• 李宏毅-机器学习hw4-self-attention结构-辨别600个speaker的身份

    一、慢慢分析+学习pytorch中的各个模块的参数含义、使用方法、功能：
    1.encoder编码器中的nhead参数：

    self.encoder_layer = nn.TransformerEncoderLayer( d_model=d_model, dim_feedforward=256, nhead=2)

    所以说，这个nhead的意思，就是有window窗口的大小，也就是一个b由几个a得到

    2.tensor.permute改变维度的用法示例：
    1. #尝试使用permute函数进行测试:可以通过tensor张量直接调用
    2. import torch
    3. import numpy as np
    4. x = np.array([[[1,1,1],[2,2,2]],[[3,3,3],[4,4,4]]])
    5. y = torch.tensor(x)
    6. #y.shape
    7. z=y.permute(2,1,0)
    8. z.shape
    9. print(z) #permute之后变成了3*2*2的维度
    10. print(y) #本来是一个2*2*3从外到内的维度
    3.tensor.mean求均值：从1个向量 到 1个数值:

    4.python中字典（映射）的使用：

    二、model的neural network设计部分：
    1. import torch
    2. import torch.nn as nn
    3. import torch.nn.functional as F
    4.  
    5.  
    6. class Classifier(nn.Module):
    7. 	def __init__(self, d_model=80, n_spks=600, dropout=0.1):
    8. 		super().__init__()
    9. 		# Project the dimension of features from that of input into d_model.
    10. 		self.prenet = nn.Linear(40, d_model) #通过一个线性的输入层，从40个维度，变成d_model个
    11.  
    12. 		#展示不需要使用这个conformer进行实验
    13. 		# TODO:
    14. 		#   Change Transformer to Conformer. 
    15. 		#   https://arxiv.org/abs/2005.08100
    16. 		#这里是不需要自己设计 self-attention层的，因为transformer的encoder层用到self-attention层
    17. 		self.encoder_layer = nn.TransformerEncoderLayer( 
    18. 			d_model=d_model, dim_feedforward=256, nhead=2 #输入维度是上面的d_model,输出维度是256，这2个nhead是啥？一个b由几个a得到
    19. 		)
    20. 	  #下面这个暂时用不到
    21. 		# self.encoder = nn.TransformerEncoder(self.encoder_layer, num_layers=2)
    22.  
    23. 		# Project the the dimension of features from d_model into speaker nums.
    24. 		#predict_layer
    25. 		self.pred_layer = nn.Sequential( #这里其实就相当于是一个线性输出层了,最终输出的是一个n_soks维度600的向量
    26. 			nn.Linear(d_model, d_model),
    27. 			nn.ReLU(),
    28. 			nn.Linear(d_model, n_spks),
    29. 		)
    30.  
    31. 	def forward(self, mels):
    32. 		"""
    33. 		args:
    34. 			mels: (batch size, length, 40) #我来试图解释一下这个东西,反正就是一段声音信号处理后得到的3维tensor，最里面那一维是40
    35. 		return:
    36. 			out: (batch size, n_spks) #最后只要输出每个batch中的行数 + 每一行中的n_spks的数值
    37. 		"""
    38. 		# out: (batch size, length, d_model) #原来out设置的3个维度的数据分别是batchsize , 
    39. 		out = self.prenet(mels) #通过一个prenet层之后，最里面的那一维空间 就变成了一个d_model维度
    40. 		# out: (length, batch size, d_model)
    41. 		out = out.permute(1, 0, 2) #利用permute将0维和1维进行交换
    42. 		# The encoder layer expect features in the shape of (length, batch size, d_model).
    43. 		out = self.encoder_layer(out)
    44. 		# out: (batch size, length, d_model)
    45. 		out = out.transpose(0, 1) #重新得到原来的维度，这次用transpose和上一次用permute没有区别
    46. 		# mean pooling
    47. 		stats = out.mean(dim=1) #对维度1（第二个维度）计算均值，也就是将整个向量空间-->转成1个数值
    48. 		#得到的是batch,d_model (len就是一行的数据，从这一行中取均值,就是所谓的均值池化)
    49.  
    50. 		# out: (batch, n_spks)
