|
a |
|
b/src/hybrid/hybrid.py |
|
|
1 |
import torch |
|
|
2 |
import torch.nn as nn |
|
|
3 |
from src.rnn.rnn_utils import create_emb_layer |
|
|
4 |
|
|
|
5 |
|
|
|
6 |
class hybrid(nn.Module): |
|
|
7 |
def __init__(self, vocabulary, sequence_length, weights_matrix, hidden_size, num_layers=2, num_classes=10): |
|
|
8 |
super().__init__() |
|
|
9 |
|
|
|
10 |
self.num_layers = num_layers |
|
|
11 |
self.hidden_size = hidden_size |
|
|
12 |
|
|
|
13 |
self.conv1 = nn.Sequential(nn.Conv1d(len(vocabulary)+1, |
|
|
14 |
128, |
|
|
15 |
kernel_size=7, |
|
|
16 |
padding=0), |
|
|
17 |
nn.ReLU(), |
|
|
18 |
nn.MaxPool1d(3) |
|
|
19 |
) |
|
|
20 |
|
|
|
21 |
self.conv2 = nn.Sequential(nn.Conv1d(128, 128, kernel_size=7, padding=0), |
|
|
22 |
nn.ReLU(), |
|
|
23 |
nn.MaxPool1d(3) |
|
|
24 |
) |
|
|
25 |
|
|
|
26 |
self.conv3 = nn.Sequential(nn.Conv1d(128, 128, kernel_size=3, padding=0), |
|
|
27 |
nn.ReLU() |
|
|
28 |
) |
|
|
29 |
|
|
|
30 |
self.conv4 = nn.Sequential(nn.Conv1d(128, 128, kernel_size=3, padding=0), |
|
|
31 |
nn.ReLU() |
|
|
32 |
) |
|
|
33 |
input_shape = (1, len(vocabulary)+1, sequence_length) |
|
|
34 |
self.output_dimension = self._get_conv_output(input_shape) |
|
|
35 |
|
|
|
36 |
# define linear layers |
|
|
37 |
|
|
|
38 |
self.fc1 = nn.Sequential( |
|
|
39 |
nn.Linear(self.output_dimension, 256), |
|
|
40 |
nn.ReLU(), |
|
|
41 |
) |
|
|
42 |
|
|
|
43 |
|
|
|
44 |
self.embeddings, num_embeddings, embedding_size = create_emb_layer(weights_matrix, True) |
|
|
45 |
self.gru1 = nn.GRU(embedding_size, hidden_size, num_layers, bidirectional = True, batch_first=True) |
|
|
46 |
|
|
|
47 |
self.fc2 = nn.Sequential( |
|
|
48 |
nn.Linear(2*hidden_size, 256), |
|
|
49 |
nn.ReLU(), |
|
|
50 |
) |
|
|
51 |
|
|
|
52 |
|
|
|
53 |
self.fc3 = nn.Linear(512,10) |
|
|
54 |
self.act = nn.Sigmoid() |
|
|
55 |
|
|
|
56 |
|
|
|
57 |
def _get_conv_output(self, shape): |
|
|
58 |
x = torch.rand(shape) |
|
|
59 |
x = self.conv1(x) |
|
|
60 |
x = self.conv2(x) |
|
|
61 |
x = self.conv3(x) |
|
|
62 |
x = self.conv4(x) |
|
|
63 |
x = x.view(x.size(0), -1) |
|
|
64 |
output_dimension = x.size(1) |
|
|
65 |
return output_dimension |
|
|
66 |
|
|
|
67 |
def forward(self,rnninput, cnninput): |
|
|
68 |
cnn_out = self.conv1(cnninput) |
|
|
69 |
cnn_out = self.conv2(cnn_out) |
|
|
70 |
cnn_out = self.conv3(cnn_out) |
|
|
71 |
cnn_out = self.conv4(cnn_out) |
|
|
72 |
cnn_out = cnn_out.view(cnn_out.size(0),-1) |
|
|
73 |
cnn_out = self.fc1(cnn_out) |
|
|
74 |
|
|
|
75 |
rnn_out = self.embeddings(rnninput) |
|
|
76 |
rnn_out,_ = self.gru1(rnn_out) |
|
|
77 |
rnn_out = self.fc2(rnn_out[:,-1,:]) |
|
|
78 |
|
|
|
79 |
|
|
|
80 |
x = torch.cat((cnn_out,rnn_out),dim=1) |
|
|
81 |
out = self.fc3(x) |
|
|
82 |
out = self.act(out) |
|
|
83 |
return out |