File size: 22,876 Bytes
bae5726 | 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 | """
Neural network models
"""
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
def torchActivation(activation='elu'):
'''Return an activation function from torch.nn'''
if activation == 'relu':
return nn.ReLU()
elif activation == 'leaky_relu':
return nn.LeakyReLU()
elif activation == 'elu':
return nn.ELU()
elif activation == 'selu':
return nn.SELU()
elif activation == 'gelu':
return nn.GELU()
def count_parameters(model):
'''Return a count of parameters and tensor shape of PyTorch model'''
counted = {}
total = 0
for name, param in model.named_parameters():
if param.requires_grad:
count = param.numel()
total += count
counted[name] = count
counted['FULL_MODEL'] = total
return counted
class ResidualDense(nn.Module):
'''A single dense layer with residual connection'''
def __init__(self, dim=2560, dropout=0.1, activation='elu', random_seed=0):
super(ResidualDense, self).__init__()
_ = torch.manual_seed(random_seed)
self.dense = nn.Linear(dim, dim)
self.batchnorm = nn.BatchNorm1d(dim)
self.activation = torchActivation(activation)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
x0 = x
x = self.dense(x)
x = self.batchnorm(x)
x = self.activation(x)
x = self.dropout(x)
x = x0 + x
return x
class LightAttention(nn.Module):
'''Convolution model with attention to learn pooled representations from embeddings'''
def __init__(self, dim=1280, kernel_size=7, random_seed=0):
super(LightAttention, self).__init__()
_ = torch.manual_seed(random_seed)
samepad = kernel_size // 2
self.values_conv = nn.Conv1d(dim, dim, kernel_size=kernel_size, padding=samepad)
self.weights_conv = nn.Conv1d(dim, dim, kernel_size=kernel_size, padding=samepad)
self.softmax = nn.Softmax(dim=-1)
def forward(self, x, mask=None):
if mask is None:
mask = torch.ones(x.shape[0], x.shape[2], dtype=torch.int32) # Don't mask out
values = self.values_conv(x)
values = values.masked_fill(mask[:,None,:]==0, -1e6)
weights = self.weights_conv(x)
weights = weights.masked_fill(mask[:,None,:]==0, -1e6)
weights = self.softmax(weights)
x_sum = torch.sum(values * weights, dim=-1) # Attention-weighted pooling
x_max, _ = torch.max(values, dim=-1) # Max pooling
x = torch.cat([x_sum, x_max], dim=1)
return x, weights
class ResidualLightAttention(nn.Module):
'''Model consisting of light attention followed by residual dense layers'''
def __init__(self, dim=1280, kernel_size=9, dropout=0.5,
activation='relu', res_blocks=4, random_seed=0):
super(ResidualLightAttention, self).__init__()
torch.manual_seed(random_seed)
self.light_attention = LightAttention(dim, kernel_size, random_seed)
self.batchnorm = nn.BatchNorm1d(2 * dim)
self.dropout = nn.Dropout(dropout)
self.residual_dense = nn.ModuleList()
for i in range(res_blocks):
self.residual_dense.append(
ResidualDense(2 * dim, dropout, activation, random_seed)
)
self.output = nn.Linear(2 * dim, 1)
def forward(self, x, mask=None):
x, weights = self.light_attention(x, mask)
x = self.batchnorm(x)
x = self.dropout(x)
for layer in self.residual_dense:
x = layer(x)
y = self.output(x).flatten()
return [y, x, weights]
class FeedForwardNetwork(nn.Module):
'''A feed-forward neural network (fully-connected)'''
def __init__(self, input_dim=1280, hidden_dim=128, num_layers=1, dropout=0.25, activation='relu',
residual=False):
super().__init__()
self.num_layers = num_layers
self.residual = residual
# First dense layer
# Separate first dense layer from the other dense layers to allow residual connection
self.first_dense = nn.ModuleList(
[
nn.Linear(input_dim, hidden_dim),
nn.BatchNorm1d(hidden_dim),
torchActivation(activation),
nn.Dropout(dropout) if activation != 'selu' else nn.AlphaDropout(dropout)
]
)
# Other dense layers after first dense and residual connection
if num_layers > 1:
self.other_dense = nn.ModuleList()
for i in range(1, num_layers):
self.other_dense.extend(
[
nn.Linear(hidden_dim, hidden_dim),
nn.BatchNorm1d(hidden_dim),
torchActivation(activation),
nn.Dropout(dropout) if activation != 'selu' else nn.AlphaDropout(dropout)
]
)