    51. 		out = self.pred_layer(stats) #这里得到n_spks还不是one-hot vec
    52. 		return out
    三、warming up 的设计过程：
    1. import math
    2.  
    3. import torch
    4. from torch.optim import Optimizer
    5. from torch.optim.lr_scheduler import LambdaLR
    6.  
    7. #这部分的代码感觉有一点诡异，好像是设计了一个learning rate的warmup过程，算了，之后再回来阅读好了
    8.  
    9.  
    10. def get_cosine_schedule_with_warmup(
    11. 	optimizer: Optimizer,
    12. 	num_warmup_steps: int,
    13. 	num_training_steps: int,
    14. 	num_cycles: float = 0.5,
    15. 	last_epoch: int = -1,
    16. ):
    17. 	"""
    18. 	Create a schedule with a learning rate that decreases following the values of the cosine function between the
    19. 	initial lr set in the optimizer to 0, after a warmup period during which it increases linearly between 0 and the
    20. 	initial lr set in the optimizer.
    22. 	Args:
    23. 		optimizer (:class:`~torch.optim.Optimizer`):
    24. 		The optimizer for which to schedule the learning rate.
    25. 		num_warmup_steps (:obj:`int`):
    26. 		The number of steps for the warmup phase.
    27. 		num_training_steps (:obj:`int`):
    28. 		The total number of training steps.
    29. 		num_cycles (:obj:`float`, `optional`, defaults to 0.5):
    30. 		The number of waves in the cosine schedule (the defaults is to just decrease from the max value to 0
    31. 		following a half-cosine).
    32. 		last_epoch (:obj:`int`, `optional`, defaults to -1):
    33. 		The index of the last epoch when resuming training.
    35. 	Return:
    36. 		:obj:`torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule.
    37. 	"""
    38. 	def lr_lambda(current_step):
    39. 		# Warmup
    40. 		if current_step < num_warmup_steps:
    41. 			return float(current_step) / float(max(1, num_warmup_steps))
    42. 		# decadence
    43. 		progress = float(current_step - num_warmup_steps) / float(
    44. 			max(1, num_training_steps - num_warmup_steps)
    45. 		)
    46. 		return max(
    47. 			0.0, 0.5 * (1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress))
    48. 		)
    49.  
    50. 	return LambdaLR(optimizer, lr_lambda, last_epoch)
    四、train中每个batch进行的处理:
    1. import torch
    2.  
    3. #这里面其实就是原来train部分的代码处理一个batch的操作
    4.  
    5. def model_fn(batch, model, criterion, device): #这个函数的参数是batch数据，model，loss_func,设备
    6. 	"""Forward a batch through the model."""
    7.  
    8. 	mels, labels = batch #获取mels参数 和 labels参数
    9. 	mels = mels.to(device)
    10. 	labels = labels.to(device)
    11.  
    12. 	outs = model(mels) #得到的输出结果
    13.  
    14. 	loss = criterion(outs, labels) #通过和labels进行比较得到loss
    15.  
    16. 	# Get the speaker id with highest probability.
    17. 	preds = outs.argmax(1) #按照列的方向 计算出最大的索引位置
    18. 	# Compute accuracy.
    19. 	accuracy = torch.mean((preds == labels).float()) #通过将preds和labels进行比较得到acc的数值
    20.  
    21. 	return loss, accuracy

