-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmodels_of_rl.py
198 lines (184 loc) · 8.55 KB
/
models_of_rl.py
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
"""
classes for ChatGLM RLHF
Critic model
Action model is ChatGLM, 所以可省略
Reward model
"""
import copy
import torch
from torch import nn
from transformers import BertTokenizer, BertModel, AutoModelForCausalLM
import numpy as np
from functools import partial
from accelerate import Accelerator
accelerator = Accelerator()
"""
critic 的词表最好和action模型的词表一样这样便于对每个生成的token进行打分，
不一致的词表会导致打分不对齐，所以选择用一样的模型但是加一下打分的输出
为了减小打分模型的大小，可以把原来的模型的layers缩减层数。
这样直接继承了，原来的token embedding
"""
class Critic(nn.Module):
def __init__(self, model_name_or_path, device) -> None:
super().__init__()
layers_keep = 1
model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True)
del model.lm_head
if "moss" in model_name_or_path:
model = model.transformer
model.h = model.h[:layers_keep]
elif "llama" in model_name_or_path:
model = model.model
model.layers = model.layers[:layers_keep]
# solve RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'
if "cuda" in device:
model = model.half().cuda(device) # half for gpu only
# model.ln_f._hf_hook.execution_device=0
elif "cpu" == device:
model = model.bfloat16()
else:
model = model.float()
self.model = model
self.layers = self.model.layers if "llama" in model_name_or_path else self.model.h
intermediate_size = 2**(3*4)
self.output_linear = nn.Linear(self.model.config.hidden_size, intermediate_size, device=self.model.device, dtype=self.model.dtype, bias=False)
self.output_linear2 = nn.Sequential()
current_size = intermediate_size
while current_size>1:
in_features = current_size
out_features = in_features//2
fc_in = nn.Linear(in_features, out_features, device=self.model.device, dtype=self.model.dtype, bias=False)
# act = torch.nn.GELU()
act = torch.nn.ReLU()
fc_out_size = out_features//2
fc_out = nn.Linear(out_features, fc_out_size, device=self.model.device, dtype=self.model.dtype, bias=False)
dropout = nn.Dropout(p=0.15)
v_out_act = torch.nn.GLU()
current_size = fc_out_size//2
self.output_linear2.extend([fc_in, act, dropout, fc_out, v_out_act])
self.dtype = self.model.dtype
self.device = self.model.device
self.model_name_or_path = model_name_or_path
def forward(self, **kwargs):
output = self.model(**kwargs)
v = self.output_linear(output.last_hidden_state)
v2 = self.output_linear2(v)
values = v2
# values = torch.sigmoid(v)
# values = torch.tanh(v)
return values.squeeze(-1)
"""
一样的原因，不需要再把生成的token ids转成文字在再转到目标ids，
所以也用chatglm直接做基模型，
只是这里只取最后的token算出对整句生成的奖励分数，具体取哪个位置可以
后续在代码里面指定，比如用torch.gather
"""
Reward = Critic
# Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
token_embeddings = model_output[0] # First element of model_output contains all token embeddings
input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
def jaccard(s1, s2):
"""
可能有字符串重合但是语义不一致问题，
TODO 可以用多阶的jaccard来解决
"""
assert len(s1)+len(s2)>0
s1 = set(s1)
s2 = set(s2)
s_or = s1 | s2
s_and = s1 & s2
jaccard_distance = len(s_and)/len(s_or)
return jaccard_distance
class RewardBySimilarity(nn.Module):
def __init__(self, device="cpu", model_name_or_path='shibing624/text2vec-base-chinese') -> None:
"""
model_name_or_path:
chinese: 'shibing624/text2vec-base-chinese',
'twnlp/chinese-macbert-base-similarity',
'GanymedeNil/text2vec-large-chinese',
'GanymedeNil/text2vec-base-chinese'
english: 'l3cube-pune/hindi-sentence-similarity-sbert',
