File size: 13,217 Bytes
5032722
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Mon Jan 27 12:44:26 2025


About:
=======
Concatenated outputs from both sequence embedders and postprocess for 
downstream blocks that create logits, features, etc.


classes available:
==================
1.) SelectMask: concatenate outputs from sequence embedders, possibly norm
2.) FeedforwardPostproc: repeat blocks of norm -> dense -> act -> dropout

"""
from flax import linen as nn
import jax
import jax.numpy as jnp

from typing import Optional

from utils.BaseClasses import ModuleBase


class SelectMask(ModuleBase):
    config: dict
    name: str
    
    def setup(self):
        self.use_anc_emb = self.config['use_anc_emb']
        self.use_desc_emb = self.config['use_desc_emb']
        self.use_prev_align_info = self.config['use_prev_align_info']
        self.use_t_per_sample = self.config.get('t_per_sample', False)
        self.normalize_seq_embeddings_before_block = self.config.get('normalize_seq_embeddings_before_block', False)
        
        if self.normalize_seq_embeddings_before_block:
            self.norm = nn.LayerNorm( reduction_axes= -1, 
                                      feature_axes=-1,
                                      name = f'{self.name}/InstanceNorm')
    
    def _mask_padding_tokens( self,
                             x: jnp.array, 
                             mask: jnp.array ):
        expanded_mask = jnp.broadcast_to( mask[...,None], x.shape ) 
        return jnp.multiply(expanded_mask, x)
    
    
    def _concatenate_and_norm( self,
                               padding_mask: jnp.array,
                               anc_emb: Optional[jnp.array] = None,
                               desc_causal_emb: Optional[jnp.array] = None,
                               prev_align_one_hot_vec: Optional[jnp.array] = None,
                               t_array: Optional[jnp.array] = None,
                               norm_fn = None ):
        ### combine sequence embeddings
        to_concat = []
        if self.use_anc_emb and (anc_emb is not None):
            to_concat.append( anc_emb )
            
        if self.use_desc_emb and (desc_causal_emb is not None):
            to_concat.append( desc_causal_emb )
        
        # concatenate, mask; embeddings_datamat could be:
        #   (B, L_align, H): (use_anc_emb | use_desc_emb) 
        #   (B, L_align, 2H): (use_anc_emb & use_desc_emb) 
        embeddings_datamat = jnp.concatenate( to_concat, axis = -1 ) # (B, L, n*H)
        del to_concat
        
        # possibly normalize and mask
        embeddings_datamat = norm_fn(embeddings_datamat) # (B, L, n*H)
        embeddings_datamat = self._mask_padding_tokens( x = embeddings_datamat,
                                                        mask = padding_mask ) # (B, L, n*H)
            
        
        ### possibly concat other things (outside of normalization)
        to_concat = [embeddings_datamat]
        
        if self.use_prev_align_info and (prev_align_one_hot_vec is not None):
            to_concat.append(prev_align_one_hot_vec) #(B, L, 5)
        
        if self.use_t_per_sample and (t_array is not None):
            B = t_array.shape[0]
            L = embeddings_datamat.shape[1]
            out_shape = ( (B, L, 1) )
            t_array_exp = jnp.broadcast_to( t_array[..., None, None], out_shape ) #(B, L, 1)
            to_concat.append( t_array_exp )
            
        datamat = jnp.concatenate( [] + to_concat, axis=-1 ) #(B, L, H_out)
        
        # mask again
        datamat = self._mask_padding_tokens( x = datamat,
                                             mask = padding_mask )
        
        return datamat
    
        
    def __call__(self,
                 sow_flax_intermeds: bool,
                 anc_emb: Optional[jnp.array] = None,
                 desc_causal_emb: Optional[jnp.array] = None,
                 prev_align_one_hot_vec: Optional[jnp.array] = None,
                 t_array: Optional[jnp.array] = None,
                 padding_mask: Optional[jnp.array] = None,
                 *args,
                 **kwargs):
        """
        B: batch size
        L_align: length of alignment
        
