chatGPT学习---Transformer代码实现1
标签: chatGPT学习---Transformer代码实现1 JavaScript博客 51CTO博客
2023-04-01 18:23:42 101浏览
这里写自定义目录标题
- 1. 创建词汇表
- 2. 创建数据集
- 3. Bigram语言模型
- 4. 代码生成
- 5. 网络训练
为了更好的理解Transformer的概念,我们可以自己动手来实现一个小型的Transformer。在这里,我们以最近大火的能写代码的chatGPT为例,自己动手写一个能写代码的小型Transformer。这部分内容大部分内存来自于Karpathy的2小时教程,主要的改动在于将他在实现细节中一些跟大家典型习惯不符合的地方,改为我们更习惯的方式。
1. 创建词汇表
我们首先需要创建一个词汇表,我在这里将我写的一个基于Yolov8的3D检测模型的所有源码,全部写到一个文本文件中,将其作为我们训练数据,这个文件大小为797K,看来Yolov8的代码量还是不小的。
在这里与Karpathy的教程中略有不同,我们把词汇表记录到AppRegistry类中,文件的组织形式为:
hwcgpt
|---core
| |---app_registry.py
|---dss
| |---yolo3d_ds.py
|---app_main.py
载入词汇表:
# dss/yolo3d_ds.py
class Yolo3dDs(Dataset):
def __init__(self):
super(Yolo3dDs, self).__init__()
@staticmethod
def generate_vocab(ds_rfn: str) -> None:
with open(ds_rfn, 'r', encoding='utf-8') as rfd:
text = rfd.read()
AppRegistry.chars = sorted(list(set(text)))
AppRegistry.vocab_size = len(AppRegistry.chars)
print(f'词汇表大小:{AppRegistry.vocab_size}')
print('词汇表:')
print(''.join(AppRegistry.chars))
AppRegistry.stoi = { ch: i for i, ch in enumerate(AppRegistry.chars) }
AppRegistry.itos = { i: ch for i, ch in enumerate(AppRegistry.chars) }
AppRegistry.encode = lambda s: [AppRegistry.stoi[c] for c in s]
AppRegistry.decode = lambda l: ''.join([AppRegistry.itos[i] for i in l])
print(AppRegistry.encode('def main(arg={}):'))
print(AppRegistry.decode(AppRegistry.encode('def main(arg={}):')))
执行的结果如下所示:
2. 创建数据集
本部分代码:
git clone https://gitee.com/yt7589/hwcgpt.git
cd hwcgpt
git checkout v0.0.2
我们假定用90%的数据作为训练集,10%的数据作为测试集,每次输入模型的长度为block_size=8,我们的任务是根据这8个单词,预测出第9个单词。
我们的数据还是yolo3d.txt,这时我们一次读入block_size+1=9个字符,前8个字符作为模型的输入,第9个字符作为模型预测输出的真值。
我们假设文本为:
REQUIREMENTS = [f'{x.name}{x.specifier}' for x in pkg.parse_requirements((PARENT / 'requirements.txt').read_text())]
# 样本数据
X: [R, E, Q, U, I, R, E, M]
y: [E, Q, U, I, R, E, M, E]
我们生成的样本为:
序号 |
X |
y |
1 |
R |
E |
2 |
[R, E] |
[Q] |
3 |
[R, E, Q] |
[U] |
4 |
[R, E, Q, U] |
[I] |
5 |
[R, E, Q, U, I] |
[R] |
6 |
[R, E, Q, U, I, R] |
[E] |
7 |
{R, E, Q, U, I, R, E] |
[M] |
8 |
[R, E, Q, U, I, R, E, M] |
[E] |
程序代码如下所示: |
class Yolo3dDs(Dataset):
def __init__(self, ds_rfn: str):
super(Yolo3dDs, self).__init__()
with open(ds_rfn, 'r', encoding='utf-8') as rfd:
text = rfd.read()
data = torch.tensor(AppRegistry.encode(text), dtype=torch.long)
# 90%作为训练集,10%作为测试集
n = int(0.9*len(data))
self.train_data = data[:n]
self.val_data = data[n:]
def disp_first_sample(self):
# 取出第一个样本
X = self.train_data[:AppRegistry.block_size]
y = self.train_data[1:AppRegistry.block_size+1]
for t in range(AppRegistry.block_size):
context = X[:t+1]
target = y[t]
print(f'X: {context} => {target}')
# app_main.py
def main(args={}):
print('最简GPT代码生成器 v0.0.1')
ds_rfn = 'datasets/yolo3d.txt'
