Stable-string commited on
Commit
a28dcd7
·
1 Parent(s): f22f79f

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +211 -102
README.md CHANGED
@@ -1,106 +1,215 @@
1
  这是一个基于cpm训练的模型,你可以使用这个模型来续写文章,文章大体接近于古代🤗。
2
- 您可以将模型保存到./mdoel,使用下面的一个示例来运行这个模型🥳:
3
- ```python
4
- import torch
5
- import torch.nn.functional as F
6
- from transformers import GPT2LMHeadModel, CpmTokenizer
7
-
8
- def generate_next_token(input_ids):
9
- input_ids = input_ids[:, -context_len:]
10
- outputs = model(input_ids=input_ids)
11
- logits = outputs.logits
12
- next_token_logits = logits[0, -1, :]
13
- next_token_logits = next_token_logits / temperature
14
- next_token_logits[unk_id] = -float('Inf')
15
- filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=topk, top_p=topp)
16
- next_token_id = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
17
- return next_token_id
18
-
19
- def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')):
20
- assert logits.dim() == 1
21
- top_k = min(top_k, logits.size(-1))
22
- if top_k > 0:
23
- indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
24
- logits[indices_to_remove] = filter_value
25
-
26
- if top_p > 0.0:
27
- sorted_logits, sorted_indices = torch.sort(logits, descending=True)
28
- cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
29
- sorted_indices_to_remove = cumulative_probs > top_p
30
- sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
31
- sorted_indices_to_remove[..., 0] = 0
32
- indices_to_remove = sorted_indices[sorted_indices_to_remove]
33
- logits[indices_to_remove] = filter_value
34
- return logits
35
-
36
- def set_logger(log_path):
37
- logger = logging.getLogger(__name__)
38
- logger.setLevel(logging.INFO)
39
- formatter = logging.Formatter(
40
- '%(asctime)s - %(levelname)s - %(message)s')
41
- file_handler = logging.FileHandler(
42
- filename=log_path)
43
- file_handler.setFormatter(formatter)
44
- file_handler.setLevel(logging.INFO)
45
- logger.addHandler(file_handler)
46
- console = logging.StreamHandler()
47
- console.setLevel(logging.DEBUG)
48
- console.setFormatter(formatter)
49
- logger.addHandler(console)
50
- return logger
51
-
52
- # 参数
53
- device = 'cuda' if torch.cuda.is_available() else 'cpu'
54
- temperature = 1
55
- topk = 0
56
- topp = 0.85
57
- context_len = 800
58
- model_path = "./model"
59
-
60
- # 加载tokenizer
61
- tokenizer = CpmTokenizer(vocab_file=model_path+"/chinese_vocab.model")
62
- eod_id = tokenizer.convert_tokens_to_ids("<eod>") # 文档结束符
63
- sep_id = tokenizer.sep_token_id
64
- unk_id = tokenizer.unk_token_id
65
-
66
- # 加载模型
67
- model = GPT2LMHeadModel.from_pretrained(model_path)
68
- model.eval()
69
- model = model.to(device)
70
-
71
- def generate_zuowen(title, context, max_len):
72
- title_ids = tokenizer.encode(title, add_special_tokens=False)
73
- context_ids = tokenizer.encode(context, add_special_tokens=False)
74
- input_ids = title_ids + [sep_id] + context_ids
75
- cur_len = len(input_ids)
76
- last_token_id = input_ids[-1]
77
- input_ids = torch.tensor([input_ids], dtype=torch.long, device=device)
78
-
79
- while True:
80
- next_token_id = generate_next_token(input_ids)
81
- input_ids = torch.cat((input_ids, next_token_id.unsqueeze(0)), dim=1)
82
- cur_len += 1
83
- word = tokenizer.convert_ids_to_tokens(next_token_id.item())
84
- if cur_len >= max_len and last_token_id == 8 and next_token_id == 3:
85
- break
86
- if cur_len >= max_len and word in [".", "。", "!", "!", "?", "?", ",", ","]:
87
- break
88
- if next_token_id == eod_id:
89
- break
90
- word_mok = word.replace('▁', '').replace('▃', '\n ')
91
- print(word_mok, end='', flush=True)
92
- result = tokenizer.decode(input_ids.squeeze(0))
93
- content = result.split("<sep>")[1]
94
- return content
95
-
96
- if __name__ == '__main__':
97
- title='天下大势'#标题
98
- context='论天下大势'#上文,可以为空
99
- max_len=100#续写的字数
100
- max_len += len(context)
101
- print(context,end='')
102
- content = generate_zuowen(title, context, max_len)
103
