File size: 18,684 Bytes
0b11a42
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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

import logging
import math
import random
from typing import Dict, Optional

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from omegaconf import DictConfig
from torch.nn.modules.normalization import LayerNorm

logger = logging.getLogger(__name__)

def circulant_mask(n: int, window: int) -> torch.Tensor:
    """Calculate the relative attention mask, calculated once when model instatiated, as a subset of this matrix
    will be used for a input length less than max.
    i,j represent relative token positions in this matrix and in the attention scores matrix,
     this mask enables attention scores to be set to 0 if further than the specified window length

        :param n: a fixed parameter set to be larger than largest max sequence length across batches
        :param window: [window length],
        :return relative attention mask
    """
    circulant_t = torch.zeros(n, n)
    # [0, 1, 2, ..., window, -1, -2, ..., window]
    offsets = [0] + [i for i in range(window + 1)] + [-i for i in range(window + 1)]
    if window >= n:
        return torch.ones(n, n)
    for offset in offsets:
        # size of the 1-tensor depends on the length of the diagonal
        circulant_t.diagonal(offset=offset).copy_(torch.ones(n - abs(offset)))
    return circulant_t


class SelfAttention(nn.Module):

    """normal query, key, value based self attention but with relative attention functionality
    and a learnable bias encoding relative token position which is added to the attention scores before the softmax"""

    def __init__(self, config: DictConfig, relative_attention: int):
        """init self attention weight of each key, query, value and output projection layer.

        :param config: model config
        :type config: ConveRTModelConfig
        """
        super().__init__()

        self.config = config
        self.query = nn.Linear(config.num_embed_hidden, config.num_attention_project)
        self.key = nn.Linear(config.num_embed_hidden, config.num_attention_project)
        self.value = nn.Linear(config.num_embed_hidden, config.num_attention_project)

        self.softmax = nn.Softmax(dim=-1)
        self.output_projection = nn.Linear(
            config.num_attention_project, config.num_embed_hidden
        )
        self.bias = torch.nn.Parameter(torch.randn(config.n), requires_grad=True)
        stdv = 1.0 / math.sqrt(self.bias.data.size(0))
        self.bias.data.uniform_(-stdv, stdv)
        self.relative_attention = relative_attention
        self.n = self.config.n
        self.half_n = self.n // 2
        self.register_buffer(
            "relative_mask",
            circulant_mask(config.tokens_len, self.relative_attention),
        )

    def forward(
        self, attn_input: torch.Tensor, attention_mask: torch.Tensor
    ) -> torch.Tensor:
        """calculate self-attention of query, key and weighted to value at the end.
        self-attention input is projected by linear layer at the first time.
        applying attention mask for ignore pad index attention weight. Relative attention mask
        applied and a learnable bias added to the attention scores.
        return value after apply output projection layer to value * attention

        :param attn_input: [description]
        :type attn_input: [type]
        :param attention_mask: [description], defaults to None
        :type attention_mask: [type], optional
        :return: [description]
        :rtype: [type]
        """
        self.T = attn_input.size()[1]
        # input is B x max seq len x n_emb
        _query = self.query.forward(attn_input)
        _key = self.key.forward(attn_input)
        _value = self.value.forward(attn_input)

        # scaled dot product
        attention_scores = torch.matmul(_query, _key.transpose(1, 2))
        attention_scores = attention_scores / math.sqrt(
            self.config.num_attention_project
        )

        # Relative attention

        # extended_attention_mask = attention_mask.to(attention_scores.device)  # fp16 compatibility
        extended_attention_mask = (1.0 - attention_mask.unsqueeze(-1)) * -10000.0
        attention_scores = attention_scores + extended_attention_mask

        # fix circulant_matrix to matrix of size 60 x60 (max token truncation_length,
        # register as buffer, so not keep creating masks of different sizes.

        attention_scores = attention_scores.masked_fill(
            self.relative_mask.unsqueeze(0)[:, : self.T, : self.T] == 0, float("-inf")
        )

        # Learnable bias vector is used of max size,for each i, different subsets of it are added to the scores, where the permutations
        # depend on the relative position (i-j). this way cleverly allows no loops. bias vector is 2*max truncation length+1
        # so has a learnable parameter for each eg. (i-j) /in {-60,...60} .

