train_txt2sql.py

import json
import random
from typing import List, Dict, Tuple
from adaptor.evaluators.generative import ROUGE, BLEU
from adaptor.lang_module import LangModule
from adaptor.objectives.seq2seq import Sequence2Sequence
from adaptor.utils import AdaptationArguments, StoppingStrategy
from adaptor.schedules import ParallelSchedule
from adaptor.adapter import Adapter
import wandb

# Dataset creation

## Define paths to JSON files
db_path = 'db_schemas.json'
spider_dataset_train_path = 'spider/train_spider.json'
spider_dataset_dev_path = 'spider/dev.json'
spider_syn_train_path = 'Spider-Syn/train_spider.json'
spider_syn_dev_path = 'Spider-Syn/dev.json'

## Open files
with open(db_path, 'r') as file_db:
    database_schemas = json.load(file_db)

with open(spider_dataset_train_path, 'r') as file_spider:
    spider_train_dataset = json.load(file_spider)

with open(spider_dataset_dev_path, 'r') as file_spider:
    spider_dev_dataset = json.load(file_spider)

with open(spider_syn_train_path, 'r') as file_spider:
    spider_syn_train_dataset = json.load(file_spider)

with open(spider_syn_dev_path, 'r') as file_spider:
    spider_syn_dev_dataset = json.load(file_spider)

## Include spider questions with synonyms (questions include text which is not in DB columns)
spider_train_dataset.extend([question for question in spider_syn_train_dataset if question['SpiderQuestion']!=question['SpiderSynQuestion']])
spider_dev_dataset.extend([question for question in spider_syn_dev_dataset if question['SpiderQuestion']!=question['SpiderSynQuestion']])

random.shuffle(spider_train_dataset)
random.shuffle(spider_dev_dataset)

def create_prompt(question: str, schema: str) -> str:
    return " ".join(["Question: ",question, "Schema:", schema])

def create_vals_and_labels(dataset: List[dict], db_dict: Dict[str, str]) -> Tuple[List[str], List[str]]:
    list_labels = [data["query"] for data in dataset]
    list_prompts = [create_prompt(data["question"], db_dict[data["db_id"]]) 
                    if "question" in data else create_prompt(data["SpiderSynQuestion"], db_dict[data["db_id"]]) for data in dataset]
    return list_prompts, list_labels

## Training prompts and labels
prompts_train, labels_train = create_vals_and_labels(spider_train_dataset, database_schemas)
assert len(prompts_train)==len(labels_train)

## Evaluation prompts and labels
prompts_eval, labels_eval = create_vals_and_labels(spider_dev_dataset, database_schemas)
assert len(prompts_eval)==len(labels_eval)

# Training

lang_module = LangModule("google/t5-large-lm-adapt")
evaluators = [BLEU(), ROUGE(decides_convergence=True)]

wandb.init(project="chatbot")

seq_qa = Sequence2Sequence(lang_module,
                           texts_or_path=prompts_train,
                           labels_or_path=labels_train,
                           val_texts_or_path=prompts_eval,
                           val_labels_or_path=labels_eval,
                           batch_size=4,
                           val_evaluators=evaluators,
                           objective_id="txt2SQL_Spider")

training_arguments = AdaptationArguments(output_dir="checkpoints-txt2sql",
                                         learning_rate=5e-5,
                                         stopping_strategy=StoppingStrategy.ALL_OBJECTIVES_CONVERGED,
                                         stopping_patience=8,
                                         save_total_limit=8,
                                         do_train=True,
                                         do_eval=True,
                                         bf16=True,
                                         warmup_steps=100,
                                         gradient_accumulation_steps=8,
                                         logging_steps=10,
                                         eval_steps=200,
                                         save_steps=200,
                                         num_train_epochs=10,
                                         evaluation_strategy="steps")


schedule = ParallelSchedule(objectives=[seq_qa],
                            args=training_arguments)
adapter = Adapter(lang_module, schedule, args=training_arguments)


adapter.train()