teng-ml/teng_ml/rnn/rnn.py

78 lines
3.7 KiB
Python
Raw Permalink Normal View History

2023-05-10 22:44:14 +02:00
import torch
import torch.nn as nn
class RNN(nn.Module):
"""
(Bi)LSTM for name classification
"""
def __init__(self, input_size, hidden_size, num_layers, num_classes, bidirectional):
super(RNN, self).__init__()
self.num_layers = num_layers
self.hidden_size = hidden_size
self.is_bidirectional = bidirectional
self.lstm = nn.LSTM(input_size=input_size, hidden_size=hidden_size, num_layers=num_layers, batch_first=True, bidirectional=bidirectional)
if bidirectional == True:
self.fc = nn.Linear(hidden_size * 2, num_classes)
else:
self.fc = nn.Linear(hidden_size, num_classes)
self.softmax = nn.Softmax(dim=1)
2023-05-26 14:01:15 +02:00
self.D = 2 if self.is_bidirectional == True else 1
2023-05-10 22:44:14 +02:00
2023-08-10 17:29:09 +02:00
def forward(self, x, unpadded_lengths=None):
"""
@param x:
Tensor (seq_length, features) for unbatched inputs
Tensor (batch_size, seq_length, features) for batch inputs
PackedSequence for padded batched inputs
@param unpadded_lengths: Tensor(batch_size) with lengths of the unpadded sequences, when using padding but without PackedSequence
@returns (batch_size, num_classes) with batch_size == 1 for unbatched inputs
"""
# if type(x) == torch.Tensor:
# device = x.device
# # h0: initial hidden states
# # c0: initial cell states
# if len(x.shape) == 2: # x: (seq_length, features)
# h0 = torch.zeros(self.D * self.num_layers, self.hidden_size).to(device)
# c0 = torch.zeros(self.D * self.num_layers, self.hidden_size).to(device)
# elif len(x.shape) == 3: # x: (batch, seq_length, features)
# batch_size = x.shape[0]
# h0 = torch.zeros(self.D * self.num_layers, batch_size, self.hidden_size).to(device)
# c0 = torch.zeros(self.D * self.num_layers, batch_size, self.hidden_size).to(device)
# else:
# raise ValueError(f"RNN.forward: invalid input shape: {x.shape}. Must be (batch, seq_length, features) or (seq_length, features)")
# elif type(x) == nn.utils.rnn.PackedSequence:
# device = x.data.device
# h0 = torch.zeros(self.D * self.num_layers, self.hidden_size).to(device)
# c0 = torch.zeros(self.D * self.num_layers, self.hidden_size).to(device)
# else:
# raise ValueError(f"RNN.forward: invalid input type: {type(x)}. Must be Tensor or PackedSequence")
2023-05-10 22:44:14 +02:00
2023-05-26 14:01:15 +02:00
# lstm: (batch_size, seq_length, features) -> (batch_size, hidden_size)
2023-08-10 17:29:09 +02:00
# or: packed_sequence -> packed_sequence
# out, (h_n, c_n) = self.lstm(x, (h0, c0))
out, (h_n, c_n) = self.lstm(x) # (h0, c0) defaults to zeros
2023-05-10 22:44:14 +02:00
2023-08-10 17:29:09 +02:00
# select the last state of lstm's neurons
if type(out) == nn.utils.rnn.PackedSequence:
# padding has to be considered
out, lengths = nn.utils.rnn.pad_packed_sequence(out, batch_first=True)
# the unpadded length of batch i is lengths[i], so that is the last non-zero state
out = torch.stack([out[i,lengths[i].item()-1,:] for i in range(len(lengths))])
elif unpadded_lengths is not None:
out = torch.stack([out[i,unpadded_lengths[i].item()-1,:] for i in range(len(unpadded_lengths))])
2023-05-10 22:44:14 +02:00
else:
2023-08-10 17:29:09 +02:00
if out.shape[0] == 3: # batched
out = out[:,-1,:]
else: # unbatched
# softmax requires (batch_size, *)
out = torch.stack([out[-1,:]])
2023-05-10 22:44:14 +02:00
2023-05-26 14:01:15 +02:00
# fc fully connected layer: (*, hidden_size) -> (*, num_classes)
out = self.fc(out)
2023-05-10 22:44:14 +02:00
2023-08-10 17:29:09 +02:00
# softmax: (batch_size, *) -> (batch_size, *)
2023-05-26 14:01:15 +02:00
out = self.softmax(out)
2023-05-10 22:44:14 +02:00
return out