        ii, jj = torch.meshgrid(torch.arange(self.T), torch.arange(self.T))
        B_matrix = self.bias[self.n // 2 - ii + jj]

        attention_scores = attention_scores + B_matrix.unsqueeze(0)

        attention_scores = self.softmax(attention_scores)
        output = torch.matmul(attention_scores, _value)

        output = self.output_projection(output)

        return [output,attention_scores]  # B x T x num embed hidden 



class FeedForward1(nn.Module):
    def __init__(
        self, input_hidden: int, intermediate_hidden: int, dropout_rate: float = 0.0
    ):
        #          512         2048

        super().__init__()

        self.linear_1 = nn.Linear(input_hidden, intermediate_hidden)
        self.dropout = nn.Dropout(dropout_rate)
        self.linear_2 = nn.Linear(intermediate_hidden, input_hidden)

    def forward(self, x: torch.Tensor) -> torch.Tensor:

        x = F.gelu(self.linear_1(x))
        return self.linear_2(self.dropout(x))


class SharedInnerBlock(nn.Module):
    def __init__(self, config: DictConfig, relative_attn: int):
        super().__init__()

        self.config = config
        self.self_attention = SelfAttention(config, relative_attn)
        self.norm1 = LayerNorm(config.num_embed_hidden)  # 512
        self.dropout = nn.Dropout(config.dropout)
        self.ff1 = FeedForward1(
            config.num_embed_hidden, config.feed_forward1_hidden, config.dropout
        )
        self.norm2 = LayerNorm(config.num_embed_hidden)

    def forward(self, x: torch.Tensor, attention_mask: int) -> torch.Tensor:

        new_values_x,attn_scores = self.self_attention(x, attention_mask=attention_mask)
        x = x+new_values_x
        x = self.norm1(x)
        x = x + self.ff1(x)
        return self.norm2(x),attn_scores


# pretty basic, just single head. but done many times, stack to have another dimension (4 with batches).# so get stacks of B x H of attention scores T x T..
# then matrix multiply these extra stacks with the v
# (B xnh)x T xT . (Bx nh xTx hs) gives (B Nh) T x hs stacks. now  hs is set to be final dimension/ number of heads, so reorder the stacks (concatenating them)
# can have optional extra projection layer, but doing that later


class MultiheadAttention(nn.Module):
    def __init__(self, config: DictConfig):
        super().__init__()
        self.num_attention_heads = config.num_attention_heads
        self.num_attn_proj = config.num_embed_hidden * config.num_attention_heads
        self.attention_head_size = int(self.num_attn_proj / self.num_attention_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = nn.Linear(config.num_embed_hidden, self.num_attn_proj)
        self.key = nn.Linear(config.num_embed_hidden, self.num_attn_proj)
        self.value = nn.Linear(config.num_embed_hidden, self.num_attn_proj)

        self.dropout = nn.Dropout(config.dropout)

    def forward(
        self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None
    ) -> torch.Tensor:
        B, T, _ = hidden_states.size()
        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
        k = (
            self.key(hidden_states)
            .view(B, T, self.num_attention_heads, self.attention_head_size)
            .transpose(1, 2)
        )  # (B, nh, T, hs)
        q = (
            self.query(hidden_states)
            .view(B, T, self.num_attention_heads, self.attention_head_size)
            .transpose(1, 2)
        )  # (B, nh, T, hs)
        v = (
            self.value(hidden_states)
            .view(B, T, self.num_attention_heads, self.attention_head_size)
            .transpose(1, 2)
        )  # (B, nh, T, hs)

        attention_scores = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))

        if attention_mask is not None:
            attention_mask = attention_mask[:, None, None, :]
            attention_mask = (1.0 - attention_mask) * -10000.0

            attention_scores = attention_scores + attention_mask

        attention_scores = F.softmax(attention_scores, dim=-1)

        attention_scores = self.dropout(attention_scores)

        y = attention_scores @ v

        y = y.transpose(1, 2).contiguous().view(B, T, self.num_attn_proj)

        return y


class PositionalEncoding(nn.Module):
    def __init__(self, model_config: DictConfig,):
        super(PositionalEncoding, self).__init__()
        self.dropout = nn.Dropout(p=model_config.dropout)
        self.num_embed_hidden = model_config.num_embed_hidden
        pe = torch.zeros(model_config.tokens_len, self.num_embed_hidden)
        position = torch.arange(
            0, model_config.tokens_len, dtype=torch.float
        ).unsqueeze(1)
        div_term = torch.exp(
            torch.arange(0, self.num_embed_hidden, 2).float()
            * (-math.log(10000.0) / self.num_embed_hidden)
        )
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0)
        self.register_buffer("pe", pe)

