Azure Machine learning Reinforcement learning PPO with GPT2 Pretrained model

Balamurugan Balakreshnan

·

Follow

3 min read

·

Aug 26, 2023

--

Listen

Share

Introduction

• Fine tune TRL model in Azure ML
• Using Azure ML
• Using NVidia A100 GPU
• SKU NCADSA100v4
• I had to request quota increase using Azure ML to achieve this experiment
• using open source data set
• Following this experiment from here
• Tune GPT2 to generate positive reviews

Code

• First install necesary packages

%load_ext autoreload
%autoreload 2

%pip install transformers trl

!pip install --upgrade bitsandbytes datasets accelerate loralib

• import libraries

import torch
from tqdm import tqdm
import pandas as pd

tqdm.pandas()from transformers import pipeline, AutoTokenizer
from datasets import load_datasetfrom trl import PPOTrainer, PPOConfig, AutoModelForCausalLMWithValueHead
from trl.core import LengthSampler

• Cconfigure Reinforcement learning PPO trainer

config = PPOConfig(
    model_name="lvwerra/gpt2-imdb",
    learning_rate=1.41e-5,
    #log_with="wandb",
)

sent_kwargs = {"return_all_scores": True, "function_to_apply": "none", "batch_size": 16}

• Load the data set

def build_dataset(config, dataset_name="imdb", 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.
    """
    tokenizer = AutoTokenizer.from_pretrained(config.model_name)
    tokenizer.pad_token = tokenizer.eos_token
    # load imdb with datasets
    ds = load_dataset(dataset_name, split="train")
    ds = ds.rename_columns({"text": "review"})
    ds = ds.filter(lambda x: len(x["review"]) > 200, batched=False)    input_size = LengthSampler(input_min_text_length, input_max_text_length)    def tokenize(sample):
        sample["input_ids"] = tokenizer.encode(sample["review"])[: input_size()]
        sample["query"] = tokenizer.decode(sample["input_ids"])
        return sample    ds = ds.map(tokenize, batched=False)
    ds.set_format(type="torch")
    return ds

dataset = build_dataset(config)

def collator(data):
    return dict((key, [d[key] for d in data]) for key in data[0])

• Load pretrained GPT2 model

model = AutoModelForCausalLMWithValueHead.from_pretrained(config.model_name)
ref_model = AutoModelForCausalLMWithValueHead.from_pretrained(config.model_name)
tokenizer = AutoTokenizer.from_pretrained(config.model_name)

tokenizer.pad_token = tokenizer.eos_token

• Initial PPOTainer

ppo_trainer = PPOTrainer(config, model, ref_model, tokenizer, dataset=dataset, data_collator=collator)

• Load bert classifier

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
sentiment_pipe = pipeline("sentiment-analysis", model="lvwerra/distilbert-imdb", device=device)

text = "this movie was really bad!!"
sentiment_pipe(text, **sent_kwargs)

• Generation Settings

gen_kwargs = {"min_length": -1, "top_k": 0.0, "top_p": 1.0, "do_sample": True, "pad_token_id": tokenizer.eos_token_id}

• Optimize Model
• Training loop will take about 2 hours to complete.

output_min_length = 4
output_max_length = 16
output_length_sampler = LengthSampler(output_min_length, output_max_length)

generation_kwargs = {
    "min_length": -1,
    "top_k": 0.0,
    "top_p": 1.0,
    "do_sample": True,
    "pad_token_id": tokenizer.eos_token_id,
}
for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
    query_tensors = batch["input_ids"]    #### Get response from gpt2
    response_tensors = []
    for query in query_tensors:
        gen_len = output_length_sampler()
        generation_kwargs["max_new_tokens"] = gen_len
        response = ppo_trainer.generate(query, **generation_kwargs)
        response_tensors.append(response.squeeze()[-gen_len:])
    batch["response"] = [tokenizer.decode(r.squeeze()) for r in response_tensors]    #### Compute sentiment score
    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]    #### Run PPO step
    stats = ppo_trainer.step(query_tensors, response_tensors, rewards)
    ppo_trainer.log_stats(stats, batch, rewards)

• Model Inspection

#### get a batch from the dataset
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 = [], []#### get response from gpt2 and gpt2_ref
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, **gen_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, **gen_kwargs
    ).squeeze()[-gen_len:]
    response_tensors.append(output)#### decode responses
game_data["response (before)"] = [tokenizer.decode(response_tensors_ref[i]) for i in range(bs)]
game_data["response (after)"] = [tokenizer.decode(response_tensors[i]) for i in range(bs)]#### sentiment analysis of query/response pairs before/after
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)]# store results in a dataframe
df_results = pd.DataFrame(game_data)
df_results

• Print Results

print("mean:")
display(df_results[["rewards (before)", "rewards (after)"]].mean())
print()
print("median:")
display(df_results[["rewards (before)", "rewards (after)"]].median())

• Save model

model.save_pretrained("gpt2-imdb-pos-v2")

Original article — Samples2023/AzureML/finetunetrl.md at main · balakreshnan/Samples2023 (github.com)

WRITER at MLearning.ai / Premier AI Video / AI art Copyright