File size: 8,793 Bytes
0a7c540 | 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 | import torch
import torch.nn as nn
from typing import List, Tuple
from torch.nn import TransformerEncoder, TransformerEncoderLayer
from transformers.models.bert import BertModel
from fastNLP.modules.torch import MLP,ConditionalRandomField,allowed_transitions
from torch.nn import CrossEntropyLoss
class ConvFeatureExtractionModel(nn.Module):
def __init__(
self,
conv_layers: List[Tuple[int, int, int]],
conv_dropout: float = 0.0,
conv_bias: bool = False,
):
super().__init__()
def block(n_in, n_out, k, stride=1, conv_bias=False):
padding = k // 2
return nn.Sequential(
nn.Conv1d(in_channels=n_in, out_channels=n_out, kernel_size=k, stride=stride, padding=padding, bias=conv_bias),
nn.Dropout(conv_dropout),
# nn.BatchNorm1d(n_out),
nn.ReLU(),
# nn.MaxPool1d(kernel_size=2, stride=2)
)
in_d = 1
self.conv_layers = nn.ModuleList()
for _, cl in enumerate(conv_layers):
assert len(cl) == 3, "invalid conv definition: " + str(cl)
(dim, k, stride) = cl
self.conv_layers.append(
block(in_d, dim, k, stride=stride, conv_bias=conv_bias))
in_d = dim
def forward(self, x):
# x = x.unsqueeze(1)
for conv in self.conv_layers:
x = conv(x)
return x
class ModelWiseCNNClassifier(nn.Module):
def __init__(self, id2labels, dropout_rate=0.1):
super(ModelWiseCNNClassifier, self).__init__()
feature_enc_layers = [(64, 5, 1)] + [(128, 3, 1)] * 3 + [(64, 3, 1)]
self.conv = ConvFeatureExtractionModel(
conv_layers=feature_enc_layers,
conv_dropout=0.0,
conv_bias=False,
)
embedding_size = 4 *64
self.norm = nn.LayerNorm(embedding_size)
self.label_num = len(id2labels)
self.dropout = nn.Dropout(dropout_rate)
self.classifier = nn.Sequential(nn.Linear(embedding_size, self.label_num))
self.crf = ConditionalRandomField(num_tags=self.label_num, allowed_transitions=allowed_transitions(id2labels))
self.crf.trans_m.data *= 0
def conv_feat_extract(self, x):
out = self.conv(x)
out = out.transpose(1, 2)
return out
def forward(self, x, labels):
x = x.transpose(1, 2)
out1 = self.conv_feat_extract(x[:, 0:1, :])
out2 = self.conv_feat_extract(x[:, 1:2, :])
out3 = self.conv_feat_extract(x[:, 2:3, :])
out4 = self.conv_feat_extract(x[:, 3:4, :])
outputs = torch.cat((out1, out2, out3, out4), dim=2)
outputs = self.norm(outputs)
dropout_outputs = self.dropout(outputs)
logits = self.classifier(dropout_outputs)
if self.training:
loss_fct = CrossEntropyLoss(ignore_index=-1)
loss = loss_fct(logits.view(-1, self.label_num), labels.view(-1))
output = {'loss': loss, 'logits': logits}
else:
mask = labels.gt(-1)
paths, scores = self.crf.viterbi_decode(logits=logits, mask=mask)
paths[mask==0] = -1
output = {'preds': paths, 'logits': logits}
pass
return output
class ModelWiseTransformerClassifier(nn.Module):
def __init__(self, id2labels, seq_len, intermediate_size = 512, num_layers=2, dropout_rate=0.1):
super(ModelWiseTransformerClassifier, self).__init__()
# feature_enc_layers = [(512, 10, 5)] + [(512, 3, 2)] * 4 + [(512,2,2)] + [(512,2,2)]
feature_enc_layers = [(64, 5, 1)] + [(128, 3, 1)] * 3 + [(64, 3, 1)]
self.conv = ConvFeatureExtractionModel(
conv_layers=feature_enc_layers,
conv_dropout=0.0,
conv_bias=False,
)
self.seq_len = seq_len # MAX Seq_len
embedding_size = 4 *64
self.encoder_layer = TransformerEncoderLayer(
d_model=embedding_size,
nhead=16,
dim_feedforward=intermediate_size,
dropout=dropout_rate,
batch_first=True)
self.encoder = TransformerEncoder(encoder_layer=self.encoder_layer,