Yolo3dDs.generate_vocab(ds_rfn=ds_rfn)
ds = Yolo3dDs(ds_rfn=ds_rfn)
ds.disp_first_sample()
运行结果为:
通常在实际应用中,我们都是以一个批次的形式进行训练,所以我们需要一次读入一个批次,我们设批次大小为batch=4,获取一个batch的代码如下所示:
def disp_batch(self, Xb, yb):
print(f'Xb: {Xb.shape}, yb: {yb.shape};')
for b in range(AppRegistry.batch_size):
print(f'batch={b}:')
for t in range(AppRegistry.block_size):
context = Xb[b, :t+1]
target = yb[b, t]
print(f' {context} => {target};')
def get_batch(self, mode: str='train') -> Tuple[torch.Tensor, torch.Tensor]:
data = self.train_data if mode=='train' else self.val_data
# 生成0至len(data)-AppRegistry.block_size-1之间随机数,共生成batch_size个
idxs = torch.randint(len(data)-AppRegistry.block_size, (AppRegistry.batch_size,))
X = torch.stack([data[i:i+AppRegistry.block_size] for i in idxs])
y = torch.stack([data[i+1:i+AppRegistry.block_size+1] for i in idxs])
return X, y
3. Bigram语言模型
下面我们建一个语言模型,我们看到一个单词,根据出现概率预测下一个单词,这个就是Bigramm语言模型。当然,如果我们看到更多的单词,预测下一个单词一定会更准确,但是这个就会比较复杂,由于我们是一个Hello World级别的例子,所以我们只考虑Bigramm语言模型。
我们定义并使用BigramLanguageModel:
import torch
import torch.nn as nn
from torch.nn import functional as F
class BigramLanguageModel(nn.Module):
def __init__(self, vocab_size):
super().__init__()
# 词汇数,单词维度
self.token_embedding_table = nn.Embedding(vocab_size, vocab_size)
def forward(self, idx):
logits = self.token_embedding_table(idx) # (B, T, C) C=vocab_size
return logits
class HwcApp(object):
def __init__(self):
self.name = 'hwc_app.HwcApp'
# hwc_app.py::HwcApp.startup
def startup(self, args={}):
torch.manual_seed(1337)
ds_rfn = 'datasets/yolo3d.txt'
Yolo3dDs.generate_vocab(ds_rfn=ds_rfn)
ds = Yolo3dDs(ds_rfn=ds_rfn)
Xb, yb = ds.get_batch(mode='train')
ds.disp_batch(Xb=Xb, yb=yb)
m = BigramLanguageModel(AppRegistry.vocab_size)
criterion = nn.CrossEntropyLoss()
# 推理过程
out = m(Xb)
print(f'out: {out.shape}')
B, T, C = out.shape
out = out.reshape(B*T, C)
yb = yb.reshape(B*T)
loss = criterion(out, yb)
print(f'loss: {loss};')
完整代码请参考:
git clone https://gitee.com/yt7589/hwcgpt.git
cd hwcgpt
git checkout v0.0.3_1
在这里跟Karpathy教程中的内容有所出入,我们将计算Loss的过程放到了模型之外,同时计算Loss时使用的是nn.CrossEntropy而不是F.cross_entropy。其中B代表batch_size,T为序列长度,C为单词维主,所以out的形状打出来是(4, 8, 211)。
4. 代码生成
下面我们在模型没有经过任何学习的情况下,尝试一下代码生成。在BigramLanguageModel类中添加generate方法:
# ann/bigram_language_model.py::BigramLanguageModel.generate
def generate(self, idx, max_new_tokens):
for _ in range(max_new_tokens):
logits = self(idx)
logits = logits[:, -1, :] # (B, T, C) => (B, C)
probs = F.softmax(logits, dim=-1) # (B, C)
idx_next = torch.multinomial(probs, num_samples=1)
idx = torch.cat((idx, idx_next), dim=1)
return idx
# hwc_app.py::HwcApp.startup
def startup(self, args={}):
torch.manual_seed(1337)
ds_rfn = 'datasets/yolo3d.txt'
Yolo3dDs.generate_vocab(ds_rfn=ds_rfn)
ds = Yolo3dDs(ds_rfn=ds_rfn)
Xb, yb = ds.get_batch(mode='train')
ds.disp_batch(Xb=Xb, yb=yb)
m = BigramLanguageModel(AppRegistry.vocab_size)
criterion = nn.CrossEntropyLoss()
# 推理过程
out = m(Xb)
print(f'out: {out.shape}')
B, T, C = out.shape
out = out.reshape(B*T, C)
yb = yb.reshape(B*T)
loss = criterion(out, yb)
print(f'loss: {loss};')
soc = torch.zeros((1, 1), dtype=torch.long) # start of sentense
rsts = m.generate(idx=soc, max_new_tokens=100) # (B, max_new_tokens)=(1, 100)
rst = rsts[0].tolist()
gen_code = AppRegistry.decode(rst)
print(gen_code)
# 生成结果
?:因·S`表且数?均N现移?未d平7~度号o因作8J2s像是 m>的高在检g表t宽Y是p偏e@ 用参Q€f离距`:二差*量证F?平无)e为w×Q'e.¡框是$置偏Rhg成N?别/rh平方作
¡六R¨
如上所示,只是随机生成了一些无意义的字符,这是因为我们的BigramLanguageModel没有经过任何训练,只能输出随机结果。完整代码请参考:
git clone https://gitee.com/yt7589/hwcgpt.git
cd hwcgpt
git checkout v0.0.4
5. 网络训练
接下来,我们对这个简单的神经网络进行训练,然后再让其生成文本:
def startup(self, args={}):
torch.manual_seed(1337)
AppRegistry.device = 'cuda' if torch.cuda.is_available() else 'cpu'
# self.train()
self.predict()
def train(self):
ds_rfn = 'datasets/yolo3d.txt'
Yolo3dDs.generate_vocab(ds_rfn=ds_rfn)
ds = Yolo3dDs(ds_rfn=ds_rfn)
m = BigramLanguageModel(AppRegistry.vocab_size)
m = m.to(AppRegistry.device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.AdamW(m.parameters(), lr=1e-3)
for epoch in range(AppRegistry.epochs):
X, y = ds.get_batch('train')
X = X.to(AppRegistry.device)
y = y.to(AppRegistry.device)
y_hat = m(X)
B, T, C = y_hat.shape
y_hat = y_hat.reshape(B*T, C)
y = y.reshape(B*T)
loss = criterion(y_hat, y)
optimizer.zero_grad(set_to_none=True)
loss.backward()
optimizer.step()
print(f'{epoch}: {loss};')
# 保存模型
torch.save(m.state_dict(), './work/ckpts/hwc.pth')
soc = torch.zeros((1, 1), dtype=torch.long).to(AppRegistry.device) # start of sentense
rsts = m.generate(idx=soc, max_new_tokens=100) # (B, max_new_tokens)=(1, 100)
rst = rsts[0].tolist()
gen_code = AppRegistry.decode(rst)
print(gen_code)
def predict(self):
ds_rfn = 'datasets/yolo3d.txt'
Yolo3dDs.generate_vocab(ds_rfn=ds_rfn)
ds = Yolo3dDs(ds_rfn=ds_rfn)
m = BigramLanguageModel(AppRegistry.vocab_size)
m = m.to(AppRegistry.device)
# 载入模型
m.load_state_dict(torch.load('./work/ckpts/hwc.pth'))
soc = torch.zeros((1, 1), dtype=torch.long).to(AppRegistry.device) # start of sentense
rsts = m.generate(idx=soc, max_new_tokens=100) # (B, max_new_tokens)=(1, 100)
rst = rsts[0].tolist()
gen_code = AppRegistry.decode(rst)
print('生成代码:')
print(gen_code)
########################## 生成结果 #######################################
ayige, nlforet[ippabonchorso d():
c, s =1, ='fon(xed CARACLnddilintoraseefiseBb s[16,
上面生成的“代码”虽然依然很差,但是已经比最原始的版本强不少了,而我们训练50000个Epoch,也用不了几分钟。
完整代码请参考:
git clone https://gitee.com/yt7589/hwcgpt.git
cd hwcgpt
git checkout v0.0.5_2
好博客就要一起分享哦!分享海报
此处可发布评论
评论(0)展开评论
展开评论
您可能感兴趣的博客