From 255421fe3943ebfa92be2640a4cf08309f2c7c27 Mon Sep 17 00:00:00 2001 From: ldwang Date: Sun, 29 Sep 2024 10:46:50 +0800 Subject: [PATCH] Add CCI3 scripts. Signed-off-by: ldwang --- examples/CCI3/README.md | 0 examples/CCI3/qwen2_api.py | 104 ++ examples/CCI3/run_classification.py | 1032 ++++++++++++++++++ examples/CCI3/run_classification_trainval.sh | 88 ++ 4 files changed, 1224 insertions(+) create mode 100644 examples/CCI3/README.md create mode 100644 examples/CCI3/qwen2_api.py create mode 100644 examples/CCI3/run_classification.py create mode 100644 examples/CCI3/run_classification_trainval.sh diff --git a/examples/CCI3/README.md b/examples/CCI3/README.md new file mode 100644 index 00000000..e69de29b diff --git a/examples/CCI3/qwen2_api.py b/examples/CCI3/qwen2_api.py new file mode 100644 index 00000000..12a80bfb --- /dev/null +++ b/examples/CCI3/qwen2_api.py @@ -0,0 +1,104 @@ +# !/usr/bin/env python +# -*- coding:utf-8 -*- + +import os +import openai +import json +import time +import random +import os +import requests +import base64 + +prompt_template = '''Below is an extract from a web page. Evaluate whether the page has a high educational value and could be useful in an educational setting for teaching from primary school to college levels using the additive 5-point scoring system described below. Points are accumulated based on the satisfaction of each criterion: + +- Add 1 point if the extract provides some basic information relevant to educational topics, even if it includes some irrelevant or non-academic content like advertisements and promotional material. +- Add another point if the extract addresses certain elements pertinent to education but does not align closely with educational standards. It might mix educational content with non-educational material, offering a superficial overview of potentially useful topics, or presenting information in a disorganized manner and incoherent writing style. +- Award a third point if the extract is appropriate for educational use and introduces key concepts relevant to school curricula. It is coherent though it may not be comprehensive or could include some extraneous information. It may resemble an introductory section of a textbook or a basic tutorial that is suitable for learning but has notable limitations like treating concepts that are too complex for grade school students. +- Grant a fourth point if the extract highly relevant and beneficial for educational purposes for a level not higher than grade school, exhibiting a clear and consistent writing style. It could be similar to a chapter from a textbook or a tutorial, offering substantial educational content, including exercises and solutions, with minimal irrelevant information, and the concepts aren't too advanced for grade school students. The content is coherent, focused, and valuable for structured learning. +- Bestow a fifth point if the extract is outstanding in its educational value, perfectly suited for teaching either at primary school or college. It follows detailed reasoning, the writing style is easy to follow and offers profound and thorough insights into the subject matter, devoid of any non-educational or complex content. + +The extract: +<{EXAMPLE}>. + +After examining the extract: +- Briefly justify your total score, up to 100 words. +- Conclude with the score using the format: "Educational score: "''' + +def get_from_gpt4(ques, engine="gpt-4"): + uer_prompt = prompt_template.format(EXAMPLE=ques) + messages = [] + messages.append({"role": "user", "content": uer_prompt}) + + openai_api_key = "EMPTY" + openai_api_base = "http://localhost:8000/v1" + openai_api_base = "http://120.92.91.62:9301/v1" + + from openai import OpenAI + client = OpenAI( + api_key=openai_api_key, + base_url=openai_api_base, + ) + + models = client.models.list() + model = models.data[0].id + + rsp = None + try: + rsp = client.chat.completions.create( + model=model, + messages=messages, + temperature=0.7, + max_tokens=500, + top_p=0.95, + frequency_penalty=0, + presence_penalty=0, + stop=None) + except: + import traceback + traceback.print_exc() + pass + #return rsp["choices"][0]["message"]["content"] + return rsp + +def parse_score_from_review(review): + try: + import re + match = re.search(r'Educational score:\s+(\d+\.?