### 使用 Transformer 模型进行音乐生成的方法 #### 1. 数据预处理 为了使音乐数据适应于Transformer模型，需先执行一系列的数据预处理操作。具体来说，要从 MIDI 文件中抽取音符和音高信息，并转化为适合模型解析的形式[^1]。 ```python import pretty_midi def preprocess_midi(file_path): midi_data = pretty_midi.PrettyMIDI(file_path) notes = [] for instrument in midi_data.instruments: for note in instrument.notes: start_time = note.start end_time = note.end pitch = note.pitch # 将音符信息整理为元组形式 (起始时间, 结束时间, 音高) notes.append((start_time, end_time, pitch)) return sorted(notes) midi_file = 'example.mid' processed_notes = preprocess_midi(midi_file) print(processed_notes[:5]) # 打印前五个音符的信息 ``` #### 2. 构建 Transformer 模型架构 构建用于音乐生成的Transformer模型时，会特别关注如何表示音符及其持续时间这两个要素。此过程涉及调整标准的Transformer框架以更好地捕捉音乐特性[^2]。 ```python import torch.nn as nn class MusicTransformer(nn.Module): def __init__(self, vocab_size, d_model=512, nhead=8, num_encoder_layers=6, num_decoder_layers=6, dim_feedforward=2048, dropout=0.1): super(MusicTransformer, self).__init__() self.embedding = nn.Embedding(vocab_size, d_model) self.transformer = nn.Transformer(d_model=d_model, nhead=nhead, num_encoder_layers=num_encoder_layers, num_decoder_layers=num_decoder_layers, dim_feedforward=dim_feedforward, dropout=dropout) self.fc_out = nn.Linear(d_model, vocab_size) def forward(self, src, tgt): embedded_src = self.embedding(src) embedded_tgt = self.embedding(tgt) output = self.transformer(embedded_src, embedded_tgt) prediction = self.fc_out(output) return prediction ``` #### 3. 训练阶段 一旦完成了上述准备工作，则可着手准备训练集并启动训练流程。这一步骤旨在让模型学会模仿给定风格或类型的乐曲创作模式。 ```python from torch.utils.data import DataLoader, Dataset # 假设已经定义了一个名为MusicDataset的数据集类 train_dataset = MusicDataset(training_files_list) train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True) model = MusicTransformer(vocab_size=len(note_to_index)) optimizer = torch.optim.Adam(model.parameters(), lr=0.001) criterion = nn.CrossEntropyLoss() for epoch in range(num_epochs): model.train() total_loss = 0 for i, (src, tgt) in enumerate(train_loader): optimizer.zero_grad() output = model(src, tgt[:-1]) loss = criterion(output.view(-1, output.size(-1)), tgt[1:].reshape(-1)) loss.backward() optimizer.step() total_loss += loss.item() avg_loss = total_loss / len(train_loader) print(f'Epoch [{epoch+1}/{num_epochs}], Loss: {avg_loss:.4f}') ``` #### 4. 后处理与生成新作品 当模型经过充分训练之后，就可以尝试让它创造全新的旋律片段了。最后还需将这些抽象化的数值还原成实际听得见的声音文件。 ```python generated_sequence = [] with torch.no_grad(): current_note = initial_seed_tensor # 初始种子张量 for _ in range(max_length_of_generated_music): next_prediction = model(current_note.unsqueeze(0), None).argmax(dim=-1)[-1].item() generated_sequence.append(next_prediction) current_note = torch.cat([current_note, torch.tensor([[next_prediction]])], dim=0)[1:] output_midi = pretty_midi.PrettyMIDI() piano_program = pretty_midi.instrument_name_to_program('Acoustic Grand Piano') piano = pretty_midi.Instrument(program=piano_program) for idx, note_number in enumerate(generated_sequence): piano.notes.append(pretty_midi.Note( velocity=100, pitch=int(note_number), start=(idx * duration_per_step), end=((idx + 1) * duration_per_step))) output_midi.instruments.append(piano) output_midi.write('generated_music.mid') ```