'hiiamsid/sentence_similarity_hindi',
'Maite89/Roberta_finetuning_semantic_similarity_stsb_multi_mt'
"""
super().__init__()
# Load model from HuggingFace Hub
tokenizer = BertTokenizer.from_pretrained(model_name_or_path)
model = BertModel.from_pretrained(model_name_or_path)
model.eval()
self.model = model.to(device)
self.tokenizer = tokenizer
self.device = device
def forward(self, gen_texts=["你好"],
good_answers=['你好', "hello"],
bad_answers=['再见', 'bye bye'],
weight_for_cos_and_jaccard = [0.5, 0.5]):
examples = good_answers + bad_answers
example_num = len(examples)
assert len(gen_texts)>0 and example_num>0
reward_direction = torch.ones(example_num, device=self.model.device)
reward_direction[len(good_answers):] = -1
sentences = gen_texts + examples
# Tokenize sentences
encoded_input = self.tokenizer(sentences, padding=True, return_tensors='pt')
ids = self.tokenizer.batch_encode_plus(sentences, add_special_tokens=False)["input_ids"]
# temporary truncate position_ids
batch_size, max_seq_len = encoded_input["input_ids"].shape
if max_seq_len > self.model.config.max_position_embeddings:
encoded_input["position_ids"] = torch.arange(max_seq_len).expand((1, -1)).repeat(batch_size, 1)
encoded_input["position_ids"] = encoded_input["position_ids"]/max_seq_len*self.model.config.max_position_embeddings
encoded_input["position_ids"] = encoded_input["position_ids"].floor().long()
# Compute token embeddings
with torch.no_grad():
encoded_input = encoded_input.to(self.model.device)
model_output = self.model(**encoded_input)
# Perform pooling. In this case, max pooling.
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
gen_text_vecs = sentence_embeddings[:len(gen_texts)]
answers_vecs = sentence_embeddings[len(gen_texts):]
reward_ = []
for i in range(gen_text_vecs.shape[0]):
gen_text_vecs_ = gen_text_vecs[i:i+1]
# 用一下广播计算cos
coses = torch.cosine_similarity(gen_text_vecs_, answers_vecs, dim=1)
# 余弦截断
coses[(coses<0)] = 0
# 计算 jaccard距离
jaccard_s1 = partial(jaccard, ids[i])
jaccards = torch.tensor(np.vectorize(jaccard_s1)(np.array(ids[-len(examples):], dtype=object)), dtype=coses.dtype, device=coses.device)
similarity = weight_for_cos_and_jaccard[0]*coses + weight_for_cos_and_jaccard[1]*jaccards
value, index = similarity.max(dim=-1)
reward_.append(value*reward_direction[index])
reward = torch.stack(reward_)
return reward
def test_reward_by_similarity():
reward_model = RewardBySimilarity()
reward = reward_model()
print(reward)
def test_critic():
# model chooose
model_name_or_path = ["fnlp/moss-moon-003-sft", "decapoda-research/llama-7b-hf"][0]
critic_device = "cuda:0" # "cpu"
critic = Critic(model_name_or_path=model_name_or_path, device=critic_device)
input_ids = torch.ones((2, 4), dtype=torch.long, device=critic.device)
output = critic(input_ids=input_ids)
print(output.shape)
def test_reward():
# with torch.no_grad():
# input_ids_RM = sequences.to(RM_device)
# rewards_ = reward_model(input_ids = input_ids_RM)
# # 由于只对最后的整句进行reward，所以只有最后一个action后的state有reward
# rewards = torch.zeros_like( sequences, dtype=rewards_.dtype, device=rewards_.device)
# pad_id = tokenizer.convert_tokens_to_ids("<pad>")
# masks = ( sequences!=pad_id).long().to(RM_device)
# final_position = masks.sum(dim=-1)-1
# index=final_position.unsqueeze(-1)
# reward = rewards_.gather(dim=1, index=index)
# rewards.scatter_(dim=1, index=index, src=reward)
pass
if __name__ == "__main__":
test_reward_by_similarity()
# test_critic()
pass