# Output layer
self.output_layer = nn.Linear(hidden_dim, 1)
def forward(self, x):
# First dense layer
for layer in self.first_dense:
x = layer(x)
# Other dense layers
if self.num_layers > 1:
x0 = x
for layer in self.other_dense:
x = layer(x)
# Residual connection
if self.residual:
x = x0 + x
# Output layer
y = self.output_layer(x).flatten()
return y
class CNN(nn.Module):
def __init__(self, input_channel=20, input_length=1024, start_conv_channel=32, kernel_size=3,
num_conv_layers=8, conv_dropout=0.25, pooltype='max', dense_dim=128, num_dense_layers=1,
dense_dropout=0.25, activation='relu'):
super().__init__()
# Conv layers
self.conv_layers = nn.ModuleList()
for i in range(num_conv_layers):
in_channel = input_channel if i==0 else out_channel
out_channel = start_conv_channel * (2 ** i)
self.conv_layers.append(nn.Conv1d(in_channel, out_channel, kernel_size=kernel_size,
padding=(kernel_size // 2)))
self.conv_layers.append(nn.BatchNorm1d(out_channel))
self.conv_layers.append(torchActivation(activation))
if pooltype=='max' and i>0:
self.conv_layers.append(nn.MaxPool1d(kernel_size=2, stride=2))
elif pooltype=='average' and i>0:
self.conv_layers.append(nn.AvgPool1d(kernel_size=2, stride=2))
self.conv_layers.append(nn.Dropout(conv_dropout))
# FNN layers
self.flatsize = start_conv_channel * input_length
self.fnn = FeedForwardNetwork(input_dim=self.flatsize, hidden_dim=dense_dim, num_layers=num_dense_layers,
dropout=dense_dropout, activation=activation, residual=True)
def forward(self, x):
for layer in self.conv_layers:
x = layer(x)
x = x.view(-1, self.flatsize)
y = self.fnn(x)
return y
class CausalConv1d(nn.Module):
'''Causal convolution'''
def __init__(self, in_channels, out_channels, kernel_size, dilation=1):
super().__init__()
self.pad = int(np.ceil(dilation * (kernel_size - 1) / 2)) # Padding to ensure causal convolution
self.conv_layer = nn.Conv1d(in_channels, out_channels, padding=0,kernel_size=kernel_size,
dilation=dilation)
def forward(self, x):
len1 = x.shape[-1]
x = F.pad(x, (self.pad * 2, 0))
x = self.conv_layer(x)
len2 = x.shape[-1]
start = len2 - len1
x = x[:,:,start:len2]
return x
class DCNNBlock(nn.Module):
def __init__(self,channel_size=32, num_layers=6,dropout=0.5, activation='elu'):
super().__init__()
dilation_rates = (2. ** np.arange(num_layers)).astype(np.int32)
self.layers = nn.ModuleList()
for d in dilation_rates:
self.layers.append(CausalConv1d(channel_size, channel_size, kernel_size=2, dilation=d))
self.layers.append(nn.BatchNorm1d(channel_size))
self.layers.append(torchActivation(activation))
self.layers.append(nn.Dropout(dropout))
def forward(self, x):
x0 = x
for layer in self.layers:
x = layer(x)
x = x0 + x
return x
class DCNN(nn.Module):
def __init__(self, input_channel=20, conv_channel=48, num_blocks=6, layers_per_block=10,
block_dropout=0.5, dense_dim=128, num_dense_layers=2, dense_dropout=0.25,
activation='elu'):
super().__init__()
self.embed_layer = nn.Conv1d(input_channel, conv_channel, kernel_size=1, bias=False)
# DCNN blocks
self.dcnn_blocks = nn.ModuleList()
for block in range(num_blocks):
self.dcnn_blocks.append(DCNNBlock(channel_size=conv_channel, num_layers=layers_per_block,
dropout=block_dropout, activation=activation))
self.fnn = FeedForwardNetwork(input_dim=conv_channel, hidden_dim=dense_dim,
num_layers=num_dense_layers, dropout=dense_dropout,
activation=activation, residual=True)
def forward(self, x, mask=None):
x = self.embed_layer(x)
# Dilated convolutions
for layer in self.dcnn_blocks:
x = layer(x)
if mask is None:
mask = torch.ones(x.shape[0], x.shape[2], dtype=torch.int32)
# Average pooling
x = x.masked_fill(mask[:,None,:] == 0, 0)
x = torch.divide(torch.sum(x, dim=-1), torch.sum(mask, dim=-1)[:,None])
# FNN
y = self.fnn(x).flatten()
return y
class RNN(nn.Module):
def __init__(self, input_channel=20, input_length=1024, gru_dim=1024, conv_downsample=1,
conv_dropout=0.25, dense_dim=128, dense_dropout=0.25, activation='relu'):
super().__init__()
# Convolutional layers to downsample
self.conv_downsample = conv_downsample
self.conv_layers = nn.ModuleList()
pad = (conv_downsample // 2 - 1)
pad = max(pad, 0)
