| from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence | |
| import torch.nn as nn | |
| import numpy as np | |
| import torch | |
| class Vector2MIDI(nn.Module): | |
| def __init__(self, input_dim, hidden_dim, n_vocab, dropout=0.2): | |
| super().__init__() # 부모 클래스 생성자 호출 | |
| self.input_fc = nn.Linear(input_dim, hidden_dim) # 입력 차원에서 은닉 차원으로 변환 | |
| self.lstm = nn.LSTM(hidden_dim, hidden_dim, num_layers=2, batch_first=True, dropout=dropout) # 과적합 방지 드롭아웃 LSTM | |
| self.fc_mid = nn.Linear(hidden_dim, 256) | |
| self.fc_out = nn.Linear(256, n_vocab) | |
| def forward(self, x, lengths): | |
| # x: [batch, seq_len, input_dim] | |
| x = self.input_fc(x) | |
| # Token 길이가 Midi마다 다르니까 PackedSequence 변환 후 LSTM 처리 | |
| packed = pack_padded_sequence(x, lengths.cpu(), batch_first=True, enforce_sorted=False) # 패딩 변환 | |
| LSTM_out, _ = self.lstm(packed) # LSTM 처리 | |
| padded, _ = pad_packed_sequence(LSTM_out, batch_first=True) # 패딩 복원 | |
| # 최종 출력 | |
| x = self.fc_mid(padded) | |
| return self.fc_out(x) # [B, T, vocab_size] | |