训练流程
环境设置：
- 训练在Google Colab上完成，利用其免费提供的16G显存（实际可用约13～15G）。
数据预处理：
- 使用BPE（Byte Pair Encoding）算法处理词表，词表大小约为5万。
- 对每篇文章mask最后一个字用作预测，并计算loss。
数据集定义：
- 定义了
MyDataSet类，用于处理数据集，包括数据的获取、解码输入和输出的生成，以及数据的填充。
- 定义了
注意力机制：
- 实现了
get_attn_pad_mask和get_attn_subsequence_mask函数，用于在注意力计算中忽略padding和序列顺序。 - 定义了
ScaledDotProductAttention和MultiHeadAttention类，用于计算Q和K的相似度矩阵，并应用多头注意力机制。
- 实现了
前馈网络：
- 实现了
PoswiseFeedForwardNet类，用于在解码器层中进行前馈网络计算。
- 实现了
解码器层和解码器：
- 定义了
DecoderLayer和Decoder类，用于构建解码器层和整个解码器。
- 定义了
GPT模型：
- 定义了
GPT类，包含解码器和投影层，用于将解码器的输出映射到词汇表大小。
- 定义了
模型训练：
- 设置了批处理大小、训练轮数和学习率。
- 实现了
train_step函数，用于执行单个epoch的训练。 - 实现了
train函数，用于执行整个训练过程，并在每个epoch后保存模型权重。
模型生成：
GPT类中的generate方法用于生成新的文本序列。
训练辅助函数：
- 提供了
epoch_time、print_num_parameters、split_array、get_dataset_from_mk和get_dataset_from_json等辅助函数，用于训练过程中的时间计算、参数统计、数据分割和数据集加载。
- 提供了
训练执行：
- 在主函数中，初始化数据集、数据加载器、模型，并执行训练过程。
模型调优和推理：
- 文章最后提到，尽管代码已经跑通，但效果尚未调优。作者认为可能需要在通用语言文本上进行预训练，然后再在垂直领域数据上进行微调，以提高模型在家用CPU上的推理能力。
整个流程涵盖了从数据预处理到模型训练、保存和生成文本的完整步骤，是一个典型的自然语言处理模型训练流程。
代码部分
# -*- coding: utf-8 -*-
"""gpt-tiny.ipynb
Automatically generated by Colab.
Original file is located at
https://colab.research.google.com/drive/1zSuS4zBqsIh7BqBv9eA8CZYcdR0RO6Lw
"""
!nvidia-smi
!ls -lh /content/
!mkdir model
"""## 制作词表"""
# 1. 遍历文件夹，制作词表
data_list = ["<pad>", "<unk>", "<sep>", "<cls>"]
data_set = set(data_list)
import glob
import json
with open("data.json") as f:
data = json.load(f)
for lines in data:
for word in lines:
if word not in data_set:
data_set.add(word)
data_list.append(word)
print("vocab_size", len(data_set))
import json
with open("vec.json", 'w') as f:
json.dump(data_list, f, ensure_ascii=False)
"""## 超参数配置"""
import json
import torch
max_pos = 1024 # 一段话最多字
d_model = 768 # Embedding Size
d_ff = 2048 # FeedForward dimension
d_k = d_v = 64 # dimension of K(=Q), V
n_layers = 6 # number of Encoder of Decoder Layer
n_heads = 8 # number of heads in Multi-Head Attention
device = "cuda" if torch.cuda.is_available() else "cpu"
vocab_data = {}
f = json.load(open("vec.json"))
for i, v in enumerate(f):
vocab_data[v] = i
vocab_data_reverse = {}
for k, v in vocab_data.items():
vocab_data_reverse[v] = k
vocab_size = len(f) # 词典大小
special_char_pad = 0
special_char_sep = 2
def encoder(text):
ret = []
for x in text:
ret.append(vocab_data[x])
return ret
def decoder(encoder_list):
ret = ""
for x in encoder_list:
ret += vocab_data_reverse[x]
return ret
"""## 模型层
"""
# 构建数据集和模型
import random
import numpy as np
import torch
import torch.utils.data as Data
from torch import nn
