eval_nli.py

# -*- coding: utf-8 -*-

import sys
import os
import logging
import fcntl
import time
import argparse
logging.basicConfig(format='%(asctime)s : %(message)s', level=logging.DEBUG)

import torch
from prettytable import PrettyTable
from transformers import AutoTokenizer
from angle_emb import AnglE



# Import SentEval
sys.path.insert(0, './SentEval')
import senteval


PATH_TO_DATA = './SentEval/data'


def print_table(task_names, scores):
    tb = PrettyTable()
    tb.field_names = task_names
    tb.add_row(scores)
    print(tb)


def lock_and_write_file(file_path, content):
    with open(file_path, 'a') as file:
        while True:
            try:
                # Acquire an exclusive lock (non-blocking)
                fcntl.flock(file, fcntl.LOCK_EX | fcntl.LOCK_NB)

                # Perform your write operations here
                file.write(content + '\n')
                file.flush()

            except IOError as e:
                print("File is locked by another process. Can't write.")
                time.sleep(1)
            finally:
                # Release the lock
                fcntl.flock(file, fcntl.LOCK_UN)
                break


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--model_name_or_path", type=str,
                        help="Transformers' model name or path", required=True)
    parser.add_argument('--prompt', type=str, default='Summarize sentence "{text}" in one word:"')
    parser.add_argument("--pooling_strategy", type=str, required=True)
    parser.add_argument("--is_llm", type=int, choices=[0, 1], default=0)
    parser.add_argument("--max_length", type=int, default=512,
                        help="max length")
    parser.add_argument("--mode", type=str,
                        choices=['dev', 'test', 'fasttest'],
                        default='test',
                        help="What evaluation mode to use (dev: fast mode, dev results; test: full mode, test results); fasttest: fast mode, test results")
    parser.add_argument("--task_set", type=str,
                        choices=['sts', 'transfer', 'full', 'na'],
                        default='sts',
                        help="What set of tasks to evaluate on. If not 'na', this will override '--tasks'")
    parser.add_argument('--lora_name_or_path', type=str, default=None)
    parser.add_argument('--pretrained_model_path', type=str, default=None)


    args = parser.parse_args()

    if args.is_llm:
        model = AnglE.from_pretrained(
            args.model_name_or_path,
            pretrained_lora_path=args.lora_name_or_path,
            pooling_strategy=args.pooling_strategy,
            torch_dtype=torch.bfloat16,
            is_llm=True).cuda()
    else:
        args.prompt = None
        model = AnglE.from_pretrained(
            args.model_name_or_path,
            pretrained_model_path=args.pretrained_model_path,
            pooling_strategy=args.pooling_strategy).cuda()
    print('>>> prompt:', args.prompt)

    tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)

    # Set up the tasks
    if args.task_set == 'sts':
        args.tasks = ['STS12', 'STS13', 'STS14', 'STS15', 'STS16', 'STSBenchmark', 'SICKRelatedness']
        if args.mode == 'dev':
            args.tasks = ['STSBenchmark-dev']
    elif args.task_set == 'transfer':
        args.tasks = ['MR', 'CR', 'MPQA', 'SUBJ', 'SST2', 'TREC', 'MRPC']
    elif args.task_set == 'full':
        args.tasks = ['STS12', 'STS13', 'STS14', 'STS15', 'STS16', 'STSBenchmark', 'SICKRelatedness']
        args.tasks += ['MR', 'CR', 'MPQA', 'SUBJ', 'SST2', 'TREC', 'MRPC']

    # Set params for SentEval
    if args.mode == 'dev' or args.mode == 'fasttest':
        # Fast mode
        params = {'task_path': PATH_TO_DATA, 'usepytorch': True, 'kfold': 5, 'batch_size': 32}
        params['classifier'] = {'nhid': 0, 'optim': 'rmsprop', 'batch_size': 32,
                                         'tenacity': 3, 'epoch_size': 2}
    elif args.mode == 'test':
        # Full mode
        params = {'task_path': PATH_TO_DATA, 'usepytorch': True, 'kfold': 10, 'batch_size': 2}
        params['classifier'] = {'nhid': 0, 'optim': 'adam', 'batch_size': 64,
                                         'tenacity': 5, 'epoch_size': 4}
    else:
        raise NotImplementedError

