth.
2024-03-27 16:59:06,318 [INFO] Start training
2024-03-27 16:59:06,368 [INFO] Start training epoch 9, 2 iters per inner epoch.
Train: data epoch: [9]  [0/2]  eta: 0:00:06  lr: 0.000012  loss: 4.1729  time: 3.4050  data: 0.0000  max mem: 9556
Train: data epoch: [9]  [1/2]  eta: 0:00:02  lr: 0.000012  loss: 3.5403  time: 2.1497  data: 0.0000  max mem: 9556
Train: data epoch: [9] Total time: 0:00:04 (2.1513 s / it)
2024-03-27 16:59:10,672 [INFO] Averaged stats: lr: 0.0000  loss: 3.8566
2024-03-27 16:59:10,677 [INFO] No validation splits found.
2024-03-27 16:59:10,716 [INFO] Saving checkpoint at epoch 9 to /cyb/LAVIS/output/BLIP2/Pretrain_stage2/20240327165/checkpoint_9.pth.
2024-03-27 16:59:13,151 [INFO] No validation splits found.
2024-03-27 16:59:13,151 [INFO] Training time 0:01:13

dataloader

vg_captioin

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
class ImageTextPairDataset(BaseDataset, __DisplMixin):
    def __init__(self, vis_processor, text_processor, vis_root, ann_paths):
        """
        vis_root (string): Root directory of images (e.g. coco/images/)
        ann_root (string): directory to store the annotation file
        """
        super().__init__(vis_processor, text_processor, vis_root, ann_paths)

    def __getitem__(self, index):

        # TODO this assumes image input, not general enough
        ann = self.annotation[index]

        image_path = os.path.join(self.vis_root, ann["image"])
        image = Image.open(image_path).convert("RGB")

        image = self.vis_processor(image)
        caption = self.text_processor(ann["caption"])

        return {"image": image, "text_input": caption}

coco_caption

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
class CaptionDataset(BaseDataset, __DisplMixin):
    def __init__(self, vis_processor, text_processor, vis_root, ann_paths):
        """
        vis_root (string): Root directory of images (e.g. coco/images/)
        ann_root (string): directory to store the annotation file
        """
        super().__init__(vis_processor, text_processor, vis_root, ann_paths)

        self.img_ids = {}
        n = 0
        for ann in self.annotation:
            img_id = ann["image_id"]
            if img_id not in self.img_ids.keys():
                self.img_ids[img_id] = n
                n += 1

    def __getitem__(self, index):

        # TODO this assumes image input, not general enough
        ann = self.annotation[index]

        image_path = os.path.join(self.vis_root, ann["image"])
        image = Image.open(image_path).convert("RGB")

        image = self.vis_processor(image)
        caption = self.text_processor(ann["caption"])

        return {
            "image": image,
            "text_input": caption,
            # "image_id": self.img_ids[ann["image_id"]],
        }

