Models/Vector2MIDI.py

from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
import torch.nn as nn
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

        # 과적합 방지 드롭아웃 LSTM
        self.lstm = nn.LSTM(hidden_dim, hidden_dim, num_layers=2, batch_first=True, dropout=dropout)

        self.fc_mid = nn.Linear(hidden_dim, 256)
        self.fc_out = nn.Linear(256, n_vocab)

    def forward(self, x, lengths, total_length=None):
        print("input to forward:", x.shape)
        B, feat_dim = x.size()
        T = lengths.max()

        # [B, 1, feat_dim] → [B, T, feat_dim]
        x = x.unsqueeze(1).expand(B, T, feat_dim)

        x = self.input_fc(x)

        packed_x = nn.utils.rnn.pack_padded_sequence(
            x, lengths.cpu(), batch_first=True, enforce_sorted=False
        )
        packed_out, _ = self.lstm(packed_x)

        out, _ = nn.utils.rnn.pad_packed_sequence(
            packed_out, batch_first=True, total_length=total_length
        )

        out = self.fc_mid(out)
        out = self.fc_out(out)  # [B, max_len, vocab_size]
        return out

    def generate(self, x, lengths, total_length=None):
        out = self.forward(x, lengths, total_length)

        preds = torch.argmax(out, dim=-1)  # [B, T], 가장 큰 점수 클래스 선택
        external = preds - 2                  # 내부 표현 → 외부 표현
        external[external == -2] = 0          # PAD 처리
        return external