        Arguments
        ----------
        sow_flax_intermeds : bool

        anc_emb : ArrayLike, (B, L, H) or None
        
        desc_causal_emb : ArrayLike, (B, L, H) or None
        
        prev_align_one_hot_vec : ArrayLike, (B, L, 5) or None
        
        t_array : ArrayLike, (B,) or None
            > note: this is ONE unique time per branch length!

        padding_mask : ArrayLike, (B, L) or None
        
        Returns
        --------
        datamat : ArrayLike, (B, L_align, H_out)
            concatenated/masked data
            
        """
        datamat = self._concatenate_and_norm( padding_mask = padding_mask,
                                              anc_emb = anc_emb, 
                                              desc_causal_emb = desc_causal_emb,
                                              prev_align_one_hot_vec = prev_align_one_hot_vec,
                                              t_array = t_array,
                                              norm_fn = lambda x: x if not self.normalize_seq_embeddings_before_block else self.norm )
        
        # possibly sow values
        self.maybe_sow(sow_flax_intermeds = sow_flax_intermeds,
                       vals = datamat,
                       label = f'{self.name}/concatenated_embeddings_features',
                       include_min_max = True,
                       include_perc_zeros = True)
        
        return datamat
        

class FeedforwardPostproc(SelectMask):
    """
    apply this blocks as many times as specified by layer_sizes: 
        [norm -> dense -> activation -> dropout]
    """
    config: dict
    name: str
    
    def setup(self):
        """
        B = batch size
        L_align = alignment length
        A = alphabet size
        
        """
        ### read config
        # required
        self.layer_sizes = self.config['layer_sizes']
        self.use_anc_emb = self.config['use_anc_emb']
        self.use_desc_emb = self.config['use_desc_emb']
        self.use_prev_align_info = self.config['use_prev_align_info']
        
        # optional
        self.use_t_per_sample = self.config.get('use_t_per_sample', False)
        self.normalize_seq_embeddings_before_block = self.config.get("normalize_seq_embeddings_before_block", True)
        self.dropout = self.config.get("dropout", 0.0)
        use_bias = self.config.get("use_bias", True)
        
        
        ### set up parameterized layers of the MLP
        dense_layers = []
        norm_layers = []
        
        # first layer: possible instance norm and dense
        if self.normalize_seq_embeddings_before_block:
            norm_layers.append( nn.LayerNorm( reduction_axes= -1, 
                                              feature_axes=-1,
                                              name = f'{self.name}/instance norm 0') )
        
        elif not self.normalize_seq_embeddings_before_block:
            norm_layers.append( lambda x: x )
        
        dense_layers.append( nn.Dense(features = self.layer_sizes[0], 
                                      use_bias = use_bias, 
                                      kernel_init = nn.initializers.lecun_normal(),
                                      name=f'{self.name}/feedforward layer 0') )
        
        # subsequent normalization and dense layers
        for i, hid_dim in enumerate(self.layer_sizes[1:]):
            layer_idx = i + 1
            norm_layers.append( nn.LayerNorm( reduction_axes= -1, 
                                              feature_axes=-1,
                                              name = f'{self.name}/instance norm {layer_idx}') )
            dense_layers.append( nn.Dense(features = hid_dim, 
                                          use_bias = use_bias, 
                                          kernel_init = nn.initializers.lecun_normal(),
                                          name=f'{self.name}/feedforward layer {layer_idx}') )
            
        self.dense_layers = dense_layers
        self.norm_layers = norm_layers
        self.act= nn.silu 
    
    @nn.compact
    def __call__(self, 
                 sow_flax_intermeds: bool,
                 training: bool,
                 anc_emb: Optional[jnp.array] = None,
                 desc_causal_emb: Optional[jnp.array] = None,
                 prev_align_one_hot_vec: Optional[jnp.array] = None,
                 t_array: Optional[jnp.array] = None,
                 padding_mask: Optional[jnp.array] = None,
                 *args,
                 **kwargs):
        """
        B: batch size
        L_align: length of alignment
        H_in, H_out: size of embedding dimension in/out of this block
        