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
104
  如果您认为生成到后期变慢,可以设置context_len(参考的上文的长度)小一些🤓。
105
  运行上面的例子,还需要一些环境👽。
106
  在cmd输入以下命令:
 
1
  这是一个基于cpm训练的模型,你可以使用这个模型来续写文章,文章大体接近于古代🤗。
2
+ 您可以将模型保存到./mdoel,使用下面的两个示例中的一个来运行这个模型🥳:
3
+ <details>
4
+ <summary>示例1</summary>
5
+ ```python
6
+ import torch
7
+ import torch.nn.functional as F
8
+ from transformers import GPT2LMHeadModel, CpmTokenizer
9
+
10
+ def generate_next_token(input_ids):
11
+ input_ids = input_ids[:, -context_len:]
12
+ outputs = model(input_ids=input_ids)
13
+ logits = outputs.logits
14
+ next_token_logits = logits[0, -1, :]
15
+ next_token_logits = next_token_logits / temperature
16
+ next_token_logits[unk_id] = -float('Inf')
17
+ filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=topk, top_p=topp)
18
+ next_token_id = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
19
+ return next_token_id
20
+
21
+ def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')):
22
+ assert logits.dim() == 1
23
+ top_k = min(top_k, logits.size(-1))
24
+ if top_k > 0:
25
+ indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
26
+ logits[indices_to_remove] = filter_value
27
+
28
+ if top_p > 0.0:
29
+ sorted_logits, sorted_indices = torch.sort(logits, descending=True)
30
+ cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
31
+ sorted_indices_to_remove = cumulative_probs > top_p
32
+ sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
33
+ sorted_indices_to_remove[..., 0] = 0
34
+ indices_to_remove = sorted_indices[sorted_indices_to_remove]
35
+ logits[indices_to_remove] = filter_value
36
+ return logits
37
+
38
+ def set_logger(log_path):
39
+ logger = logging.getLogger(__name__)
40
+ logger.setLevel(logging.INFO)
41
+ formatter = logging.Formatter(
42
+ '%(asctime)s - %(levelname)s - %(message)s')
43
+ file_handler = logging.FileHandler(
44
+ filename=log_path)
45
+ file_handler.setFormatter(formatter)
46
+ file_handler.setLevel(logging.INFO)
47
+ logger.addHandler(file_handler)
48
+ console = logging.StreamHandler()
49
+ console.setLevel(logging.DEBUG)
50
+ console.setFormatter(formatter)
51
+ logger.addHandler(console)
52
+ return logger
53
+
54
+ # 参数
55
+ device = 'cuda' if torch.cuda.is_available() else 'cpu'
56
+ temperature = 1
57
+ topk = 0
58
+ topp = 0.85
59
+ context_len = 800
60
+ model_path = "./model"
61
+
62
+ # 加载tokenizer
63
+ tokenizer = CpmTokenizer(vocab_file=model_path+"/chinese_vocab.model")
64
+ eod_id = tokenizer.convert_tokens_to_ids("<eod>") # 文档结束符
65
+ sep_id = tokenizer.sep_token_id
66
+ unk_id = tokenizer.unk_token_id
67
+
68
+ # 加载模型
69
+ model = GPT2LMHeadModel.from_pretrained(model_path)
70
+ model.eval()
71
+ model = model.to(device)
72
+
73
+ def generate_zuowen(title, context, max_len):
74
+ title_ids = tokenizer.encode(title, add_special_tokens=False)
75
+ context_ids = tokenizer.encode(context, add_special_tokens=False)
76
+ input_ids = title_ids + [sep_id] + context_ids
77
+ cur_len = len(input_ids)
78
+ last_token_id = input_ids[-1]
79
+ input_ids = torch.tensor([input_ids], dtype=torch.long, device=device)
80
+
81
+ while True:
82
+ next_token_id = generate_next_token(input_ids)
83
+ input_ids = torch.cat((input_ids, next_token_id.unsqueeze(0)), dim=1)
84
+ cur_len += 1
85
+ word = tokenizer.convert_ids_to_tokens(next_token_id.item())
86
+ if cur_len >= max_len and last_token_id == 8 and next_token_id == 3:
87
+ break
88
+ if cur_len >= max_len and word in [".", "。", "!", "!", "?", "?", ",", ","]:
89
+ break
90
+ if next_token_id == eod_id:
91
+ break
92
+ word_mok = word.replace('▁', '').replace('▃', '\n ')
93
+ print(word_mok, end='', flush=True)
94
+ result = tokenizer.decode(input_ids.squeeze(0))
95
+ content = result.split("<sep>")[1]
96
+ return content
97
+
98
+ if __name__ == '__main__':
99
+ title='天下大势'#标题
100
+ context='论天下大势'#上文,可以为空
101
+ max_len=100#续写的字数
102
+ max_len += len(context)
103
+ print(context,end='')
104
+ content = generate_zuowen(title, context, max_len)
105
+ ```