\d*)', review) + if match: + return float(match.group(1)) + else: + print(f"Failed to parse scores from {review}") + return -1 + except: + print(f"Failed to parse scores from {review}") + return -1 + +if __name__ == "__main__": + import argparse + parser = argparse.ArgumentParser() + parser.add_argument('--file-path', type=str, default="", help="file path", required=True) + args = parser.parse_args() + + import jsonlines + file_path = args.file_path + output_file_path = f'{file_path}.jsonl.qwen' + writer = jsonlines.open(output_file_path, mode='w') + with jsonlines.open(file_path) as reader: + for line in reader: + ques = line['text'] + score = -1.0 + response = get_from_gpt4(ques, engine='Qwen2-72B-Instruct') + content = '' + print(response) + if response is None: + pass + else: + for choice in response.choices: + content = choice.message.content + score = parse_score_from_review(content) + break + line['score'] = score + line['content'] = content + writer.write(line) + diff --git a/examples/CCI3/run_classification.py b/examples/CCI3/run_classification.py new file mode 100644 index 00000000..d254003b --- /dev/null +++ b/examples/CCI3/run_classification.py @@ -0,0 +1,1032 @@ +# !/usr/bin/env python +# -*- coding:utf-8 -*- +# coding=utf-8 +# Copyright 2020 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" Finetuning the library models for text classification.""" +# You can also adapt this script on your own text classification task. Pointers for this are left as comments. + +import logging +import os +import random +import sys +import warnings +from dataclasses import dataclass, field +from typing import List, Optional + +import datasets +import evaluate +import numpy as np +from datasets import Value, load_dataset + +import transformers +from transformers import ( + AutoConfig, + AutoModelForSequenceClassification, + AutoTokenizer, + DataCollatorWithPadding, + EvalPrediction, + HfArgumentParser, + Trainer, + TrainingArguments, + default_data_collator, + set_seed, +) +from transformers.trainer_utils import get_last_checkpoint +from transformers.utils import check_min_version, send_example_telemetry +from transformers.utils.versions import require_version + + +# Will error if the minimal version of Transformers is not installed. Remove at your own risks. +# check_min_version("4.41.0.dev0") + +require_version( + "datasets>=1.8.0", + "To fix: pip install -r examples/pytorch/text-classification/requirements.txt", +) + + +logger = logging.getLogger(__name__) + + +@dataclass +class DataTrainingArguments: + """ + Arguments pertaining to what data we are going to input our model for training and eval. + + Using `HfArgumentParser` we can turn this class + into argparse arguments to be able to specify them on + the command line. + """ + + dataset_name: Optional[str] = field( + default=None, + metadata={"help": "The name of the dataset to use (via the datasets library)."}, + ) + dataset_config_name: Optional[str] = field( + default=None, + metadata={ + "help": "The configuration name of the dataset to use (via the datasets library)." + }, + ) + do_regression: bool = field( + default=None, + metadata={ + "help": "Whether to do regression instead of classification. If None, will be inferred from the dataset." + }, + ) + text_column_names: Optional[str] = field( + default=None, + metadata={ + "help": ( + "The name of the text column in the input dataset or a CSV/JSON file. " + 'If not specified, will use the "sentence" column for single/multi-label classification task.' + ) + }, + ) + text_column_delimiter: Optional[str] = field( + default=" ", + metadata={ + "help": "THe delimiter to use to join text columns into a single sentence." + }, + ) + train_split_name: Optional[str] = field( + default=None, + metadata={ + "help": 'The name of the train split in the input dataset. If not specified, will use the "train" split when do_train is enabled' + }, + ) + validation_split_name: Optional[str] = field( + default=None, + metadata={ + "help": 'The name of the validation split in the input dataset. If not specified, will use the "validation" split when do_eval is enabled' + }, + ) + test_split_name: Optional[str] = field( + default=None, + metadata={ + "help": 'The name of the test split in the input dataset. If not specified, will use the "test" split when do_predict is enabled' + }, + ) + remove_splits: Optional[str] = field( + default=None, + metadata={ + "help": "The splits to remove from the dataset. Multiple splits should be separated by commas." + }, + ) + remove_columns: Optional[str] = field( + default=None, + metadata={ + "help": "The columns to remove from the dataset. Multiple columns should be separated by commas." + }, + ) + label_column_name: Optional[str] = field( + default=None, + metadata={ + "help": ( + "The name of the label column in the input dataset or a CSV/JSON file. " + 'If not specified, will use the "label" column for single/multi-label classification task' + ) + }, + ) + max_seq_length: int = field( + default=128, + metadata={ + "help": ( + "The maximum total input sequence length after tokenization. Sequences longer " + "than this will be truncated, sequences shorter will be padded." + ) + }, + ) + overwrite_cache: bool = field( + default=False, + metadata={"help": "Overwrite the cached preprocessed datasets or not."