        Arguments
        ----------
        sow_flax_intermeds : bool

        anc_emb : ArrayLike, (B, L, H) or None
        
        desc_causal_emb : ArrayLike, (B, L, H) or None
        
        prev_align_one_hot_vec : ArrayLike, (B, L, 5) or None
        
        t_array : ArrayLike, (B,) or None
            > note: this is ONE unique time per branch length!

        padding_mask : ArrayLike, (B, L) or None
        
        Returns
        --------
        datamat : ArrayLike, (B, L_align, H_out)
            concatenated and post-processed data
            > n=1, if only using ancestor embedding OR descendant embedding
            > n=2, if using both embeddings
        """
        ### First block
        # 1.) select (potentially concat+norm) the ancestor and descendant embeddings
        # padding tokens are masked in this function
        datamat = self._concatenate_and_norm( padding_mask = padding_mask,
                                              anc_emb = anc_emb, 
                                              desc_causal_emb = desc_causal_emb,
                                              prev_align_one_hot_vec = prev_align_one_hot_vec,
                                              t_array = t_array,
                                              norm_fn = None if not self.normalize_seq_embeddings_before_block else self.norm_layers[0] ) #(B, L, H_out)
        
        # 2.) dense layer, mask
        datamat = self.dense_layers[0](datamat) #(B, L, hid_dim[0] )
        datamat = self._mask_padding_tokens( x=datamat,
                                             mask=padding_mask ) #(B, L, hid_dim[0] )
        
        # 3.) activation
        datamat = self.act(datamat) #(B, L, hid_dim[0] )
        
        # 4.) dropout
        datamat = nn.Dropout(rate = self.dropout)(datamat,
                                            deterministic = not training)  #(B, L, layer_sizes[0] )
         
        # record results from first block
        self.maybe_sow(sow_flax_intermeds = sow_flax_intermeds,
                       vals = datamat,
                       label = f'{self.name}/final_feedforward_layer_0/after_dropout',
                       include_min_max = True,
                       include_perc_zeros = False)
        
        # mask tokens after block
        datamat = self._mask_padding_tokens( x=datamat,
                                             mask=padding_mask ) #(B, L, layer_sizes[0] )
        
        
        ### Subsequent blocks: norm -> dense -> activation -> dropout
        for i in range( len(self.layer_sizes[1:]) ):
            layer_idx = i + 1
             
            # 1.) instance norm, mask
            datamat = self.norm_layers[layer_idx](datamat) #(B, L, layer_sizes[layer_idx-1])
            datamat = self._mask_padding_tokens( x=datamat,
                                                 mask=padding_mask ) #(B, L, layer_sizes[layer_idx-1])
            
            # 2.) dense, mask
            datamat = self.dense_layers[layer_idx](datamat) #(B, L, layer_sizes[layer_idx-1])
            datamat = self._mask_padding_tokens( x=datamat,
                                                 mask=padding_mask ) #(B, L, layer_sizes[layer_idx-1])
            
            # 3.) activation
            datamat = self.act(datamat) #(B, L, layer_sizes[layer_idx-1])
            
            # 4.) dropout
            datamat = nn.Dropout(rate = self.dropout)(datamat,
                                                deterministic = not training) #(B, L, layer_sizes[layer_idx-1])
            
            # record results from first block
            self.maybe_sow(sow_flax_intermeds = sow_flax_intermeds,
                           vals = datamat,
                           label = (f'{self.name}/'+
                                    f'final_feedforward_layer_{layer_idx}/'+
                                    f'after_dropout'),
                           include_min_max = True,
                           include_perc_zeros = False)
        
            # mask out padding tokens after block
            datamat = self._mask_padding_tokens( x=datamat,
                                                 mask=padding_mask ) #(B, L, layer_sizes[-1])
            
        return datamat  #(B, L, layer_sizes[-1])