import torch.nn.functional as F
CLIP = 1
# 定义数据集
class MyDataSet(Data.Dataset):
def __init__(self, datas):
self.datas = datas
def __getitem__(self, item):
data_item = self.datas[item]
decoder_input = data_item[:-1]
decoder_output = data_item[1:]
return {"decoder_input": decoder_input,
"decoder_output": decoder_output}
def padding_batch(self, batch): #
for d in batch: # 对当前batch的每一个decoder_input和decoder_output数据填充"<pad>"，填充到和batch里面的有的最大长度为止
input_len = len(d["decoder_input"])
output_len = len(d["decoder_output"])
d["decoder_input"].extend([special_char_pad] * (max_pos - input_len))
d["decoder_output"].extend([special_char_pad] * (max_pos - output_len))
decoder_inputs = torch.tensor([d["decoder_input"] for d in batch], dtype=torch.long) # 转type
decoder_outputs = torch.tensor([d["decoder_output"] for d in batch], dtype=torch.long)
return decoder_inputs, decoder_outputs # 形状[b,decoder_input_maxlen], [b,decoder_output_maxlen] type为torch.long
def __len__(self):
return len(self.datas)
"""
======================================================================================================================================================================
下面是模型构建
"""
# 把数据里面<pad>对应的字符给mask掉，让后面Q和K相似度矩阵的softmax中这些pad都为0，就不会被后续的V考虑
def get_attn_pad_mask(seq_q, seq_k): # 形状都是[b, tgt_len <300]
batch_size, len_q = seq_q.size() # len_q = len_k = tgt_len
batch_size, len_k = seq_k.size()
# eq(zero) is PAD token.就是把数据里面<pad>对应的字符给mask掉，让后面Q和K的softmax不考虑这些<pad>
pad_attn_mask = seq_k.data.eq(0).unsqueeze(
1) # [b, 1, tgt_len], id为0(也就是<pad>的id)的位置为True，其他位置为False。后面会把Ture位置的mask掉
return pad_attn_mask.expand(batch_size, len_q, len_k) # [b, tgt_len, tgt_len]
def get_attn_subsequence_mask(seq): # seq: [b, tgt_len]
attn_shape = [seq.size(0), seq.size(1), seq.size(1)] # [b, tgt_len, tgt_len]
subsequence_mask = np.triu(np.ones(attn_shape), k=1) # Upper triangular matrix(上三角矩阵)
subsequence_mask = torch.from_numpy(subsequence_mask).byte()
subsequence_mask = subsequence_mask.to(device)
return subsequence_mask # [b, tgt_len, tgt_len] 上三角矩阵,下0上1,dtype=torch.uint8
class ScaledDotProductAttention(nn.Module): # 计算Q和K的相似度矩阵，然后乘V。对应笔记里的图
def __init__(self):
super(ScaledDotProductAttention, self).__init__()
def forward(self, Q, K, V,
attn_mask): # 前三者形状相同[b, n_heads, tgt_len, d_k=64]，attn_mask:[b, n_heads, tgt_len, tgt_len]
scores = torch.matmul(Q, K.transpose(-1, -2)) / np.sqrt(d_k) # Q和K的相似度矩阵scores : [b, n_heads, tgt_len, tgt_len]
scores.masked_fill_(attn_mask, -1e9) # Fills elements of self tensor with value where mask is True.