}, + ) + pad_to_max_length: bool = field( + default=True, + metadata={ + "help": ( + "Whether to pad all samples to `max_seq_length`. " + "If False, will pad the samples dynamically when batching to the maximum length in the batch." + ) + }, + ) + shuffle_train_dataset: bool = field( + default=False, metadata={"help": "Whether to shuffle the train dataset or not."} + ) + shuffle_seed: int = field( + default=42, + metadata={ + "help": "Random seed that will be used to shuffle the train dataset." + }, + ) + max_train_samples: Optional[int] = field( + default=None, + metadata={ + "help": ( + "For debugging purposes or quicker training, truncate the number of training examples to this " + "value if set." + ) + }, + ) + max_eval_samples: Optional[int] = field( + default=None, + metadata={ + "help": ( + "For debugging purposes or quicker training, truncate the number of evaluation examples to this " + "value if set." + ) + }, + ) + max_predict_samples: Optional[int] = field( + default=None, + metadata={ + "help": ( + "For debugging purposes or quicker training, truncate the number of prediction examples to this " + "value if set." + ) + }, + ) + metric_name: Optional[str] = field( + default=None, metadata={"help": "The metric to use for evaluation."} + ) + train_file: Optional[str] = field( + default=None, + metadata={"help": "A csv or a json file containing the training data."}, + ) + validation_file: Optional[str] = field( + default=None, + metadata={"help": "A csv or a json file containing the validation data."}, + ) + test_file: Optional[str] = field( + default=None, + metadata={"help": "A csv or a json file containing the test data."}, + ) + + def __post_init__(self): + if self.dataset_name is None: + if self.train_file is None or self.validation_file is None: + raise ValueError(" training/validation file or a dataset name.") + + train_extension = self.train_file.split(".")[-1] + assert train_extension in [ + "csv", + "json", + "jsonl", + ], "`train_file` should be a csv or a json file." + validation_extension = self.validation_file.split(".")[-1] + assert ( + validation_extension == train_extension + ), "`validation_file` should have the same extension (csv or json) as `train_file`." + + +@dataclass +class ModelArguments: + """ + Arguments pertaining to which model/config/tokenizer we are going to fine-tune from. + """ + + model_name_or_path: str = field( + metadata={ + "help": "Path to pretrained model or model identifier from huggingface.co/models" + } + ) + config_name: Optional[str] = field( + default=None, + metadata={ + "help": "Pretrained config name or path if not the same as model_name" + }, + ) + tokenizer_name: Optional[str] = field( + default=None, + metadata={ + "help": "Pretrained tokenizer name or path if not the same as model_name" + }, + ) + cache_dir: Optional[str] = field( + default=None, + metadata={ + "help": "Where do you want to store the pretrained models downloaded from huggingface.co" + }, + ) + use_fast_tokenizer: bool = field( + default=True, + metadata={ + "help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not." + }, + ) + model_revision: str = field( + default="main", + metadata={ + "help": "The specific model version to use (can be a branch name, tag name or commit id)." + }, + ) + token: str = field( + default=None, + metadata={ + "help": ( + "The token to use as HTTP bearer authorization for remote files. If not specified, will use the token " + "generated when running `huggingface-cli login` (stored in `~/.huggingface`)." + ) + }, + ) + use_auth_token: bool = field( + default=None, + metadata={ + "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token` instead." + }, + ) + trust_remote_code: bool = field( + default=False, + metadata={ + "help": ( + "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option " + "should only be set to `True` for repositories you trust and in which you have read the code, as it will " + "execute code present on the Hub on your local machine." + ) + }, + ) + ignore_mismatched_sizes: bool = field( + default=False, + metadata={ + "help": "Will enable to load a pretrained model whose head dimensions are different." + }, + ) + + freeze_backbone: int = field( + default=0, metadata={"help": "Whether to freeze the backbone or not."