    五、validation的处理函数:
    1. from tqdm import tqdm
    2. import torch
    3.  
    4.  
    5. def valid(dataloader, model, criterion, device):  #感觉就是整个validationset中的数据都进行了操作
    6. 	"""Validate on validation set."""
    7.  
    8. 	model.eval() #开启evaluation模式
    9. 	running_loss = 0.0
    10. 	running_accuracy = 0.0
    11. 	pbar = tqdm(total=len(dataloader.dataset), ncols=0, desc="Valid", unit=" uttr") #创建进度条，实现可视化process_bar
    12.  
    13. 	for i, batch in enumerate(dataloader): #下标i，batch数据存到batch中
    14. 		with torch.no_grad(): #先说明不会使用SGD
    15. 			loss, accuracy = model_fn(batch, model, criterion, device) #调用上面定义的batch处理函数得到loss 和 acc
    16. 			running_loss += loss.item()
    17. 			running_accuracy += accuracy.item()
    18.  
    19. 		pbar.update(dataloader.batch_size) #这些处理进度条的内容可以暂时不用管 
    20. 		pbar.set_postfix(
    21. 			loss=f"{running_loss / (i+1):.2f}",
    22. 			accuracy=f"{running_accuracy / (i+1):.2f}",
    23. 		)
    24.  
    25. 	pbar.close()
    26. 	model.train()
    27.  
    28. 	return running_accuracy / len(dataloader) #返回正确率
    六、train的main调用：
    1. from tqdm import tqdm
    2.  
    3. import torch
    4. import torch.nn as nn
    5. from torch.optim import AdamW
    6. from torch.utils.data import DataLoader, random_split
    7.  
    8.  
    9. def parse_args(): #定义一个给config赋值的函数
    10. 	"""arguments"""
    11. 	config = {
    12. 		"data_dir": "./Dataset",
    13. 		"save_path": "model.ckpt",
    14. 		"batch_size": 32,
    15. 		"n_workers": 1, #这个参数太大的时候，我的这个会error
    16. 		"valid_steps": 2000,
    17. 		"warmup_steps": 1000,
    18. 		"save_steps": 10000,
    19. 		"total_steps": 70000,
    20. 	}
    21.  
    22. 	return config
    23.  
    24.  
    25. def main( #可以直接用上面定义那些参数作为这个main里面的参数
    26. 	data_dir,
    27. 	save_path,
    28. 	batch_size,
    29. 	n_workers,
    30. 	valid_steps,
    31. 	warmup_steps,
    32. 	total_steps,
    33. 	save_steps,
    34. ):
    35. 	"""Main function."""
    36. 	device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    37. 	print(f"[Info]: Use {device} now!")
    38.  
    39. 	train_loader, valid_loader, speaker_num = get_dataloader(data_dir, batch_size, n_workers) #获取所需的data,调用get_dataloader函数
    40. 	train_iterator = iter(train_loader) #定义一个train_data的迭代器
    41. 	print(f"[Info]: Finish loading data!",flush = True)
    42.  
    43. 	model = Classifier(n_spks=speaker_num).to(device) #构造一个model的实例
    44. 	criterion = nn.CrossEntropyLoss() #分别构造loss_func 和 optimizer的实例
    45. 	optimizer = AdamW(model.parameters(), lr=1e-3)
    46. 	scheduler = get_cosine_schedule_with_warmup(optimizer, warmup_steps, total_steps) #构造warmup的实例
    47. 	print(f"[Info]: Finish creating model!",flush = True)
    48.  
    49. 	best_accuracy = -1.0
    50. 	best_state_dict = None
    51.  
    52. 	pbar = tqdm(total=valid_steps, ncols=0, desc="Train", unit=" step") #process_bar相关的东西，不用管它
    53.  
    54. 	for step in range(total_steps): #一共需要的步数进行for循环
    55. 		# Get data
    56. 		try:
    57. 			batch = next(train_iterator) #从train_data中获取到下一个batch的数据
    58. 		except StopIteration:
    59. 			train_iterator = iter(train_loader)
    60. 			batch = next(train_iterator)
    61.  
    62. 		loss, accuracy = model_fn(batch, model, criterion, device) #传递对应的数据、模型参数，得到这个batch的loss和acc
    63. 		batch_loss = loss.item()
    64. 		batch_accuracy = accuracy.item()
    65.  
    66. 		# Updata model
    67. 		loss.backward()
    68. 		optimizer.step()
    69. 		scheduler.step()
    70. 		optimizer.zero_grad() #更新进行Gradient descend 更新模型，并且将grad清空
    71.  
    72. 		# Log
    73. 		pbar.update() #process_bar的东西先不管
    74. 		pbar.set_postfix(
    75. 			loss=f"{batch_loss:.2f}",
    76. 			accuracy=f"{batch_accuracy:.2f}",
    77. 			step=step + 1,
    78. 		)
    79.  
    80. 		# Do validation
    81. 		if (step + 1) % valid_steps == 0:
    82. 			pbar.close()
    83.  
    84. 			valid_accuracy = valid(valid_loader, model, criterion, device) #调用valid函数计算这一次validation的正确率
    85.  
    86. 			# keep the best model
    87. 			if valid_accuracy > best_accuracy: #总是保持最好的valid_acc
    88. 				best_accuracy = valid_accuracy
    89. 				best_state_dict = model.state_dict()
    90.  
    91. 			pbar = tqdm(total=valid_steps, ncols=0, desc="Train", unit=" step")
    92.  
    93. 		# Save the best model so far.
    94. 		if (step + 1) % save_steps == 0 and best_state_dict is not None:
    95. 			torch.save(best_state_dict, save_path) #保存最好的model参数
    96. 			pbar.write(f"Step {step + 1}, best model saved. (accuracy={best_accuracy:.4f})")
    97.  
    98. 	pbar.close()
    99.  
    100.  
    101. if __name__ == "__main__": #调用这个main函数
    102. 	main(**parse_args())