    # SentEval prepare and batcher
    def prepare(params, samples):
        return

    def batcher(params, batch, max_length=None):
        # Handle rare token encoding issues in the dataset
        if len(batch) >= 1 and len(batch[0]) >= 1 and isinstance(batch[0][0], bytes):
            batch = [[word.decode('utf-8') for word in s] for s in batch]

        sentences = [' '.join(s) for s in batch]
        if max_length == 500:
            sentences = [tokenizer.decode(tokenizer.encode(s, add_special_tokens=False)[:max_length]) for s in sentences]
            max_length = 512

        if args.prompt is not None:
            for i, s in enumerate(sentences):
                if len(s) > 0 and s[-1] not in '.?"\'': s += '.'
                s = s.replace('"', '\'')
                if len(s) > 0 and '?' == s[-1]: s = s[:-1] + '.'
                sentences[i] = args.prompt.format(text=s)

        return model.encode(sentences, to_numpy=True, max_length=args.max_length)

    results = {}
    for task in args.tasks:
        se = senteval.engine.SE(params, batcher, prepare)
        result = se.eval(task)
        results[task] = result

    # Print evaluation results
    if args.mode == 'dev':
        print("------ %s ------" % (args.mode))

        task_names = []
        scores = []
        for task in ['STSBenchmark-dev']:
            task_names.append(task)
            if task in results:
                scores.append("%.2f" % (results[task]['dev']['spearman'][0] * 100))
            else:
                scores.append("0.00")
        print_table(task_names, scores)

        task_names = []
        scores = []
        for task in ['MR', 'CR', 'SUBJ', 'MPQA', 'SST2', 'TREC', 'MRPC']:
            task_names.append(task)
            if task in results:
                scores.append("%.2f" % (results[task]['devacc']))    
            else:
                scores.append("0.00")
        task_names.append("Avg.")
        scores.append("%.2f" % (sum([float(score) for score in scores]) / len(scores)))
        print_table(task_names, scores)


    elif args.mode == 'test' or args.mode == 'fasttest':
        print("------ %s ------" % (args.mode))

        task_names = []
        scores = []
        for task in ['STS12', 'STS13', 'STS14', 'STS15', 'STS16', 'STSBenchmark', 'SICKRelatedness']:
            task_names.append(task)
            if task in results:
                if task in ['STS12', 'STS13', 'STS14', 'STS15', 'STS16']:
                    scores.append("%.2f" % (results[task]['all']['spearman']['all'] * 100))
                else:
                    scores.append("%.2f" % (results[task]['test']['spearman'].correlation * 100))
            else:
                scores.append("0.00")
        task_names.append("Avg.")
        scores.append("%.2f" % (sum([float(score) for score in scores]) / len(scores)))
        print_table(task_names, scores)
        #
        # write results and template to file
        if args.prompt is not None and args.task_set != 'transfer':
            with open('./sts-org-results', 'a') as f:
                model_name = args.model_name_or_path.split('/')[-1]
                f.write(args.prompt.replace(' ', '_') + ' ' + model_name + ' ' + ' '.join([str(s) for s in scores]) + '\n')

        task_names = []
        scores = []
        for task in ['MR', 'CR', 'SUBJ', 'MPQA', 'SST2', 'TREC', 'MRPC']:
            task_names.append(task)
            if task in results:
                scores.append("%.2f" % (results[task]['acc']))
            else:
                scores.append("0.00")
        task_names.append("Avg.")
        scores.append("%.2f" % (sum([float(score) for score in scores]) / len(scores)))
        print_table(task_names, scores)


if __name__ == "__main__":
    main()