-
Notifications
You must be signed in to change notification settings - Fork 22
/
tuning_lm_with_rl.py
254 lines (217 loc) · 9.04 KB
/
tuning_lm_with_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
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
254
import os
import torch
from dataclasses import dataclass, field
from typing import Optional
from accelerate import Accelerator
from datasets import load_dataset
from peft import LoraConfig
from tqdm import tqdm
from transformers import (
Adafactor,
AutoTokenizer,
LlamaTokenizer,
HfArgumentParser,
pipeline
)
from trl import AutoModelForCausalLMWithValueHead, PPOConfig, PPOTrainer, set_seed
from trl.core import LengthSampler
DEFAULT_PAD_TOKEN = "[PAD]"
DEFAULT_EOS_TOKEN = "</s>"
DEFAULT_BOS_TOKEN = "</s>"
DEFAULT_UNK_TOKEN = "</s>"
tqdm.pandas()
@dataclass
class ScriptArguments:
"""
The name of the Casual LM model we wish to fine with PPO
"""
# NOTE: gpt2 models use Conv1D instead of Linear layers which are not yet supported in 8 bit mode
# models like gpt-neo* models are more suitable.
model_name: Optional[str] = field(default="", metadata={"help": "the model name"})
tokenizer_name: Optional[str] = field(default="", metadata={"help": "the tokenizer name"})
reward_model_name: Optional[str] = field(default="", metadata={"help": "the reward model name"})
dataset_name: Optional[str] = field(default="", metadata={"help": "the dataset name"})
log_with: Optional[str] = field(default=None, metadata={"help": "use 'wandb' to log with wandb"})
learning_rate: Optional[float] = field(default=1.41e-5, metadata={"help": "the learning rate"})
max_length: Optional[int] = field(default=512, metadata={"help": "maximum length for input"})
output_max_length: Optional[int] = field(default=128, metadata={"help": "maximum length for generation"})
mini_batch_size: Optional[int] = field(default=1, metadata={"help": "the PPO minibatch size"})
batch_size: Optional[int] = field(default=32, metadata={"help": "the batch size"})
ppo_epochs: Optional[int] = field(default=4, metadata={"help": "the number of ppo epochs"})
gradient_accumulation_steps: Optional[int] = field(
default=4, metadata={"help": "the number of gradient accumulation steps"}
)
adafactor: Optional[bool] = field(default=False, metadata={"help": "whether to use the adafactor optimizer"})
early_stopping: Optional[bool] = field(default=False, metadata={"help": "whether to early stop"})
target_kl: Optional[float] = field(default=0.1, metadata={"help": "kl target for early stopping"})
reward_baseline: Optional[float] = field(
default=0.0,
metadata={"help": "a baseline value that is subtracted from the reward"},
)
batched_gen: Optional[bool] = field(default=False, metadata={"help": "whether to use the batched text gen"})
save_freq: Optional[int] = field(default=None, metadata={"help": "n steps to save the model"})
output_dir: Optional[str] = field(default="./checkpoints/tuning_llama_rl/",
metadata={"help": "n steps to save the model"})
seed: Optional[int] = field(default=0, metadata={"help": "the seed"})
parser = HfArgumentParser(ScriptArguments)
script_args: ScriptArguments = parser.parse_args_into_dataclasses()[0]
set_seed(script_args.seed)
# Below is an example function to build the dataset. In our case, we use the IMDB dataset
# from the `datasets` library. One should customize this function to train the model on
# its own dataset.
def build_dataset(
tokenizer, dataset_name, input_min_text_length=2, input_max_text_length=8
):
"""
Build dataset for training. This builds the dataset from `load_dataset`, one should
customize this function to train the model on its own dataset.
Args:
dataset_name (`str`):
The name of the dataset to be loaded.
Returns:
dataloader (`torch.utils.data.DataLoader`):
The dataloader for the dataset.
"""
train_dataset = load_dataset(dataset_name, split="train")
original_columns = train_dataset.column_names
num_proc = 24
def preprocess_function(examples):
new_examples = {
"query": [],
"input_ids": [],
}
for question in examples["question"]:
query = "Question: " + question + "\n\nAnswer: "
tokenized_question = tokenizer(query, truncation=True)
new_examples["query"].append(query)
new_examples["input_ids"].append(tokenized_question["input_ids"])
return new_examples
ds = train_dataset.map(
preprocess_function,
batched=True,
num_proc=num_proc,
remove_columns=original_columns,
)
ds = ds.filter(lambda x: len(x["input_ids"]) < script_args.max_length, batched=False)
ds.set_format(type="torch")
return ds
def collator(data):
return dict((key, [d[key] for d in data]) for key in data[0])
reward_model_name = script_args.reward_model_name
config = PPOConfig(
model_name=script_args.model_name,
learning_rate=script_args.learning_rate,
log_with=script_args.log_with,
batch_size=script_args.batch_size,
mini_batch_size=script_args.mini_batch_size,
gradient_accumulation_steps=script_args.gradient_accumulation_steps,
optimize_cuda_cache=True,
early_stopping=script_args.early_stopping,
target_kl=script_args.target_kl,
ppo_epochs=script_args.ppo_epochs,
seed=script_args.seed,
)