# 就是scores矩阵里面和attn_mask=1对应位置的元素全部替换成-1e9，使其在下一步的softmax中变为0
attn = nn.Softmax(dim=-1)(scores) # [b, n_heads, tgt_len, tgt_len]
context = torch.matmul(attn, V) # [b, n_heads, tgt_len, d_v]
return context, attn
class MultiHeadAttention(nn.Module): # 多头注意力机制
def __init__(self):
super(MultiHeadAttention, self).__init__()
self.W_Q = nn.Linear(d_model, d_k * n_heads, bias=False) # d_model=768 , d_v = d_k = 64 , n_heads=8
self.W_K = nn.Linear(d_model, d_k * n_heads, bias=False)
self.W_V = nn.Linear(d_model, d_v * n_heads, bias=False)
self.fc = nn.Linear(n_heads * d_v, d_model, bias=False)
self.layernorm = nn.LayerNorm(d_model)
def forward(self, input_Q, input_K, input_V,
attn_mask): # 前三者形状相同，都是[b, tgt_len, d_model] , attn_mask: [b, tgt_len, tgt_len]
residual, batch_size = input_Q, input_Q.size(0) #
# [b, tgt_len, d_model] --> [b, tgt_len, d_k * n_heads] -split-> (b, tgt_len, n_heads, d_k) -trans-> (b, n_heads, tgt_len, d_k)
Q = self.W_Q(input_Q).view(batch_size, -1, n_heads, d_k).transpose(1, 2) # Q: [b, n_heads, tgt_len, d_k=64]
K = self.W_K(input_K).view(batch_size, -1, n_heads, d_k).transpose(1, 2) # K: [b, n_heads, tgt_len, d_k=64]
V = self.W_V(input_V).view(batch_size, -1, n_heads, d_v).transpose(1, 2) # V: [b, n_heads, tgt_len, d_v=64]
attn_mask = attn_mask.unsqueeze(1).repeat(1, n_heads, 1,
1) # 添加n_heads维度并复制。attn_mask : [b, n_heads, tgt_len, tgt_len]
context, attn = ScaledDotProductAttention()(Q, K, V, attn_mask) # 参考图解，context形状[b, n_heads, tgt_len, d_v]
context = context.transpose(1, 2).reshape(batch_size, -1, n_heads * d_v) # context: [b, tgt_len, n_heads * d_v]
output = self.fc(context) # [batch_size, tgt_len, d_model]
return self.layernorm(output + residual), attn
class PoswiseFeedForwardNet(nn.Module): # [b,tgt_len,d_model] -> [b,tgt_len,d_model] 输入和输出形状不变
def __init__(self):
super(PoswiseFeedForwardNet, self).__init__()
self.fc = nn.Sequential(
nn.Linear(d_model, d_ff, bias=False),
nn.ReLU(),
nn.Linear(d_ff, d_model, bias=False)
)
self.layernorm = nn.LayerNorm(d_model)
def forward(self, inputs):
'''
inputs: [batch_size, seq_len, d_model]
'''
residual = inputs
output = self.fc(inputs)
return self.layernorm(output + residual) # [batch_size, seq_len, d_model]
class DecoderLayer(nn.Module):
def __init__(self):
super(DecoderLayer, self).__init__()
self.dec_self_attn = MultiHeadAttention() # 多头注意力
# self.dec_enc_attn = MultiHeadAttention()
self.pos_ffn = PoswiseFeedForwardNet()
def forward(self, dec_inputs,
dec_self_attn_mask): # dec_inputs: [b, tgt_len, d_model] dec_self_attn_mask: [b, tgt_len, tgt_len]
# dec_outputs: [b, tgt_len, d_model], dec_self_attn: [b, n_heads, tgt_len, tgt_len]
dec_outputs, dec_self_attn = self.dec_self_attn(dec_inputs, dec_inputs, dec_inputs, dec_self_attn_mask)
dec_outputs = self.pos_ffn(dec_outputs) # [b, tgt_len, d_model]
return dec_outputs, dec_self_attn # [b, tgt_len, d_model] , [b, n_heads, tgt_len, tgt_len]
class Decoder(nn.Module):
def __init__(self):
super(Decoder, self).__init__()
self.tgt_emb = nn.Embedding(vocab_size,
d_model) # 以矩阵形式抽取一行，会比直接用mlp高效。因为mlp会多很多无用运算 emb矩阵形状(vocab_size,768)
self.pos_emb = nn.Embedding(max_pos, d_model) # 可学习的位置编码 emb矩阵形状(300,768)
self.layers = nn.ModuleList([DecoderLayer() for _ in range(n_layers)])
def forward(self, dec_inputs): # 输入dec_inputs形状[b,tgt_len]