    def forward(self, x):
        x = x + self.pe[: x.size(0), :]
        return self.dropout(x)


class RNAFFrwd(
    nn.Module
):  # params are not shared for context and reply. so need two sets of weights
    """Fully-Connected 3-layer Linear Model"""

    def __init__(self, model_config: DictConfig):
        """
        :param input_hidden: first-hidden layer input embed-dim
        :type input_hidden: int
        :param intermediate_hidden: layer-(hidden)-layer middle point weight
        :type intermediate_hidden: int
        :param dropout_rate: dropout rate, defaults to None
        :type dropout_rate: float, optional
        """
        # paper specifies,skip connections,layer normalization, and orthogonal initialization

        super().__init__()
        # 3,679,744 x2 params
        self.rna_ffwd_input_dim = (
            model_config.num_embed_hidden * model_config.num_attention_heads
        )
        self.linear_1 = nn.Linear(self.rna_ffwd_input_dim, self.rna_ffwd_input_dim)
        self.linear_2 = nn.Linear(self.rna_ffwd_input_dim, self.rna_ffwd_input_dim)

        self.norm1 = LayerNorm(self.rna_ffwd_input_dim)
        self.norm2 = LayerNorm(self.rna_ffwd_input_dim)
        self.final = nn.Linear(self.rna_ffwd_input_dim, model_config.num_embed_hidden)
        self.orthogonal_initialization()  # torch implementation works perfectly out the box,

    def orthogonal_initialization(self):
        for l in [
            self.linear_1,
            self.linear_2,
        ]: 
            torch.nn.init.orthogonal_(l.weight)

    def forward(self, x: torch.Tensor, attn_msk: torch.Tensor) -> torch.Tensor:
        sentence_lengths = attn_msk.sum(1)

        # adding square root reduction projection separately as not a shared.
        # part of the diagram torch.Size([Batch, scent_len, embedd_dim])

        # x has dims B x T x 2*d_emb
        norms = 1 / torch.sqrt(sentence_lengths.double()).float()  # 64
        # TODO: Aggregation is done on all words including the masked ones
        x = norms.unsqueeze(1) * torch.sum(x, dim=1)  # 64 x1024

        x = x + F.gelu(self.linear_1(self.norm1(x)))
        x = x + F.gelu(self.linear_2(self.norm2(x)))

        return F.normalize(self.final(x), dim=1, p=2)  # 64 512


class RNATransformer(nn.Module):
    def __init__(self, model_config: DictConfig):
        super().__init__()
        self.num_embedd_hidden = model_config.num_embed_hidden
        self.encoder = nn.Embedding(
            model_config.vocab_size, model_config.num_embed_hidden
        )
        self.model_input = model_config.model_input
        if 'baseline' not in self.model_input:
            # positional encoder
            self.pos_encoder = PositionalEncoding(model_config)

            self.transformer_layers = nn.ModuleList(
                [
                    SharedInnerBlock(model_config, int(window/model_config.window))
                    for window in model_config.relative_attns[
                        : model_config.num_encoder_layers
                    ]
                ]
            )
            self.MHA = MultiheadAttention(model_config)
            # self.concatenate = FeedForward2(model_config)

            self.rna_ffrwd = RNAFFrwd(model_config)
            self.pad_id = 0

    def forward(self, x:torch.Tensor) -> torch.Tensor:
        if x.is_cuda:
            long_tensor = torch.cuda.LongTensor
        else:
            long_tensor = torch.LongTensor
            
        embedds = self.encoder(x)
        if 'baseline' not in self.model_input:
            output = self.pos_encoder(embedds)
            attention_mask = (x != self.pad_id).int()
            
            for l in self.transformer_layers:
                output,attn_scores = l(output, attention_mask)
            output = self.MHA(output)
            output = self.rna_ffrwd(output, attention_mask)
            return output,attn_scores
        else:
            embedds = torch.flatten(embedds,start_dim=1)
            return embedds,None