} + ) + model_name: str = field( + default="xx", metadata={"help": "Whether to freeze the backbone or not."} + ) + + +def get_parameter_number(model): + total_num = sum(p.numel() for p in model.parameters()) + trainable_num = sum(p.numel() for p in model.parameters() if p.requires_grad) + logger.info( + f"Total param: {total_num}, trainable num is: {trainable_num}, trainable ratio is: {trainable_num/total_num*100.0}%" + ) + + +def get_label_list(raw_dataset, split="train") -> List[str]: + """Get the list of labels from a multi-label dataset""" + + if isinstance(raw_dataset[split]["label"][0], list): + label_list = [ + label for sample in raw_dataset[split]["label"] for label in sample + ] + label_list = list(set(label_list)) + else: + label_list = raw_dataset[split].unique("label") + # we will treat the label list as a list of string instead of int, consistent with model.config.label2id + label_list = [str(label) for label in label_list] + logger.info(f"label_list: {label_list}") + return label_list + + +def main(): + # See all possible arguments in src/transformers/training_args.py + # or by passing the --help flag to this script. + # We now keep distinct sets of args, for a cleaner separation of concerns. + + parser = HfArgumentParser( + (ModelArguments, DataTrainingArguments, TrainingArguments) + ) + if len(sys.argv) == 2 and sys.argv[1].endswith(".json"): + # If we pass only one argument to the script and it's the path to a json file, + # let's parse it to get our arguments. + model_args, data_args, training_args = parser.parse_json_file( + json_file=os.path.abspath(sys.argv[1]) + ) + else: + model_args, data_args, training_args = parser.parse_args_into_dataclasses() + + if model_args.use_auth_token is not None: + warnings.warn( + "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token` instead.", + FutureWarning, + ) + if model_args.token is not None: + raise ValueError( + "`token` and `use_auth_token` are both specified. Please set only the argument `token`." + ) + model_args.token = model_args.use_auth_token + + # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The + # information sent is the one passed as arguments along with your Python/PyTorch versions. + send_example_telemetry("run_classification", model_args, data_args) + + # Setup logging + logging.basicConfig( + format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", + datefmt="%m/%d/%Y %H:%M:%S", + handlers=[logging.StreamHandler(sys.stdout)], + ) + + if training_args.should_log: + # The default of training_args.log_level is passive, so we set log level at info here to have that default. + transformers.utils.logging.set_verbosity_info() + + log_level = training_args.get_process_log_level() + logger.setLevel(log_level) + datasets.utils.logging.set_verbosity(log_level) + transformers.utils.logging.set_verbosity(log_level) + transformers.utils.logging.enable_default_handler() + transformers.utils.logging.enable_explicit_format() + + # Log on each process the small summary: + logger.warning( + f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, " + + f"distributed training: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}" + ) + logger.info(f"Training/evaluation parameters {training_args}") + logger.info(f"model_args parameters {model_args}") + logger.info(f"data_args parameters {data_args}") + + # Detecting last checkpoint. + last_checkpoint = None + if ( + os.path.isdir(training_args.output_dir) + and training_args.do_train + and not training_args.overwrite_output_dir + ): + last_checkpoint = get_last_checkpoint(training_args.output_dir) + if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0: + raise ValueError( + f"Output directory ({training_args.output_dir}) already exists and is not empty. " + "Use --overwrite_output_dir to overcome." + ) + elif ( + last_checkpoint is not None and training_args.resume_from_checkpoint is None + ): + logger.info( + f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change " + "the `--output_dir` or add `--overwrite_output_dir` to train from scratch." + ) + + # Set seed before initializing