forked from eilamshapira/HumanChoicePrediction
-
Notifications
You must be signed in to change notification settings - Fork 0
/
SpecialLSTM.py
70 lines (60 loc) · 3.28 KB
/
SpecialLSTM.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
62
63
64
65
66
67
68
69
70
import torch
import torch.nn as nn
from utils.usersvectors import UsersVectors
class SpecialLSTM(nn.Module):
def __init__(self, n_layers, input_dim, hidden_dim, output_dim, dropout, logsoftmax=True, input_twice=False):
super().__init__()
self.n_layers = n_layers
self.input_dim = input_dim
self.hidden_dim = hidden_dim
self.output_dim = output_dim
self.dropout = dropout
self.input_twice = input_twice
self.input_fc = nn.Sequential(nn.Linear(input_dim, input_dim * 2),
nn.Dropout(dropout),
nn.ReLU(),
nn.Linear(input_dim * 2, self.hidden_dim),
nn.Dropout(dropout),
nn.ReLU())
self.main_task = nn.LSTM(input_size=self.hidden_dim,
hidden_size=self.hidden_dim,
batch_first=True,
num_layers=self.n_layers,
dropout=dropout)
seq = [nn.Linear(self.hidden_dim + (input_dim if self.input_twice else 0), self.hidden_dim // 2),
nn.ReLU(),
nn.Linear(self.hidden_dim // 2, self.output_dim)]
if logsoftmax:
seq += [nn.LogSoftmax(dim=-1)]
self.output_fc = nn.Sequential(*seq)
self.user_vectors = UsersVectors(user_dim=self.hidden_dim, n_layers=self.n_layers)
self.game_vectors = UsersVectors(user_dim=self.hidden_dim, n_layers=self.n_layers)
def init_game(self, batch_size=1):
return torch.stack([self.game_vectors.init_user] * batch_size, dim=0)
def init_user(self, batch_size=1):
return torch.stack([self.user_vectors.init_user] * batch_size, dim=0)
def forward(self, input_vec, game_vector, user_vector):
lstm_input = self.input_fc(input_vec)
# lstm_input = lstm_input.reshape(-1, self.hidden_dim)
# output = self.output_fc(lstm_input)
lstm_shape = lstm_input.shape
shape = user_vector.shape
assert game_vector.shape == shape
if len(lstm_shape) != len(shape):
lstm_input = lstm_input.reshape((1,) * (len(shape) - 1) + lstm_input.shape)
user_vector = user_vector.reshape(shape[:-1][::-1] + (shape[-1],))
game_vector = game_vector.reshape(shape[:-1][::-1] + (shape[-1],))
lstm_output, (game_vector, user_vector) = self.main_task(lstm_input.contiguous(),
(game_vector.contiguous(),
user_vector.contiguous()))
user_vector = user_vector.reshape(shape)
game_vector = game_vector.reshape(shape)
if hasattr(self, "input_twice") and self.input_twice:
lstm_output = torch.cat([lstm_output, input_vec], dim=-1)
output = self.output_fc(lstm_output)
if len(output.shape) != len(lstm_shape):
output.reshape(-1, output.shape[-1])
if self.training:
return {"output": output, "game_vector": game_vector, "user_vector": user_vector}
else:
return {"output": output, "game_vector": game_vector.detach(), "user_vector": user_vector.detach()}