-
Notifications
You must be signed in to change notification settings - Fork 0
/
model_utils.py
61 lines (44 loc) · 1.89 KB
/
model_utils.py
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import edit_distance
def find_lengths_onehot(messages: torch.Tensor, vocab_size) -> torch.Tensor:
zero = F.one_hot(torch.tensor([0]), vocab_size).float()[0]
zero = zero.to(messages.device)
max_k = messages.size(1)
zero_mask = messages == zero
zero_mask = zero_mask.sum(dim=-1) == vocab_size
lengths = max_k - (zero_mask.cumsum(dim=1) > 0).sum(dim=1)
lengths.add_(1).clamp_(max=max_k)
return lengths
def find_lengths(messages: torch.Tensor) -> torch.Tensor:
max_k = messages.size(1)
zero_mask = messages == 0
lengths = max_k - (zero_mask.cumsum(dim=1) > 0).sum(dim=1)
lengths.add_(1).clamp_(max=max_k)
return lengths
def get_edit_distance(ref, hyp):
sm = edit_distance.SequenceMatcher(a=list(ref), b=list(hyp))
return sm.distance()
def add_eos_to_messages(message, message_length, max_length):
cleaned_message = torch.zeros_like(message)
for i in range(max_length):
not_eosed = (i < message_length).float()
cleaned_message[:, i] = message[:, i] * not_eosed
return cleaned_message
class MessageClassifier(nn.Module):
def __init__(self, n_classes, message_max_length):
super(MessageClassifier, self).__init__()
self.linearlayer =nn.Linear(message_max_length, n_classes, bias=True)
def forward(self, x):
x = self.linearlayer(x)
return x
class MessageMLP(nn.Module):
def __init__(self, n_classes, message_max_length, hidden_size=10):
super(MessageMLP, self).__init__()
self.mlp =nn.Sequential(nn.Linear(message_max_length, hidden_size, bias=False),
nn.ReLU(inplace=True),
nn.Linear(hidden_size, n_classes, bias=True))
def forward(self, x):
x = self.mlp(x)
return x