model. + set_seed(training_args.seed) + + # Get the datasets: you can either provide your own CSV/JSON training and evaluation files, or specify a dataset name + # to load from huggingface/datasets. In ether case, you can specify a the key of the column(s) containing the text and + # the key of the column containing the label. If multiple columns are specified for the text, they will be joined together + # for the actual text value. + # In distributed training, the load_dataset function guarantee that only one local process can concurrently + # download the dataset. + if data_args.dataset_name is not None: + # Downloading and loading a dataset from the hub. + raw_datasets = load_dataset( + data_args.dataset_name, + data_args.dataset_config_name, + cache_dir=model_args.cache_dir, + token=model_args.token, + ) + # Try print some info about the dataset + logger.info(f"Dataset loaded: {raw_datasets}") + logger.info(raw_datasets) + else: + # Loading a dataset from your local files. + # CSV/JSON training and evaluation files are needed. + data_files = { + "train": data_args.train_file, + "validation": data_args.validation_file, + } + + # Get the test dataset: you can provide your own CSV/JSON test file + if training_args.do_predict: + if data_args.test_file is not None: + train_extension = data_args.train_file.split(".")[-1] + test_extension = data_args.test_file.split(".")[-1] + assert ( + test_extension == train_extension + ), "`test_file` should have the same extension (csv or json) as `train_file`." + data_files["test"] = data_args.test_file + else: + raise ValueError( + "Need either a dataset name or a test file for `do_predict`." + ) + + for key in data_files.keys(): + logger.info(f"load a local file for {key}: {data_files[key]}") + + if data_args.train_file.endswith(".csv"): + # Loading a dataset from local csv files + raw_datasets = load_dataset( + "csv", + data_files=data_files, + cache_dir=model_args.cache_dir, + token=model_args.token, + ) + else: + # Loading a dataset from local json files + raw_datasets = load_dataset( + "json", + data_files=data_files, + cache_dir=model_args.cache_dir, + token=model_args.token, + ) + + # See more about loading any type of standard or custom dataset at + # https://huggingface.co/docs/datasets/loading_datasets. + + if data_args.remove_splits is not None: + for split in data_args.remove_splits.split(","): + logger.info(f"removing split {split}") + raw_datasets.pop(split) + + if data_args.train_split_name is not None: + logger.info(f"using {data_args.train_split_name} as train set") + raw_datasets["train"] = raw_datasets[data_args.train_split_name] + raw_datasets.pop(data_args.train_split_name) + + if data_args.validation_split_name is not None: + logger.info(f"using {data_args.validation_split_name} as validation set") + raw_datasets["validation"] = raw_datasets[data_args.validation_split_name] + raw_datasets.pop(data_args.validation_split_name) + + if data_args.test_split_name is not None: + logger.info(f"using {data_args.test_split_name} as test set") + raw_datasets["test"] = raw_datasets[data_args.test_split_name] + raw_datasets.pop(data_args.test_split_name) + + if data_args.remove_columns is not None: + for split in raw_datasets.keys(): + for column in data_args.remove_columns.split(","): + logger.info(f"removing column {column} from split {split}") + raw_datasets[split] = raw_datasets[split].remove_columns(column) + + if ( + data_args.label_column_name is not None + and data_args.label_column_name != "label" + ): + for key in raw_datasets.keys(): + raw_datasets[key] = raw_datasets[key].rename_column( + data_args.label_column_name, "label" + ) + + # Trying to have good defaults here, don't hesitate to tweak to your needs. + + is_regression = ( + raw_datasets["train"].features["label"].dtype in ["float32", "float64"] + if data_args.do_regression is None + else data_args.do_regression + ) + + is_multi_label = False + if is_regression: + label_list = None + num_labels = 1 + # regession requires float as label type, let's cast it if needed + for split in raw_datasets.keys(): + if raw_datasets[split].features["label"].dtype not in [ + "float32", + "float64", + ]: + logger.warning( + f"Label type for {split} set to float32, was {raw_datasets[split].features['label'].dtype}" + ) + features = raw_datasets[split].features + features.update({"label": Value("float32")}) + try: + raw_datasets[split] = raw_datasets[split].cast(features) + except TypeError as error: + logger.error( + f"Unable to