num_layers=num_layers)
self.position_encoding = torch.zeros((seq_len, embedding_size))
for pos in range(seq_len):
for i in range(0, embedding_size, 2):
self.position_encoding[pos, i] = torch.sin(
torch.tensor(pos / (10000**((2 * i) / embedding_size))))
self.position_encoding[pos, i + 1] = torch.cos(
torch.tensor(pos / (10000**((2 *
(i + 1)) / embedding_size))))
self.norm = nn.LayerNorm(embedding_size)
self.label_num = len(id2labels)
self.dropout = nn.Dropout(dropout_rate)
self.classifier = nn.Sequential(nn.Linear(embedding_size, self.label_num))
self.crf = ConditionalRandomField(num_tags=self.label_num, allowed_transitions=allowed_transitions(id2labels))
self.crf.trans_m.data *= 0
def conv_feat_extract(self, x):
out = self.conv(x)
out = out.transpose(1, 2)
return out
def forward(self, x, labels):
mask = labels.gt(-1)
padding_mask = ~mask
x = x.transpose(1, 2)
out1 = self.conv_feat_extract(x[:, 0:1, :])
out2 = self.conv_feat_extract(x[:, 1:2, :])
out3 = self.conv_feat_extract(x[:, 2:3, :])
out4 = self.conv_feat_extract(x[:, 3:4, :])
out = torch.cat((out1, out2, out3, out4), dim=2)
outputs = out + self.position_encoding.to(out.device)
outputs = self.norm(outputs)
outputs = self.encoder(outputs, src_key_padding_mask=padding_mask)
dropout_outputs = self.dropout(outputs)
logits = self.classifier(dropout_outputs)
if self.training:
loss_fct = CrossEntropyLoss(ignore_index=-1)
loss = loss_fct(logits.view(-1, self.label_num), labels.view(-1))
output = {'loss': loss, 'logits': logits}
else:
paths, scores = self.crf.viterbi_decode(logits=logits, mask=mask)
paths[mask==0] = -1
output = {'preds': paths, 'logits': logits}
pass
return output
class TransformerOnlyClassifier(nn.Module):
def __init__(self, id2labels, seq_len, embedding_size=4, num_heads=2, intermediate_size=64, num_layers=2, dropout_rate=0.1):
super(TransformerOnlyClassifier, self).__init__()
self.encoder_layer = TransformerEncoderLayer(
d_model=embedding_size,
nhead=num_heads,
dim_feedforward=intermediate_size,
dropout=dropout_rate,
batch_first=True)
self.encoder = TransformerEncoder(encoder_layer=self.encoder_layer,
num_layers=num_layers)
self.position_encoding = torch.zeros((seq_len, embedding_size))
for pos in range(seq_len):
for i in range(0, embedding_size, 2):
self.position_encoding[pos, i] = torch.sin(
torch.tensor(pos / (10000**((2 * i) / embedding_size))))
self.position_encoding[pos, i + 1] = torch.cos(
torch.tensor(pos / (10000**((2 *
(i + 1)) / embedding_size))))
self.norm = nn.LayerNorm(embedding_size)
self.label_num = len(id2labels)
self.dropout = nn.Dropout(dropout_rate)
self.classifier = nn.Sequential(nn.Linear(embedding_size, self.label_num))
self.crf = ConditionalRandomField(num_tags=self.label_num, allowed_transitions=allowed_transitions(id2labels))
self.crf.trans_m.data *= 0
def forward(self, inputs, labels):
mask = labels.gt(-1)
padding_mask = ~mask
outputs = inputs + self.position_encoding.to(inputs.device)
outputs = self.norm(outputs)
outputs = self.encoder(outputs, src_key_padding_mask=padding_mask)
dropout_outputs = self.dropout(outputs)
logits = self.classifier(dropout_outputs)
if self.training:
loss_fct = CrossEntropyLoss(ignore_index=-1)
loss = loss_fct(logits.view(-1, self.label_num), labels.view(-1))
output = {'loss': loss, 'logits': logits}
else:
paths, scores = self.crf.viterbi_decode(logits=logits, mask=mask)
paths[mask==0] = -1
output = {'preds': paths, 'logits': logits}
pass
return output |