# We then define the arguments to pass to the sentiment analysis pipeline.
# We set `return_all_scores` to True to get the sentiment score for each token.
rw_kwargs = {
"return_all_scores": True,
"function_to_apply": "none",
"batch_size": 16,
"truncation": True
}
if "decapoda" in script_args.model_name.lower():
tokenizer = LlamaTokenizer.from_pretrained(script_args.model_name)
# required for llama
tokenizer.add_special_tokens(
{
"eos_token": DEFAULT_EOS_TOKEN,
"bos_token": DEFAULT_BOS_TOKEN,
"unk_token": DEFAULT_UNK_TOKEN,
"pad_token": DEFAULT_PAD_TOKEN,
}
)
else:
tokenizer = AutoTokenizer.from_pretrained(script_args.model_name)
if getattr(tokenizer, "pad_token", None) is None:
tokenizer.pad_token = tokenizer.eos_token
# We retrieve the dataloader by calling the `build_dataset` function.
dataset = build_dataset(tokenizer, script_args.dataset_name)
# Now let's build the model, the reference model, and the tokenizer.
current_device = Accelerator().local_process_index
lora_config = LoraConfig(
r=16,
lora_alpha=32,
lora_dropout=0.05,
bias="none",
task_type="CAUSAL_LM",
)
model = AutoModelForCausalLMWithValueHead.from_pretrained(
config.model_name,
load_in_8bit=True,
device_map={"": current_device},
peft_config=lora_config,
)
optimizer = None
if script_args.adafactor:
optimizer = Adafactor(
filter(lambda p: p.requires_grad, model.parameters()),
scale_parameter=False,
relative_step=False,
warmup_init=False,
lr=config.learning_rate,
)
# We then build the PPOTrainer, passing the model, the reference model, the tokenizer
ppo_trainer = PPOTrainer(
config,
model,
ref_model=None,
tokenizer=tokenizer,
dataset=dataset,
data_collator=collator,
optimizer=optimizer,
)
# We then build the sentiment analysis pipeline, passing the model name and the
# sentiment analysis pipeline arguments. Let's also make sure to set the device
# to the same device as the PPOTrainer.
device = ppo_trainer.accelerator.device
if ppo_trainer.accelerator.num_processes == 1:
device = 0 if torch.cuda.is_available() else "cpu" # to avoid a ` pipeline` bug
reward_model = pipeline(
"text-classification",
model=reward_model_name,
device_map={"": current_device},
model_kwargs={"load_in_8bit": True},
tokenizer=tokenizer,
)
# We then define the arguments to pass to the `generate` function. These arguments
# are passed to the `generate` function of the PPOTrainer, which is a wrapper around
# the `generate` function of the trained model.
generation_kwargs = {
# "min_length": -1,
"top_k": 0.0,
"top_p": 1.0,
"do_sample": True,
"pad_token_id": tokenizer.pad_token_id,
"eos_token_id": 100_000,
}
output_min_length = 32
output_max_length = script_args.output_max_length
output_length_sampler = LengthSampler(output_min_length, output_max_length)
for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
question_tensors = batch["input_ids"]
response_tensors = ppo_trainer.generate(
question_tensors,
return_prompt=False,
length_sampler=output_length_sampler,
**generation_kwargs,
)
batch["response"] = tokenizer.batch_decode(response_tensors, skip_special_tokens=True)
# Compute sentiment score
texts = [q + r for q, r in zip(batch["query"], batch["response"])]
reward_outputs = reward_model(texts, **rw_kwargs)
rewards = [torch.tensor(output[0]["score"] - script_args.reward_baseline) for output in reward_outputs]
# Run PPO step
stats = ppo_trainer.step(question_tensors, response_tensors, rewards)
ppo_trainer.log_stats(stats, batch, rewards)
if script_args.save_freq and epoch and epoch % script_args.save_freq == 0:
ppo_trainer.save_pretrained(script_args.output_dir + f"step_{epoch}")