cast {split} set to float32, please check the labels are correct, or maybe try with --do_regression=False" + ) + raise error + + else: # classification + if ( + raw_datasets["train"].features["label"].dtype == "list" + ): # multi-label classification + is_multi_label = True + logger.info("Label type is list, doing multi-label classification") + # Trying to find the number of labels in a multi-label classification task + # We have to deal with common cases that labels appear in the training set but not in the validation/test set. + # So we build the label list from the union of labels in train/val/test. + label_list = get_label_list(raw_datasets, split="train") + for split in ["validation", "test"]: + if split in raw_datasets: + val_or_test_labels = get_label_list(raw_datasets, split=split) + diff = set(val_or_test_labels).difference(set(label_list)) + if len(diff) > 0: + # add the labels that appear in val/test but not in train, throw a warning + logger.warning( + f"Labels {diff} in {split} set but not in training set, adding them to the label list" + ) + label_list += list(diff) + # if label is -1, we throw a warning and remove it from the label list + for label in label_list: + if label == -1: + logger.warning("Label -1 found in label list, removing it.") + label_list.remove(label) + + label_list.sort() + num_labels = len(label_list) + if num_labels <= 1: + raise ValueError("You need more than one label to do classification.") + + # Load pretrained model and tokenizer + # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently + # download model & vocab. + + config = AutoConfig.from_pretrained( + ( + model_args.config_name + if model_args.config_name + else model_args.model_name_or_path + ), + num_labels=num_labels, + finetuning_task="text-classification", + cache_dir=model_args.cache_dir, + revision=model_args.model_revision, + token=model_args.token, + trust_remote_code=model_args.trust_remote_code, + ) + # 回归任务不用dp，分类任务用dp + if is_regression: + if "bert" in model_args.model_name.lower() or "roberta" in model_args.model_name.lower(): + config.hidden_dropout_prob=0.0 + config.classifier_dropout=0.0 + else: + # if "bert" in model_args.model_name.lower() or "roberta" in model_args.model_name.lower(): + config.hidden_dropout_prob=0.0 + config.classifier_dropout=0.1 + + + + if is_regression: + config.problem_type = "regression" + logger.info("setting problem type to regression") + elif is_multi_label: + config.problem_type = "multi_label_classification" + logger.info("setting problem type to multi label classification") + else: + config.problem_type = "single_label_classification" + logger.info("setting problem type to single label classification") + + logger.info(f"config: {config}") + + tokenizer = AutoTokenizer.from_pretrained( + ( + model_args.tokenizer_name + if model_args.tokenizer_name + else model_args.model_name_or_path + ), + cache_dir=model_args.cache_dir, + use_fast=model_args.use_fast_tokenizer, + revision=model_args.model_revision, + token=model_args.token, + trust_remote_code=model_args.trust_remote_code, + ) + + logger.info(f"tokenizer: {tokenizer}") + + model = AutoModelForSequenceClassification.from_pretrained( + model_args.model_name_or_path, + from_tf=bool(".ckpt" in model_args.model_name_or_path), + config=config, + cache_dir=model_args.cache_dir, + revision=model_args.model_revision, + token=model_args.token, + trust_remote_code=model_args.trust_remote_code, + ignore_mismatched_sizes=model_args.ignore_mismatched_sizes + ) + + logger.info(f"model: {model}") + + if model_args.freeze_backbone > 0: + if "roberta" in model_args.model_name.lower(): + for param in model.roberta.embeddings.parameters(): + param.requires_grad = False + for param in model.roberta.encoder.parameters(): + param.requires_grad = False + elif "bert" in model_args.model_name.lower(): + for param in model.bert.embeddings.parameters(): + param.requires_grad = False + for param in model.bert.encoder.parameters(): + param.requires_grad = False + ## freeze之后qwen的可更新参数很少 + elif "qwen" in model_args.model_name: + for param in model.model.parameters(): + param.requires_grad = False + + get_parameter_number(model) + + # Padding strategy + if data_args.pad_to_max_length: + padding = "max_length" + else: + # We will pad later, dynamically at batch creation, to the max sequence length in each batch + padding = False + + # for training ,we will update the config with label infos, + # if do_train is not set, we will use the label infos in