106
+ <details>
107
+ <summary>示例2</summary>
108
+ ```python
109
+ import torch
110
+ import torch.nn.functional as F
111
+ from transformers import AutoModel, AutoTokenizer
112
+
113
+ def generate_next_token(input_ids):
114
+ input_ids = input_ids[:, -context_len:]
115
+ outputs = model(input_ids=input_ids)
116
+ logits = outputs.logits
117
+ next_token_logits = logits[0, -1, :]
118
+ next_token_logits = next_token_logits / temperature
119
+ next_token_logits[unk_id] = -float('Inf')
120
+ filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=topk, top_p=topp)
121
+ next_token_id = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
122
+ return next_token_id
123
+
124
+ def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')):
125
+ assert logits.dim() == 1
126
+ top_k = min(top_k, logits.size(-1))
127
+ if top_k > 0:
128
+ indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
129
+ logits[indices_to_remove] = filter_value
130
+
131
+ if top_p > 0.0:
132
+ sorted_logits, sorted_indices = torch.sort(logits, descending=True)
133
+ cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
134
+ sorted_indices_to_remove = cumulative_probs > top_p
135
+ sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
136
+ sorted_indices_to_remove[..., 0] = 0
137
+ indices_to_remove = sorted_indices[sorted_indices_to_remove]
138
+ logits[indices_to_remove] = filter_value
139
+ return logits
140
+
141
+ def set_logger(log_path):
142
+ logger = logging.getLogger(__name__)
143
+ logger.setLevel(logging.INFO)
144
+ formatter = logging.Formatter(
145
+ '%(asctime)s - %(levelname)s - %(message)s')
146
+ file_handler = logging.FileHandler(
147
+ filename=log_path)
148
+ file_handler.setFormatter(formatter)
149
+ file_handler.setLevel(logging.INFO)
150
+ logger.addHandler(file_handler)
151
+ # 创建一个handler,用于将日志输出到控制台
152
+ console = logging.StreamHandler()
153
+ console.setLevel(logging.DEBUG)
154
+ console.setFormatter(formatter)
155
+ logger.addHandler(console)
156
+ return logger
157
+
158
+ # 参数
159
+ device = 'cuda' if torch.cuda.is_available() else 'cpu'
160
+ temperature = 1
161
+ topk = 0
162
+ topp = 0.85
163
+ context_len = 800
164
+ model_path = "model"
165
+
166
+ # 加载tokenizer
167
+ tokenizer = AutoTokenizer.from_pretrained(model_path)
168
+ eod_id = tokenizer.convert_tokens_to_ids("<eod>") # 文档结束符
169
+ sep_id = tokenizer.sep_token_id
170
+ unk_id = tokenizer.unk_token_id
171
+
172
+ # 加载模型
173
+ model = AutoModel.from_pretrained(model_path)
174
+ model.eval()
175
+ model = model.to(device)
176
+
177
+ def generate_zuowen(title, context, max_len):
178
+ title_ids = tokenizer.encode(title, add_special_tokens=False)
179
+ context_ids = tokenizer.encode(context, add_special_tokens=False)
180
+ input_ids = title_ids + [sep_id] + context_ids
181
+ cur_len = len(input_ids)
182
+ last_token_id = input_ids[-1]
183
+ input_ids = torch.tensor([input_ids], dtype=torch.long, device=device)
184
+
185
+ while True:
186
+ next_token_id = generate_next_token(input_ids)
187
+ input_ids = torch.cat((input_ids, next_token_id.unsqueeze(0)), dim=1)
188
+ cur_len += 1
189
+ word = tokenizer.convert_ids_to_tokens(next_token_id.item())
190
+ if cur_len >= max_len and last_token_id == 8 and next_token_id == 3:
191
+ break
192
+ if cur_len >= max_len and word in [".", "。", "!", "!", "?", "?", ",", ","]:
193
+ break
194
+ if next_token_id == eod_id:
195
+ break
196
+ word_mok = word.replace('▁', '').replace('▃', '\n ')
197
+ print(word_mok, end='', flush=True)
198
+ result = tokenizer.decode(input_ids.squeeze(0))
199
+ content = result.split("<sep>")[1]
200
+ return content
201
+
202
+ if __name__ == '__main__':
203
+ title='标题'#标题
204
+ context='上文'#上文,可以为空
205
+ max_len=100#续写的字数
206
+ max_len += len(context)
207
+ print(context,end='')
208
+ content = generate_zuowen(title, context, max_len)
209
+ ```
210
+ </details>
211
+
212
+
213
  如果您认为生成到后期变慢,可以设置context_len(参考的上文的长度)小一些🤓。
214
  运行上面的例子,还需要一些环境👽。
215
  在cmd输入以下命令: