1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
| '''
Decoder-Only transformer
pre layerNorm
'''
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import requests
import tiktoken
import math
batch_size = 4
context_len = 16
d_model = 64
num_blocks = 8
num_heads = 4
learning_rate = 1e-3
dropout = 0.1
max_iters = 5000
eval_interval = 100
eval_iters = 100
device = 'cuda' if torch.cuda.is_available() else 'cpu'
TORCH_SEED = 1337
torch.manual_seed(TORCH_SEED)
if not os.path.exists('/home/lizy/graduate/Transformer_learning/sales_textbook.txt'):
url = 'https://huggingface.co/datasets/goendalf666/sales-textbook_for_convincing_and_selling/raw/main/sales_textbook.txt'
with open('/home/lizy/graduate/Transformer_learning/sales_textbook.txt','wb') as f:
f.write(requests.get(url).content)
with open('/home/lizy/graduate/Transformer_learning/sales_textbook.txt','r', encoding='utf-8') as f:
text = f.read()
encoding = tiktoken.get_encoding("cl100k_base")
vocab_size = encoding.n_vocab
tokenized_text = encoding.encode(text)
tokenized_text=torch.tensor(tokenized_text,dtype=torch.long,device=device)
train_idex = int(len(tokenized_text) * 0.9)
train_data = tokenized_text[:train_idex]
valid_data = tokenized_text[train_idex:]
class FeedforwardNetwork(nn.Module):
def __init__(self,d_model,d_ff):
super(FeedforwardNetwork,self).__init__()
self.linear1 = nn.Linear(d_model,d_ff)
self.ReLU = nn.ReLU()
self.linear2 = nn.Linear(d_ff,d_model)
self.dropout = nn.Dropout(dropout)
def forward(self,x):
x=self.linear1(x)
x=self.ReLU(x)
x=self.linear2(x)
x=self.dropout(x)
return x
class MultiHeadAttention(nn.Module):
def __init__(self):
super().__init__()
self.Wqkv = nn.Linear(d_model,d_model*3)
self.projection_layer = nn.Linear(d_model,d_model)
self.dropout = nn.Dropout(dropout)
def forward(self,x):
B,T,C = x.shape
qkv = self.Wqkv(x).reshape(B,T,3,num_heads,C // num_heads)
q,k,v = qkv.unbind(dim=2)
q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
attn = (q @ k.transpose(-2, -1)) / math.sqrt(C // num_heads)
mask = torch.tril(torch.ones(T, T)).to(x.device)
attn = attn.masked_fill(mask == 0, float('-inf'))
attn = F.softmax(attn, dim=-1)
attn = self.dropout(attn)
out = (attn @ v).transpose(1, 2).contiguous().view(B, T, C)
return self.projection_layer(out)
class TransformerBlock(nn.Module):
def __init__(self):
super().__init__()
self.multi_head_attention_layer = MultiHeadAttention()
self.ffn = FeedforwardNetwork(d_model,d_model*4)
self.layer_norm_1=nn.LayerNorm(d_model)
self.layer_norm_2=nn.LayerNorm(d_model)
self.dropout = nn.Dropout(dropout)
def forward(self,x):
x = x + self.dropout(self.multi_head_attention_layer(self.layer_norm_1(x)))
x = x + self.dropout(self.ffn(self.layer_norm_2(x)))
return x
class TransformerLanguageModel(nn.Module):
def __init__(self):
super().__init__()
self.token_embedding_lookup_table = nn.Embedding(vocab_size, d_model)
self.transformer_blocks = nn.Sequential(*(
[TransformerBlock() for _ in range(num_blocks)]
+ [nn.LayerNorm(d_model)]
))
self.language_model_out_linear_layer = nn.Linear(d_model,vocab_size)
self.register_buffer('position_embedding', self._create_position_embedding())
def _create_position_embedding(self):
position_encoding_lookup_table = torch.zeros(context_len,d_model)
position = torch.arange(0,context_len,dtype=torch.float).unsqueeze(1)
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
position_encoding_lookup_table[:, 0::2] = torch.sin(position * div_term)
position_encoding_lookup_table[:, 1::2] = torch.cos(position * div_term)
return position_encoding_lookup_table
def forward(self,idx,targets=None):
B , T = idx.shape
position_embedding = self.position_embedding[:T, :].to(device)
x = self.token_embedding_lookup_table(idx) + position_embedding
x = self.transformer_blocks(x)
logits = self.language_model_out_linear_layer(x)
if targets is not None:
B, T, C = logits.shape
logits_reshaped = logits.view(B * T, C)
targets_reshaped = targets.view(B * T)
loss = F.cross_entropy(input=logits_reshaped, target=targets_reshaped)
else:
loss = None
return logits, loss
def generate(self, idx, max_new_tokens, temperature=0.8, top_k=50):
for _ in range(max_new_tokens):
idx_crop = idx[:, -context_len:] if idx.size(1) > context_len else idx
logits, _ = self(idx_crop)
logits = logits[:, -1, :] / temperature
if top_k is not None:
v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
logits[logits < v[:, [-1]]] = float('-inf')
probs = F.softmax(logits, dim=-1)
idx_next = torch.multinomial(probs, num_samples=1)
idx_next = torch.clamp(idx_next, 0, vocab_size - 1)
idx = torch.cat((idx, idx_next), dim=1)
return idx
model = TransformerLanguageModel()
model = model.to(device)
def get_batch(split):
data = train_data if split == 'train' else valid_data
idxs = torch.randint(low=0, high=len(data) - context_len, size=(batch_size,))
x = torch.stack([data[idx:idx + context_len] for idx in idxs]).to(device)
y = torch.stack([data[idx + 1:idx + context_len + 1] for idx in idxs]).to(device)
return x, y
@torch.no_grad()
def estimate_loss():
out = {}
model.eval()
for split in ['train', 'valid']:
losses = torch.zeros(eval_iters)
for k in range(eval_iters):
x_batch, y_batch = get_batch(split)
logits, loss = model(x_batch, y_batch)
losses[k] = loss.item()
out[split] = losses.mean()
model.train()
return out
optimizer = torch.optim.AdamW(params=model.parameters(), lr=learning_rate)
tracked_losses = list()
for step in range(max_iters):
if step % eval_interval == 0 or step == max_iters - 1:
losses = estimate_loss()
tracked_losses.append(losses)
print('Step:', step, 'Training Loss:', round(losses['train'].item(), 3), 'Validation Loss:',
round(losses['valid'].item(), 3))
xb, yb = get_batch('train')
logits, loss = model(xb, yb)
optimizer.zero_grad(set_to_none=True)
loss.backward()
optimizer.step()
torch.save(model.state_dict(), './model_para/model_pre-ckpt.pt')
model.eval()
start = 'The salesperson'
start_ids = encoding.encode(start)
x = (torch.tensor(start_ids, dtype=torch.long, device=device)[None, ...])
y = model.generate(x, max_new_tokens=100)
print('---------------')
try:
generated_text = encoding.decode(y[0].tolist())
except KeyError as e:
print(f"解码时遇到无效token，尝试忽略: {e}")
valid_tokens = []
for token in y[0].tolist():
try:
if 0 <= token < vocab_size:
valid_tokens.append(token)
except:
continue
generated_text = encoding.decode(valid_tokens)
print(encoding.decode(y[0].tolist()))
print('---------------')
|
