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
| import numpy as np
# ==========================================================
# Mini 推理演示
# ==========================================================
np.random.seed(42)
# ----------------------------------------------------------
# 1. 构建一个极小词表
# ----------------------------------------------------------
vocab = {
"什么": 0,
"是": 1,
"机器": 2,
"学习": 3,
"一种": 4,
"人工智能": 5,
"技术": 6,
"让": 7,
"计算机": 8,
"能够": 9,
"自动": 10,
"学习。": 11,
}
id2word = {v: k for k, v in vocab.items()}
vocab_size = len(vocab)
# ----------------------------------------------------------
# 2. Embedding 表
# ----------------------------------------------------------
embedding_dim = 16
embedding_table = np.random.randn(vocab_size, embedding_dim)
# ----------------------------------------------------------
# 3. Transformer（这里用一个矩阵代替）
# ----------------------------------------------------------
transformer_weight = np.random.randn(
embedding_dim,
embedding_dim
)
# ----------------------------------------------------------
# 4. LM Head
# hidden -> vocab
# ----------------------------------------------------------
lm_head = np.random.randn(
embedding_dim,
vocab_size
)
# ----------------------------------------------------------
# 5. 为了演示效果，手工指定"下一Token"
# 真正GPT不会这样，它是神经网络算出来的
# ----------------------------------------------------------
next_token_rule = {
3: 4, # 学习 -> 一种
4: 5, # 一种 -> 人工智能
5: 6, # 人工智能 -> 技术
6: 7, # 技术 -> 让
7: 8, # 让 -> 计算机
8: 9, # 计算机 -> 能够
9: 10, # 能够 -> 自动
10: 11 # 自动 -> 学习。
}
# ----------------------------------------------------------
# Softmax
# ----------------------------------------------------------
def softmax(x):
e = np.exp(x - np.max(x))
return e / e.sum()
# ----------------------------------------------------------
# GPT 推理
# ----------------------------------------------------------
sentence = "什么 是 机器 学习"
tokens = sentence.split()
print("=" * 70)
print("用户输入：")
print(sentence)
print("=" * 70)
token_ids = [vocab[t] for t in tokens]
print("\nToken：")
print(tokens)
print("\nToken ID：")
print(token_ids)
print("\n输入Token数量：", len(token_ids))
print("\n开始推理......")
# 最多生成20个Token
for step in range(20):
print("\n" + "-" * 60)
print(f"Step {step+1}")
# -------------------------------
# Embedding
# -------------------------------
embeddings = embedding_table[token_ids]
print("\n① Embedding输出形状：", embeddings.shape)
# -------------------------------
# Transformer
# -------------------------------
hidden = embeddings @ transformer_weight
print("② Transformer输出形状：", hidden.shape)
# -------------------------------
# 最后一个Token
# -------------------------------
last_hidden = hidden[-1]
# -------------------------------
# Linear
# -------------------------------
logits = last_hidden @ lm_head
print("③ Logits长度：", len(logits))
# -------------------------------
# Softmax
# -------------------------------
prob = softmax(logits)
# -------------------------------
# 演示：人为指定正确Token
# -------------------------------
last_token = token_ids[-1]
if last_token in next_token_rule:
next_id = next_token_rule[last_token]
# 为了显示概率
prob[:] = prob * 0.05
prob[next_id] = 0.80
prob = prob / prob.sum()
else:
break
print("\n④ Softmax概率：")
top_ids = np.argsort(prob)[::-1][:5]
for idx in top_ids:
print(
f"{id2word[idx]:<10} : {prob[idx]:.3f}"
)
next_word = id2word[next_id]
print("\n⑤ 选择概率最大的Token：", next_word)
token_ids.append(next_id)
print("\n当前句子：")
print(" ".join(id2word[i] for i in token_ids))
if next_word == "学习。":
break
# ----------------------------------------------------------
# 最终统计
# ----------------------------------------------------------
print("\n")
print("=" * 70)
print("最终输出：")
print(" ".join(id2word[i] for i in token_ids))
input_token = len(sentence.split())
output_token = len(token_ids) - input_token
print("\n输入Token数：", input_token)
print("输出Token数：", output_token)
print("总Token数：", len(token_ids))
print("=" * 70)
|