    七、inference部分的test内容：

    1. import os
    2. import json
    3. import torch
    4. from pathlib import Path
    5. from torch.utils.data import Dataset
    6.  
    7.  
    8. class InferenceDataset(Dataset):
    9. 	def __init__(self, data_dir):
    10. 		testdata_path = Path(data_dir) / "testdata.json"
    11. 		metadata = json.load(testdata_path.open())
    12. 		self.data_dir = data_dir
    13. 		self.data = metadata["utterances"]
    14.  
    15. 	def __len__(self):
    16. 		return len(self.data)
    17.  
    18. 	def __getitem__(self, index):
    19. 		utterance = self.data[index]
    20. 		feat_path = utterance["feature_path"]
    21. 		mel = torch.load(os.path.join(self.data_dir, feat_path))
    22.  
    23. 		return feat_path, mel
    24.  
    25.  
    26. def inference_collate_batch(batch):
    27. 	"""Collate a batch of data."""
    28. 	feat_paths, mels = zip(*batch)
    29.  
    30. 	return feat_paths, torch.stack(mels)
    1. import json
    2. import csv
    3. from pathlib import Path
    4. from tqdm.notebook import tqdm
    5.  
    6. import torch
    7. from torch.utils.data import DataLoader
    8.  
    9. def parse_args():
    10. 	"""arguments"""
    11. 	config = {
    12. 		"data_dir": "./Dataset",
    13. 		"model_path": "./model.ckpt",
    14. 		"output_path": "./output.csv",
    15. 	}
    16.  
    17. 	return config
    18.  
    19.  
    20. def main(
    21. 	data_dir,
    22. 	model_path,
    23. 	output_path,
    24. ):
    25. 	"""Main function."""
    26. 	device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    27. 	print(f"[Info]: Use {device} now!")
    28.  
    29. 	mapping_path = Path(data_dir) / "mapping.json"
    30. 	mapping = json.load(mapping_path.open())
    31.  
    32. 	dataset = InferenceDataset(data_dir)
    33. 	dataloader = DataLoader(
    34. 		dataset,
    35. 		batch_size=1,
    36. 		shuffle=False,
    37. 		drop_last=False,
    38. 		num_workers=8,
    39. 		collate_fn=inference_collate_batch,
    40. 	)
    41. 	print(f"[Info]: Finish loading data!",flush = True)
    42.  
    43. 	speaker_num = len(mapping["id2speaker"])
    44. 	model = Classifier(n_spks=speaker_num).to(device)
    45. 	model.load_state_dict(torch.load(model_path))
    46. 	model.eval()
    47. 	print(f"[Info]: Finish creating model!",flush = True)
    48.  
    49. 	results = [["Id", "Category"]]
    50. 	for feat_paths, mels in tqdm(dataloader):
    51. 		with torch.no_grad():
    52. 			mels = mels.to(device)
    53. 			outs = model(mels)  #调用model计算得到outs
    54. 			preds = outs.argmax(1).cpu().numpy() #对outs进行argmax,得到的索引存储到preds中
    55. 			for feat_path, pred in zip(feat_paths, preds):
    56. 				results.append([feat_path, mapping["id2speaker"][str(pred)]]) #将每一次的结果存放的到results中
    57.  
    58. 	with open(output_path, 'w', newline='') as csvfile:
    59. 		writer = csv.writer(csvfile)
    60. 		writer.writerows(results)
    61.  
    62.  
    63. if __name__ == "__main__":
    64. 	main(**parse_args())