self.conv_layers.append(nn.Conv1d(input_channel, gru_dim, kernel_size=conv_downsample,
stride=conv_downsample, padding=pad))
self.conv_layers.append(nn.BatchNorm1d(gru_dim))
self.conv_layers.append(torchActivation(activation))
self.conv_layers.append(nn.Dropout(conv_dropout))
# RNN layer
self.gru_layer = nn.GRU(gru_dim, gru_dim, batch_first=True, bidirectional=False)
# Dense layer
self.dense_layers = nn.ModuleList()
self.dense_layers.extend(
[
nn.Linear(gru_dim, dense_dim),
nn.BatchNorm1d(dense_dim),
torchActivation(activation),
nn.Dropout(dense_dropout) if activation != 'selu' else nn.AlphaDropout(dense_dropout)
]
)
# Output layer
self.output_layer = nn.Linear(dense_dim, 1)
def forward(self, x, mask=None):
if mask is None:
mask = torch.tensor(np.ones((x.shape[0], x.shape[-1])), dtype=torch.float32)
# Downsample with convolution
for layer in self.conv_layers:
x = layer(x)
# RNN
x = x.transpose(2, 1) # Transpose from [batch, feat, seqlen] to [batch, sequence, features]
x = self.gru_layer(x)[0]
lastind = torch.sum(mask, axis=1, dtype=int) # index of last residue in sequence
lastind = (lastind // self.conv_downsample) - 1 # index after conv. downsampling
x = [x[i][lastind[i]] for i in range(len(lastind))] # get hidden state of last residue for each sequence
x = torch.stack(x)
# Dense layer
for layer in self.dense_layers:
x = layer(x)
# Output layer
x = self.output_layer(x).flatten()
return x
class AttentionTopModel(nn.Module):
'''
Class to implement a top model based on LAT architecture to predict a property from
embeddings.
Parameters
-----------
input_dim : int
Channel dimension of input embedding, 1280 for ESM-1b
attention_dim : int
Channel dimension of transformed values and attention weights
kernel_size : int
Kernel size of convolution
conv_dropout : float
Dropout applied to values embedding after convolution
perceptive : bool
If True, concatenate input embedding with average pooled embedding to derive
a perceptive embedding.
dense_dim : int
Number of units in hidden dense layer
dense_dropout : float
Dropout of dense hidden layers.
activation : str
Activation function of dense hidden layer.
random_seed : int
Random seed for reproducibillity
'''
def __init__(self,
input_dim=1280,
attention_dim=1280,
kernel_size=9,
conv_dropout=0.25,
perceptive=True,
dense_dim=512,
dense_dropout=0.25,
num_layers=1,
residual=False,
activation='relu',
random_seed=0):
super(AttentionTopModel, self).__init__()
# Initialize class parameters
self.input_dim = input_dim
self.attention_dim = attention_dim
self.kernel_size = kernel_size
self.conv_dropout = conv_dropout
self.perceptive = perceptive
self.dense_dim = dense_dim
self.dense_dropout = dense_dropout
self.num_layers = num_layers
self.residual = residual
self.activation = activation
self.random_seed = random_seed
_ = torch.manual_seed(self.random_seed)
# A convolution layer to learn values from embeddings
self.values_conv = nn.Conv1d(
2 * self.input_dim if perceptive else self.input_dim,
self.attention_dim,
kernel_size=self.kernel_size,
stride=1,
padding=(self.kernel_size // 2)
)
self.values_dropout = nn.Dropout(self.conv_dropout)
# A convolution and softmax layer to learn attention weights that sum to 1
self.attn_conv = nn.Conv1d(
2 * self.input_dim if perceptive else self.input_dim,
self.attention_dim,
kernel_size=self.kernel_size,
stride=1,
padding=(self.kernel_size // 2)
)
self.attn_softmax = nn.Softmax(dim=-1)
# Dense(residual) layers on top of the attention-learned embeddings
self.dense_layers = FeedForwardNetwork(
input_dim=(2 * self.attention_dim),
hidden_dims=[self.dense_dim] * self.num_layers,
dropout=self.dense_dropout,
activation=self.activation,
residual=self.residual,
random_seed=self.random_seed
)
def forward(self, x, mask=None):
'''
Forward pass of network.
Parameters
------------
x : torch.Tensor
Shape is [batch_size, embeddings_dim, seq_len]
mask : torch.Tensor or None
Boolean tensor indicating padded (0) and non-padded (1) positions. Shape is
[batch_size, seq_len]. If mask is None, a tensor of ones is used,
indicating that padded positions are not masked out in the forward
computations.