class GeneEmbeddModel(nn.Module):
    def __init__(
        self, main_config: DictConfig,
    ):
        super().__init__()
        self.train_config = main_config["train_config"]
        self.model_config = main_config["model_config"]
        self.device = self.train_config.device
        self.model_input = self.model_config["model_input"]
        self.false_input_perc = self.model_config["false_input_perc"]
        #adjust n (used to add rel bias on attn scores)
        self.model_config.n = self.model_config.tokens_len*2+1
        self.transformer_layers = RNATransformer(self.model_config)
        #save tokens_len of sequences to be used to split ids between transformers
        self.tokens_len = self.model_config.tokens_len
        #reassign tokens_len and vocab_size to init a new transformer
        #more clean solution -> RNATransformer and its children should 
        # have a flag input indicating which transformer
        self.model_config.tokens_len = self.model_config.second_input_token_len
        self.model_config.n = self.model_config.tokens_len*2+1
        self.seq_vocab_size = self.model_config.vocab_size
        #this differs between both models not the token_len/ss_token_len
        self.model_config.vocab_size = self.model_config.second_input_vocab_size 

        self.second_input_model = RNATransformer(self.model_config)

        #num_transformers refers to using either one model or two in parallel
        self.num_transformers = 2
        if self.model_input == 'seq':
            self.num_transformers = 1
        # could be moved to model
        self.weight_decay = self.train_config.l2_weight_decay
        if 'baseline' in self.model_input:
            self.num_transformers = 1
            num_nodes = self.model_config.num_embed_hidden*self.tokens_len
            self.final_clf_1 = nn.Linear(num_nodes,self.model_config.num_classes)
        else:
            #setting classification layer
            num_nodes = self.num_transformers*self.model_config.num_embed_hidden
            if self.num_transformers == 1:
                self.final_clf_1 = nn.Linear(num_nodes,self.model_config.num_classes)
            else:
                self.final_clf_1 = nn.Linear(num_nodes,num_nodes)
                self.final_clf_2 = nn.Linear(num_nodes,self.model_config.num_classes)
                self.relu = nn.ReLU()
                self.BN = nn.BatchNorm1d(num_nodes)
                self.dropout = nn.Dropout(0.6)

        logger.info("number of parameters: %e", sum(p.numel() for p in self.parameters()))


    def distort_input(self,x):
        for sample_idx in range(x.shape[0]):
            seq_length = x[sample_idx,-1]
            num_tokens_flipped = int(self.false_input_perc*seq_length)
            max_start_flip_idx = seq_length - num_tokens_flipped

            random_feat_idx = random.randint(0,max_start_flip_idx-1)
            x[sample_idx,random_feat_idx:random_feat_idx+num_tokens_flipped] = \
                    torch.tensor(np.random.choice(range(1,self.seq_vocab_size-1),size=num_tokens_flipped,replace=True))

            x[sample_idx,random_feat_idx+self.tokens_len:random_feat_idx+self.tokens_len+num_tokens_flipped] = \
                    torch.tensor(np.random.choice(range(1,self.model_config.second_input_vocab_size-1),size=num_tokens_flipped,replace=True))
        return x
        
    def forward(self, x,train=False):
        if self.device == 'cuda':
            long_tensor = torch.cuda.LongTensor
            float_tensor = torch.cuda.FloatTensor
        else:
            long_tensor = torch.LongTensor
            float_tensor = torch.FloatTensor
        if train:
            if self.false_input_perc > 0:
                x = self.distort_input(x)

        gene_embedd,attn_scores_first = self.transformer_layers(
            x[:, : self.tokens_len].type(long_tensor)
        )
        attn_scores_second = None
        second_input_embedd,attn_scores_second = self.second_input_model(
                    x[:, self.tokens_len :-1].type(long_tensor)
                )

        #for tcga: if seq or baseline
        if self.num_transformers == 1:
            activations = self.final_clf_1(gene_embedd)
        else:
            out_clf_1 = self.final_clf_1(torch.cat((gene_embedd, second_input_embedd), 1))
            out = self.BN(out_clf_1)
            out = self.relu(out)
            out = self.dropout(out)
            activations = self.final_clf_2(out)
        
        #create dummy attn scores for baseline
        if 'baseline' in self.model_input:
            attn_scores_first = torch.ones((1,2,2),device=x.device)

        return [gene_embedd, second_input_embedd, activations,attn_scores_first,attn_scores_second]