    inference部分的代码暂时就看看好了，这个2022版本的数据在github上404了。。。

    七、Dataset的处理过程：
    1. import os
    2. import json
    3. import torch
    4. import random
    5. from pathlib import Path
    6. from torch.utils.data import Dataset
    7. from torch.nn.utils.rnn import pad_sequence
    8.  
    9.  
    10. class myDataset(Dataset):
    11. 	def __init__(self, data_dir, segment_len=128):
    12. 		self.data_dir = data_dir
    13. 		self.segment_len = segment_len
    14. 	
    15. 		# Load the mapping from speaker neme to their corresponding id. 
    16. 		mapping_path = Path(data_dir) / "mapping.json"
    17. 		mapping = json.load(mapping_path.open()) #将这个json文件load到变量mapping中
    18. 		self.speaker2id = mapping["speaker2id"] #其实speaker2id这个变量就是mapping里面的内容
    19. 		#其实也就是原来数据集中的"id00464"变成我们这里的600个人的数据集的0-599的id
    20. 	
    21. 		# Load metadata of training data.
    22. 		metadata_path = Path(data_dir) / "metadata.json"
    23. 		metadata = json.load(open(metadata_path))["speakers"]
    24. 		#和上面类似的操作，这里的metadata就是打开那个json文件中的内容
    25. 		#我觉得按照他上课的说法，这里的n_mels的意思就是每个特征音频长度取出40就好了,?对吗
    26. 		#然后，这个json文件里面的内容就是不同speakerid所发声的音频文件的路径和mel_len
    27.  
    28.  
    29. 		# Get the total number of speaker.
    30. 		self.speaker_num = len(metadata.keys())
    31. 		self.data = [] #data就是这个class中的数据
    32. 		for speaker in metadata.keys(): #逐个遍历每个speaker
    33. 			for utterances in metadata[speaker]: #遍历每个speaker的每一段录音
    34. 				self.data.append([utterances["feature_path"], self.speaker2id[speaker]])#将每一段录音按照 (路径，新id)存入data变量中
    35.  
    36. 	def __len__(self):
    37. 			return len(self.data) #返回总共的data数量
    38.  
    39. 	def __getitem__(self, index):
    40. 		feat_path, speaker = self.data[index] #从下标位置获取到该段录音的路径 和 speakerid
    41. 		# Load preprocessed mel-spectrogram.
    42. 		mel = torch.load(os.path.join(self.data_dir, feat_path)) #从路径中获取到该mel录音文件
    43.  
    44. 		# Segmemt mel-spectrogram into "segment_len" frames.
    45. 		if len(mel) > self.segment_len: #如果大于128这个seg , 一些处理....
    46. 			# Randomly get the starting point of the segment.
    47. 			start = random.randint(0, len(mel) - self.segment_len)
    48. 			# Get a segment with "segment_len" frames.
    49. 			mel = torch.FloatTensor(mel[start:start+self.segment_len])
    50. 		else:
    51. 			mel = torch.FloatTensor(mel)
    52. 		# Turn the speaker id into long for computing loss later.
    53. 		speaker = torch.FloatTensor([speaker]).long() #将speakerid转换为long类型 
    54. 		return mel, speaker #返回这个录音mel文件和对应的speakerid
    55.  
    56. 	def get_speaker_number(self):
    57. 		return self.speaker_num