model

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
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
SpikeModel(
  (model): VisionTransformer(
    (conv1): Conv2d(3, 768, kernel_size=(32, 32), stride=(32, 32), bias=False)
    (ln_pre): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
    (transformer): Transformer(
      (resblocks): Sequential(
        (0): ResidualAttentionBlock(
          (attn): SpikeAttention(
            (product): SpikeProduct()
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x_pos): X2X_POS(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
              )
            )
          )
          (ln_1): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
          (mlp): Sequential(
            (c_fc): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (gelu): Distilled_GELU(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (c_proj): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
          )
          (ln_2): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
        )
        (1): ResidualAttentionBlock(
          (attn): SpikeAttention(
            (product): SpikeProduct()
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x_pos): X2X_POS(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
              )
            )
          )
          (ln_1): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
          (mlp): Sequential(
            (c_fc): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (gelu): Distilled_GELU(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (c_proj): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
          )
          (ln_2): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
        )
        (2): ResidualAttentionBlock(
          (attn): SpikeAttention(
            (product): SpikeProduct()
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x_pos): X2X_POS(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
              )
            )
          )
          (ln_1): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
          (mlp): Sequential(
            (c_fc): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (gelu): Distilled_GELU(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (c_proj): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
          )
          (ln_2): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
        )
        (3): ResidualAttentionBlock(
          (attn): SpikeAttention(
            (product): SpikeProduct()
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x_pos): X2X_POS(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
              )
            )
          )
          (ln_1): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
          (mlp): Sequential(
            (c_fc): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (gelu): Distilled_GELU(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (c_proj): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
          )
          (ln_2): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
        )
        (4): ResidualAttentionBlock(
          (attn): SpikeAttention(
            (product): SpikeProduct()
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x_pos): X2X_POS(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
              )
            )
          )
          (ln_1): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
          (mlp): Sequential(
            (c_fc): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (gelu): Distilled_GELU(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (c_proj): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
          )
          (ln_2): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
        )
        (5): ResidualAttentionBlock(
          (attn): SpikeAttention(
            (product): SpikeProduct()
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x_pos): X2X_POS(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
              )
            )
          )
          (ln_1): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
          (mlp): Sequential(
            (c_fc): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (gelu): Distilled_GELU(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (c_proj): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
          )
          (ln_2): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
        )
        (6): ResidualAttentionBlock(
          (attn): SpikeAttention(
            (product): SpikeProduct()
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x_pos): X2X_POS(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
              )
            )
          )
          (ln_1): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
          (mlp): Sequential(
            (c_fc): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (gelu): Distilled_GELU(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (c_proj): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
          )
          (ln_2): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
        )
        (7): ResidualAttentionBlock(
          (attn): SpikeAttention(
            (product): SpikeProduct()
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x_pos): X2X_POS(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
              )
            )
          )
          (ln_1): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
          (mlp): Sequential(
            (c_fc): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (gelu): Distilled_GELU(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (c_proj): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
          )
          (ln_2): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
        )
        (8): ResidualAttentionBlock(
          (attn): SpikeAttention(
            (product): SpikeProduct()
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x_pos): X2X_POS(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
              )
            )
          )
          (ln_1): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
          (mlp): Sequential(
            (c_fc): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (gelu): Distilled_GELU(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (c_proj): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
          )
          (ln_2): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
        )
        (9): ResidualAttentionBlock(
          (attn): SpikeAttention(
            (product): SpikeProduct()
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x_pos): X2X_POS(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
              )
            )
          )
          (ln_1): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
          (mlp): Sequential(
            (c_fc): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (gelu): Distilled_GELU(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (c_proj): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
          )
          (ln_2): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
        )
        (10): ResidualAttentionBlock(
          (attn): SpikeAttention(
            (product): SpikeProduct()
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x_pos): X2X_POS(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
              )
            )
          )
          (ln_1): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
          (mlp): Sequential(
            (c_fc): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (gelu): Distilled_GELU(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (c_proj): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
          )
          (ln_2): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
        )
        (11): ResidualAttentionBlock(
          (attn): SpikeAttention(
            (product): SpikeProduct()
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x_pos): X2X_POS(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
              )
            )
          )
          (ln_1): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
          (mlp): Sequential(
            (c_fc): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (gelu): Distilled_GELU(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (c_proj): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
          )
          (ln_2): SpikeLN(
            (module): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
            (spike_sqrtinv): Distilled_SQRTINV(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
            (spike_x2x): X2X(
              (approximator): Sequential(
                (0): SpikeLinear_ReLU(
                  (relu): ReLU()
                )
                (1): StraightThrough()
                (2): SpikeLinear_ReLU(
                  (relu): StraightThrough()
                )
              )
            )
          )
        )
      )
    )
    (ln_post): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
  )
)

model2

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
SpikeModel(
  (model): OPTForCausalLM(
    (model): OPTModel(
      (decoder): OPTDecoder(
        (embed_tokens): Embedding(50272, 2560, padding_idx=1)
        (embed_positions): OPTLearnedPositionalEmbedding(2050, 2560)
        (final_layer_norm): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
        (layers): ModuleList(
          (0-31): 32 x OPTDecoderLayer(
            (self_attn): OPTAttention(
              (k_proj): SpikeLinear_ReLU(
                (relu): StraightThrough()
              )
              (v_proj): SpikeLinear_ReLU(
                (relu): StraightThrough()
              )
              (q_proj): SpikeLinear_ReLU(
                (relu): StraightThrough()
              )
              (out_proj): SpikeLinear_ReLU(
                (relu): ReLU()
              )
            )
            (activation_fn): StraightThrough()
            (self_attn_layer_norm): SpikeLN(
              (module): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
              (spike_sqrtinv): Distilled_SQRTINV(
                (approximator): Sequential(
                  (0): SpikeLinear_ReLU(
                    (relu): ReLU()
                  )
                  (1): StraightThrough()
                  (2): SpikeLinear_ReLU(
                    (relu): StraightThrough()
                  )
                )
              )
              (spike_x2x): X2X(
                (approximator): Sequential(
                  (0): SpikeLinear_ReLU(
                    (relu): ReLU()
                  )
                  (1): StraightThrough()
                  (2): SpikeLinear_ReLU(
                    (relu): StraightThrough()
                  )
                )
              )
            )
            (fc1): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (fc2): SpikeLinear_ReLU(
              (relu): StraightThrough()
            )
            (final_layer_norm): SpikeLN(
              (module): LayerNorm((2560,), eps=1e-05, elementwise_affine=True)
              (spike_sqrtinv): Distilled_SQRTINV(
                (approximator): Sequential(
                  (0): SpikeLinear_ReLU(
                    (relu): ReLU()
                  )
                  (1): StraightThrough()
                  (2): SpikeLinear_ReLU(
                    (relu): StraightThrough()
                  )
                )
              )
              (spike_x2x): X2X(
                (approximator): Sequential(
                  (0): SpikeLinear_ReLU(
                    (relu): ReLU()
                  )
                  (1): StraightThrough()
                  (2): SpikeLinear_ReLU(
                    (relu): StraightThrough()
                  )
                )
              )
            )
          )
        )
      )
    )
    (lm_head): Linear(in_features=2560, out_features=50272, bias=False)
  )
)

参考链接