seq_len = dec_inputs.size(1) # tgt_len ，表示batch内最大长度，不会超过300
pos = torch.arange(seq_len, dtype=torch.long, device=device) # 给位编码准备的值，[0,1,2,3,...,seq_len-1]
pos = pos.unsqueeze(0).expand_as(dec_inputs) # [tgt_len] -> [b, tgt_len]
dec_outputs = self.tgt_emb(dec_inputs) + self.pos_emb(pos) # [b, tgt_len, d_model=768]
dec_self_attn_pad_mask = get_attn_pad_mask(dec_inputs, dec_inputs) # [b, tgt_len, tgt_len] 把<pad>给mask掉
dec_self_attn_subsequence_mask = get_attn_subsequence_mask(dec_inputs) # [b, tgt_len, tgt_len] 上三角矩阵
dec_self_attn_mask = torch.gt((dec_self_attn_pad_mask + dec_self_attn_subsequence_mask),
0) # [b, tgt_len, tgt_len] 矩阵大于0的全为1，否则为0
dec_self_attns = []
for layer in self.layers:
# dec_outputs: [b, tgt_len, d_model], dec_self_attn: [b, n_heads, tgt_len, tgt_len], dec_enc_attn: [b, h_heads, tgt_len, src_len]
dec_outputs, dec_self_attn = layer(dec_outputs, dec_self_attn_mask)
dec_self_attns.append(dec_self_attn)
return dec_outputs, dec_self_attns
class GPT(nn.Module):
def __init__(self):
super(GPT, self).__init__()
self.decoder = Decoder()
self.projection = nn.Linear(d_model, vocab_size) # 768->vocab_size,也就是把最后的隐藏层节点768投影到字典个数的节点上
def forward(self, dec_inputs): # 输入dec_inputs形状[b,tgt_len] tgt_len<=300 (tgt_len是batch内最大长度)
dec_outputs, dec_self_attns = self.decoder(
dec_inputs) # dec_outpus: [b, tgt_len, d_model=768], dec_self_attns: [n_layers, b, n_heads, tgt_len, tgt_len]
dec_logits = self.projection(dec_outputs) # dec_logits: [b, tgt_len, vocab_size]
return dec_logits.view(-1, dec_logits.size(-1)), dec_self_attns # 左边那个输出形状[b *tgt_len,vocab_size]
@torch.no_grad()
def generate(self, sentence, max_new_tokens, temperature=1.0, top_k=None):
"""
Take a conditioning sequence of indices idx (LongTensor of shape (b,t)) and complete
the sequence max_new_tokens times, feeding the predictions back into the model each time.
Most likely you'll want to make sure to be in model.eval() mode of operation for this.
"""
idx = torch.tensor(encoder(sentence), dtype=torch.long, device=device).unsqueeze(
0) # [n] -> [1,n] 转type，并放入指定设备
for _ in range(max_new_tokens):
# forward the model to get the logits for the index in the sequence
dec_outputs, _ = self.decoder(idx)
logits = self.projection(dec_outputs) # [1, tgt_len, vocab_size]
# pluck the logits at the final step and scale by desired temperature
logits = logits[:, -1, :] / temperature
# optionally crop the logits to only the top k options
if top_k is not None:
vv, _ = torch.topk(logits, min(top_k, logits.size(-1)))
logits[logits < vv[:, [-1]]] = -float('Inf')
# apply softmax to convert logits to (normalized) probabilities
probs = F.softmax(logits, dim=-1)
# sample from the distribution
# idx_next = torch.multinomial(probs, num_samples=1)
idx_next = torch.max(probs, dim=-1, keepdim=True)[1]
# append sampled index to the running sequence and continue
if idx_next.item() == special_char_sep:
break
idx = torch.cat(
[idx.detach(), idx_next], -1)
yield vocab_data_reverse[idx_next.item()]
"""
## 训练"""
# 模型的训练
import glob
import math
import time
from torch import optim
from tqdm import tqdm
def epoch_time(start_time, end_time): # 把秒数表示为分钟和秒
elapsed_time = end_time - start_time
elapsed_mins = int(elapsed_time / 60)
elapsed_secs = int(elapsed_time - (elapsed_mins * 60))
return elapsed_mins, elapsed_secs