    这里附带我下载的文件资源路径:

    ML2022Spring-hw4 | Kaggle

    下面dropbox的链接是可以使用的

    1. !wget https://www.dropbox.com/s/vw324newiku0sz0/Dataset.tar.gz.aa?d1=0
    2. !wget https://www.dropbox.com/s/vw324newiku0sz0/Dataset.tar.gz.aa?d1=0 
    3. !wget https://www.dropbox.com/s/z840g69e71nkayo/Dataset.tar.gz.ab?d1=0 
    4. !wget https://www.dropbox.com/s/h1081e1ggonio81/Dataset.tar.gz.ac?d1=0 
    5. !wget https://www.dropbox.com/s/fh3zd8ow668c4th/Dataset.tar.gz.ad?d1=0 
    6. !wget https://www.dropbox.com/s/ydzygoy2pv6gw9d/Dataset.tar.gz.ae?d1=0
    7. !cat Dataset.tar.gz.*  |  tar zxvf  -

    这样才能下载到你需要的数据

    怎么说呢？最后的最后，还是这个dropbox中下载的内容不全，少了一些文件

    有一个解决的方法是，直接在kaggle上面下载那个5.2GB的压缩包，不过解压之后可能有70GB,文件似乎太大了，而且下载之后，只要全部解压导入到Dataset文件夹就可以运行了

    方法三：尝试一下那个GoogleDrive上面的文件 :

    失败了，算了还是自己老老实实下载然后上传吧

    1. !gdown --id '1CtHZhJ-mTpNsO-MqvAPIi4Yrt3oSBXYV' --output Dataset.zip
    2. !gdown --id '14hmoMgB1fe6v50biIceKyndyeYABGrRq' --output Dataset.zip 
    3. !gdown --id '1e9x-Pj13n7-9tK9LS_WjiMo21ru4UBH9' --output Dataset.zip 
    4. !gdown --id '10TC0g46bcAz_jkiM165zNmwttT4RiRgY' --output Dataset.zip 
    5. !gdown --id '1MUGBvG_Jjq00C2JYHuyV3B01vaf1kWIm' --output Dataset.zip 
    6. !gdown --id '18M91P5DHwILNy01ssZ57AiPOR0OwutOM' --output Dataset.zip 
    7. !unzip Dataset.zip

  • 相关阅读:
    log4j日志漏洞问题
    WSL2快速上手
    人工智能笔记
    C++入门基础05：表达式(表达式基础、算术运算符与赋值运算符、逻辑关系运算符、成员访问运算符与条件运算符、位运算符、移位运算符与类型转换)
    Ubuntu20.04安装Ipopt的流程介绍及报错解决方法（亲测简单有效）
    界面控件DevExpress WPF Scheduler控件 - 如何实现数据的按需加载？
    IMU标定之---Allan方差
    WPF项目-按着键盘方向键，移动格子盒子效果
    逍遥自在学C语言 | 指针的基础用法
    《隐私计算简易速速上手小册》第7章：隐私计算与云计算/边缘计算（2024 最新版）
  • 原文地址：https://blog.csdn.net/xiao_ZHEDA/article/details/132715083