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
| import pandas as pd
import torch
from datasets import load_dataset
from tqdm import tqdm
from transformers import AutoTokenizer, set_seed, pipeline
from trl import AutoModelForCausalLMWithValueHead, create_reference_model, PPOConfig, PPOTrainer
from trl.core import LengthSampler
model_name = "Wenzhong-GPT2-110M"
config = PPOConfig(
model_name=model_name,
learning_rate=1.41e-5,
log_with="tensorboard",
batch_size=64,
gradient_accumulation_steps=8,
mini_batch_size=8,
)
sent_kwargs = {"return_all_scores": True, "function_to_apply": "none", "batch_size": 16}
gpt2_tokenizer = AutoTokenizer.from_pretrained(config.model_name)
gpt2_tokenizer.pad_token = gpt2_tokenizer.eos_token
set_seed(1)
def build_dataset(input_min_text_length=2, input_max_text_length=8):
ds = load_dataset(
'csv',
data_files={
"train": "train.tsv",
"dev": "dev.tsv",
},
delimiter='\t'
)['train']
ds = ds.rename_columns({"text_a": "review"})
ds = ds.filter(lambda x: len(x['review']) > 10, batched=False)
input_size = LengthSampler(input_min_text_length, input_max_text_length)
def tokenize(sample):
sample["input_ids"] = gpt2_tokenizer.encode(sample["review"][: input_size()])
sample["query"] = gpt2_tokenizer.decode(sample["input_ids"])
return sample
ds = ds.map(tokenize, batched=False)
ds.set_format(type="torch")
return ds
dataset = build_dataset()
def collator(data):
return dict((key, [d[key] for d in data]) for key in data[0])
sentiment_pipe = pipeline(
"sentiment-analysis",
model='/score_model/checkpoint-1000',
device='cuda:1'
)
text = "今天天气很好"
print(sentiment_pipe(text, **sent_kwargs))
text = '这个电影主角演的真心一般般'
print(sentiment_pipe(text, **sent_kwargs))
output_min_length = 16
output_max_length = 32
output_length_sampler = LengthSampler(output_min_length, output_max_length)
model = AutoModelForCausalLMWithValueHead.from_pretrained(config.model_name)
ref_model = create_reference_model(model)
ppo_trainer = PPOTrainer(config, model, ref_model, gpt2_tokenizer, dataset, data_collator=collator)
generation_kwargs = {
"min_length": -1,
"top_k": 0,
"top_p": 1,
"do_sample": True,
"pad_token_id": gpt2_tokenizer.eos_token_id,
"bos_token_id": gpt2_tokenizer.bos_token_id,
"eos_token_id": gpt2_tokenizer.eos_token_id,
}
device = 'cuda:0'
for i in range(3):
for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
query_tensors = batch['input_ids']
response_tensors = []
for query in query_tensors:
gen_len = output_length_sampler()
response = ppo_trainer.generate(query, **{**generation_kwargs, "max_new_tokens": gen_len})
response_tensors.append(response.squeeze()[-gen_len:])
batch["response"] = [gpt2_tokenizer.decode(r.squeeze()) for r in response_tensors]
texts = [q + r for q, r in zip(batch["query"], batch["response"])]
pipe_outputs = sentiment_pipe(texts, **sent_kwargs)
rewards = [torch.tensor(output[1]["score"]) for output in pipe_outputs]
stats = ppo_trainer.step(query_tensors, response_tensors, rewards)
ppo_trainer.log_stats(stats, batch, rewards)
bs = 16
game_data = dict()
dataset.set_format("pandas")
df_batch = dataset[:].sample(bs)
game_data["query"] = df_batch["query"].tolist()
query_tensors = df_batch["input_ids"].tolist()
response_tensors_ref, response_tensors = [], []
for i in range(bs):
gen_len = output_length_sampler()
output = ref_model.generate(
torch.tensor(query_tensors[i]).unsqueeze(dim=0).to(device), max_new_tokens=gen_len, **generation_kwargs
).squeeze()[-gen_len:]
response_tensors_ref.append(output)
output = model.generate(
torch.tensor(query_tensors[i]).unsqueeze(dim=0).to(device), max_new_tokens=gen_len, **generation_kwargs
).squeeze()[-gen_len:]
response_tensors.append(output)
game_data["response (before)"] = [gpt2_tokenizer.decode(response_tensors_ref[i]) for i in range(bs)]
game_data["response (after)"] = [gpt2_tokenizer.decode(response_tensors[i]) for i in range(bs)]
texts = [q + r for q, r in zip(game_data["query"], game_data["response (before)"])]
game_data["rewards (before)"] = [output[1]["score"] for output in sentiment_pipe(texts, **sent_kwargs)]
texts = [q + r for q, r in zip(game_data["query"], game_data["response (after)"])]
game_data["rewards (after)"] = [output[1]["score"] for output in sentiment_pipe(texts, **sent_kwargs)]
df_results = pd.DataFrame(game_data)
print(df_results)
print("mean:")
print(df_results[["rewards (before)", "rewards (after)"]].mean())
print()
print("median:")
print(df_results[["rewards (before)", "rewards (after)"]].median())
model.save_pretrained("gpt2-imdb-pos-v2", push_to_hub=False)
gpt2_tokenizer.save_pretrained("gpt2-imdb-pos-v2", push_to_hub=False)
|