Returns
-------
output: list
A list of predicted values (shape is [batch_size, 1]), embedding of the
hidden layer before the output layer (shape is [batch_size, dense_dim]), and
the masked attention weights (shape is [batch_size, attention_dim, seq_len])
'''
# Masking for padded positions
if mask is None:
# If mask is None, use all positions including padded positions
mask = torch.ones(x.shape[0], x.shape[2], dtype=torch.int32)
# Derive a perceptive tensor by concatenating mean pooled embeddings to the input
# Output shape of perceptive tensor is [batch_size, 2 * input_dim, seq_length]
if self.perceptive:
# Mask out padded positions
x_perc = x.masked_fill(mask[:,None,:] == 0, 0)
# Average pooling over unpadded positions
x_perc = torch.divide(torch.sum(x_perc, dim=-1),
torch.sum(mask, dim=1)[:,None])
# Repeat pooled tensor to match sequence length
x_perc = x_perc[:,:,None].repeat(1,1,x.shape[-1])
# Concatenate input embeddings and perceptive tensor along feature dimension
x = torch.cat((x, x_perc), axis=1)
# Derive values and attention weights from input embeddings with convolution
# Output shapes for values/weights is [batch_size, attention_dim, seq_length]
values = self.values_conv(x)
values = self.values_dropout(values)
values = values.masked_fill(mask[:,None,:] == 0, -1e6) # Mask out for max pooling
# Attention weights that sum to 1
attn = self.attn_conv(x)
attn = attn.masked_fill(mask[:,None,:] == 0, -1e6) # Mask out for average pooling
attn = self.attn_softmax(attn)
# Derive attention-learned embeddings by multiplying values and attention weights
# Output shape is [batch_size, 2 * attention_dim]
z_sum = torch.sum(values * attn, dim=-1)
z_max, _ = torch.max(values, dim=-1)
z = torch.cat([z_sum, z_max], dim=1)
# Dense layers
y, z = self.dense_layers(z)
return [y, z, attn]
class DilatedConvTopModel(nn.Module):
def __init__(self,
input_dim=1280,
input_len=1024,
attention_kernel_size=9,
conv_dim=64,
conv_kernel_size=3,
conv_dropout=0.1,
dilated_layers=9,
dense_dim=512,
dense_layers=1,
dense_dropout=0.25,
activation='relu',
random_seed=0):
super(DilatedConvTopModel, self).__init__()
_ = torch.manual_seed(random_seed)
# A convolution layer to learn transformed values from embeddings
self.values_conv_layer = nn.Conv1d(
input_dim,
conv_dim,
kernel_size=attention_kernel_size,
stride=1,
dilation=1,
padding=(attention_kernel_size // 2)
)
# A convolution and softmax layer to learn attention weights that sum to 1
# Attention-weighted values are obtained by element-wise multiplication with weights
self.attention_conv_layer = nn.Conv1d(
input_dim,
conv_dim,
kernel_size=attention_kernel_size,
stride=1,
dilation=1,
padding=(attention_kernel_size // 2)
)
self.attention_softmax = nn.Softmax(dim=-1)
self.conv_dropout_layer = nn.Dropout(conv_dropout) if activation != 'selu' \
else nn.AlphaDropout(conv_dropout)
# Convolutional layers with increasing dilation rates
self.dilated_conv_layers = nn.ModuleList()
dilations = np.power(2, np.arange(dilated_layers))
for d in dilations:
same_padding = self.get_same_padding_size(conv_kernel_size, d)
self.dilated_conv_layers.append(nn.Conv1d(conv_dim,
conv_dim,
kernel_size=conv_kernel_size,
padding=same_padding,
dilation=d)
)
self.dilated_conv_layers.append(nn.BatchNorm1d(conv_dim))
self.dilated_conv_layers.append(torchActivation(activation))
# Dense layers
self.dense_layers = FeedForwardNetwork(input_dim=(input_len * conv_dim),
hidden_dims=[dense_dim] * dense_layers,
dropout=dense_dropout,
activation=activation,
residual=(dense_layers > 1),
random_seed=random_seed)
def get_same_padding_size(self, k, d):
padsize = int(
np.ceil(
d * (k - 1) / 2
)
)
return padsize
def forward(self, x, mask=None):
# Masking for padded positions
if mask is None:
# If mask is None, use all positions including padded positions
mask = torch.ones(x.shape[0], x.shape[2], dtype=torch.int32)
# Derive attention-weighted values from input embeddings
# Shapes for values/weights is [batch_size, attention_dim, seq_length]
x_values = self.values_conv_layer(x)
x_weights = self.attention_conv_layer(x)
x_weights = x_weights.masked_fill(mask[:,None,:] == 0, -1e6)
x_weights = self.attention_softmax(x_weights) # Weights sum to 1
x = (x_values * x_weights)
x = self.conv_dropout_layer(x)
# Dilated convolution layers
x0 = x
for layer in self.dilated_conv_layers:
x = layer(x)
x = x0 + x # Residual connection
# Dense layers
x = x.flatten(start_dim=1, end_dim=2)
y, z = self.dense_layers(x)
return [y, z, x_weights]
|