def train_step(model, data_loader, optimizer, criterion, clip=1, print_every=None): # 每一个eopch的训练
model.train() # 训练模式
if print_every == 0:
print_every = 1
print_loss_total = 0 # 每次打印都重置，统计一定batch数内(默认10)的loss，每10个batch打印一次
epoch_loss = 0 # epoch的总loss
for i, (dec_inputs, dec_outputs) in enumerate(
tqdm(data_loader)): # dec_inputs: [b, tgt_len] , dec_outputs: [b, tgt_len]
optimizer.zero_grad()
dec_inputs, dec_outputs = dec_inputs.to(device), dec_outputs.to(device)
# outputs: [batch_size * tgt_len, tgt_vocab_size] tgt_len<=30
# with torch.cuda.amp.autocast(): # 半精度训练
outputs, dec_self_attns = model(dec_inputs)
loss = criterion(outputs, dec_outputs.view(
-1)) # outputs :(b * tgt_len, vocab_size),dec_outputs.view(-1) :(b * tgt_len) tgt_len<=300
print_loss_total += loss.item()
epoch_loss += loss.item()
loss.backward() # 梯度反向传播
# 梯度裁剪，防止梯度爆炸。如果loss超过clip，将梯度值缩小为原来的(loss/clip)分之一
torch.nn.utils.clip_grad_norm_(model.parameters(), clip)
optimizer.step() # 更新模型权重
if print_every and (i + 1) % print_every == 0:
print_loss_avg = print_loss_total / print_every
print_loss_total = 0
print('\tCurrent Loss: %.4f' % print_loss_avg)
return epoch_loss / len(data_loader)
def train(model, data_loader, lr):
criterion = nn.CrossEntropyLoss(ignore_index=0).to(device) # 损失函数
optimizer = optim.AdamW(model.parameters(), lr=lr) # 优化器
for epoch in range(epochs):
start_time = time.time()
train_loss = train_step(model, data_loader, optimizer, criterion, CLIP, print_every=100) # 训练一个epoch
end_time = time.time()
torch.save(model.state_dict(), r'model/GPT-%d.pt' % epoch) # 保存模型权重
epoch_mins, epoch_secs = epoch_time(start_time, end_time) # 把秒数表示为分钟和秒
print(f'Epoch: {epoch + 1:02} | Time: {epoch_mins}m {epoch_secs}s')
print(f'\tTrain Loss: {train_loss:.3f}')
def print_num_parameters(model):
# Find total parameters and trainable parameters
total_params = sum(p.numel() for p in model.parameters())
print("number of parameters: %.2fM" % (total_params / 1e6,))
total_trainable_params = sum(
p.numel() for p in model.parameters() if p.requires_grad)
print("train of parameters: %.2fM" % (total_trainable_params / 1e6))
def split_array(array, num):
length = len(array)
chunk_size = math.ceil(length / num)
result = []
for i in range(0, chunk_size):
result.append(array[:num])
array = array[num:]
return result
def get_dataset_from_mk(folder):
dataset = []
for filename in glob.glob(folder + "*.md"):
with open(filename) as f:
data = f.read()
array_ = split_array(encoder(data), max_pos)
dataset.extend(array_)
return dataset
def get_dataset_from_json(filename):
with open(filename) as f:
data = json.load(f)
dataset = []
for item in data:
dataset.append(encoder(item))
return dataset
if __name__ == '__main__':
batch_size = 16
epochs = 10
shuffle = True
lr = 1e-4
filename = "data.json"
dataset = get_dataset_from_json(filename)
data_set = MyDataSet(dataset)
data_loader = Data.DataLoader(data_set,
batch_size=batch_size,
collate_fn=data_set.padding_batch,
shuffle=shuffle) # 对每个batch单独调用collate_fn处理，因为batch内的句子长短不一，不能直接用torch的默认方法
model = GPT().to(device)
print_num_parameters(model)
train(model, data_loader, lr)
"""## **测试推理**"""
import torch
model = GPT().to(device)
model.load_state_dict(torch.load('model/GPT-3.pt', map_location=device))
model.eval() # 推理模式
context = "java反序列"
for chunk in model.generate(context, 500, 0.8,1):
print(chunk, end="")