the config + if training_args.do_train and not is_regression: # classification, training + label_to_id = {v: i for i, v in enumerate(label_list)} + logger.info(f"label_to_id: {label_to_id}") + # update config with label infos + if model.config.label2id != label_to_id: + logger.warning( + "The label2id key in the model config.json is not equal to the label2id key of this " + "run. You can ignore this if you are doing finetuning." + ) + model.config.label2id = label_to_id + model.config.id2label = {id: label for label, id in label_to_id.items()} + logger.info(f"label_to_id: {model.config.label2id}") + logger.info(f"id2label: {model.config.id2label}") + + elif not is_regression: # classification, but not training + logger.info("using label infos in the model config") + logger.info("label2id: {}".format(model.config.label2id)) + label_to_id = model.config.label2id + else: # regression + label_to_id = None + + if data_args.max_seq_length > tokenizer.model_max_length: + logger.warning( + f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the " + f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}." + ) + max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length) + logger.info(f"max_seq_length: {max_seq_length}") + + def multi_labels_to_ids(labels: List[str]) -> List[float]: + ids = [0.0] * len(label_to_id) # BCELoss requires float as target type + for label in labels: + ids[label_to_id[label]] = 1.0 + return ids + + def preprocess_function(examples): + if data_args.text_column_names is not None: + text_column_names = data_args.text_column_names.split(",") + # join together text columns into "sentence" column + examples["sentence"] = examples[text_column_names[0]] + for column in text_column_names[1:]: + for i in range(len(examples[column])): + examples["sentence"][i] += ( + data_args.text_column_delimiter + examples[column][i] + ) + # Tokenize the texts + result = tokenizer( + examples["sentence"], + padding=padding, + max_length=max_seq_length, + truncation=True, + ) + if label_to_id is not None and "label" in examples: + if is_multi_label: + result["label"] = [multi_labels_to_ids(l) for l in examples["label"]] + else: + result["label"] = [ + (label_to_id[str(l)] if l != -1 else -1) for l in examples["label"] + ] + return result + + # Running the preprocessing pipeline on all the datasets + with training_args.main_process_first(desc="dataset map pre-processing"): + raw_datasets = raw_datasets.map( + preprocess_function, + batched=True, + load_from_cache_file=not data_args.overwrite_cache, + num_proc=32, + desc="Running tokenizer on dataset", + ) + + if training_args.do_train: + if "train" not in raw_datasets: + raise ValueError("--do_train requires a train dataset.") + train_dataset = raw_datasets["train"] + if data_args.shuffle_train_dataset: + logger.info("Shuffling the training dataset") + train_dataset = train_dataset.shuffle(seed=data_args.shuffle_seed) + if data_args.max_train_samples is not None: + max_train_samples = min(len(train_dataset), data_args.max_train_samples) + train_dataset = train_dataset.select(range(max_train_samples)) + + if training_args.do_eval: + if ( + "validation" not in raw_datasets + and "validation_matched" not in raw_datasets + ): + if "test" not in raw_datasets and "test_matched" not in raw_datasets: + raise ValueError( + "--do_eval requires a validation or test dataset if validation is not defined." + ) + else: + logger.warning( + "Validation dataset not found. Falling back to test dataset for validation." + ) + eval_dataset = raw_datasets["test"] + else: + eval_dataset = raw_datasets["validation"] + + if data_args.max_eval_samples is not None: + max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples) + eval_dataset = eval_dataset.select(range(max_eval_samples)) + + if training_args.do_predict or data_args.test_file is not None: + if "test" not in raw_datasets: + raise ValueError("--do_predict requires a test dataset") + predict_dataset = raw_datasets["test"] + # remove label column if it exists + if data_args.max_predict_samples is not None: + max_predict_samples = min( + len(predict_dataset), data_args.max_predict_samples + ) + predict_dataset = predict_dataset.select(range(max_predict_samples)) + + # Log a few random samples from the training set: + if training_args.do_train: + for index in random.sample(range(len(train_dataset)), 3): + logger.info(f"Sample {index} of the training set: {train_dataset[index]}.") + + if data_args.metric_name is not None: + metric = ( + evaluate.load( + data_args.metric_name, + config_name="multilabel", + cache_dir=model_args.cache_dir, + ) + if is_multi_label + else evaluate.load(data_args.metric_name, cache_dir=model_args.cache_dir) + ) + logger.info(f"Using metric {data_args.metric_name} for evaluation.") + else: + if is_regression: + metric = evaluate.load("mse", cache_dir=model_args.cache_dir) + logger.info( + "Using mean squared error (mse) as regression score, you can use --metric_name to overwrite." + ) + else: + if is_multi_label: + metric = evaluate.load( + "f1", config_name="multilabel", cache_dir=model_args.cache_dir + ) + logger.info( + "Using multilabel F1 for multi-label classification task, you can use --metric_name to overwrite." + ) + else: + metric = evaluate.load("accuracy", cache_dir=model_args.cache_dir) + logger.info( + "Using accuracy as classification score, you can use --metric_name to overwrite." + ) + + ''' + def compute_metrics(p: EvalPrediction): + preds = p.predictions[0] if isinstance(p.predictions, tuple) else p.predictions + if is_regression: + preds = np.squeeze(preds) + result = metric.compute(predictions=preds, references=p.label_ids) + elif is_multi_label: + preds = np.array( + [np.where(p > 0, 1, 0) for p in preds] + ) # convert logits to multi-hot encoding + # Micro F1 is commonly used in multi-label classification + result = metric.compute( + predictions=preds, references=p.label_ids, average="micro" + ) + else: + preds = np.argmax(preds, axis=1) + result = metric.compute(predictions=preds, references=p.label_ids) + if len(result) > 1: + result["combined_score"] = np.mean(list(result.values())).item() + return result + ''' + + def compute_metrics(eval_pred): + p = eval_pred + preds = p.predictions[0] if isinstance(p.predictions, tuple) else p.predictions + if is_regression: + preds = np.squeeze(preds) + result = metric.compute(predictions=preds, references=p.label_ids) + elif is_multi_label: + preds = np.array( + [np.where(p > 0, 1, 0) for p in preds] + ) # convert logits to multi-hot encoding + # Micro F1 is commonly used in multi-label classification + result = metric.compute( + predictions=preds, references=p.label_ids, average="micro" + ) + else: + preds = np.argmax(preds, axis=1) + result = metric.compute(predictions=preds, references=p.label_ids) + if len(result) > 1: + result["combined_score"] = np.mean(list(result.values())).item() + + precision_metric = evaluate.load("precision") + recall_metric = evaluate.load("recall") + f1_metric = evaluate.load("f1") + accuracy_metric = evaluate.load("accuracy") + + logits, labels = eval_pred + preds = np.round(logits.squeeze()).clip(0, 5).astype(int) + labels = np.round(labels.squeeze()).astype(int) + + precision = precision_metric.compute( + predictions=preds, references=labels, average="macro" + )["precision"] + recall = recall_metric.compute( + predictions=preds, references=labels, average="macro" + )["recall"] + f1 = f1_metric.compute(predictions=preds, references=labels, average="macro")["f1"] + accuracy = accuracy_metric.compute(predictions=preds, references=labels)["accuracy"] + + from sklearn.metrics import classification_report, confusion_matrix + report = classification_report(labels, preds) + print("Validation Report:\n" + report, flush=True) + cm = confusion_matrix(labels, preds) + print("Confusion Matrix:\n" + str(cm), flush=True) + + result['precision'] = precision + result['recall'] = recall + result['f1'] = f1 + result['accuracy'] = accuracy + + ## Binary: a threshold of 3 + logits, labels = eval_pred + labels = np.round(labels.squeeze()).astype(int) + labels = np.where(labels >= 3, 0, 1) + preds = np.where(logits.squeeze() >= 3, 0, 1).clip(0, 5).astype(int) + report = classification_report(labels, preds) + print("Binary Validation Report:\n" + report, flush=True) + matrix = confusion_matrix(labels, preds) + print("Binary Confusion Matrix:\n" + str(matrix), flush=True) + TP, FP, FN, TN = matrix[0][0], matrix[1][0], matrix[0][1], matrix[1][1] + precision = (TP) / (TP + FP) + recall = (TP) / (TP + FN) + f1_score = 2 * (precision * recall) / (recall + precision) + precision_negative = TN / (TN + FN) + recall_negative = TN / (TN + FP) + f1_score_negative = 2 * (precision_negative * recall_negative) / (precision_negative + recall_negative) + + result['pos_precision'] = precision + result['pos_recall'] = recall + result['pos_f1'] = f1_score + result['neg_precision'] = precision_negative + result['neg_recall'] = recall_negative + result['neg_f1'] = f1_score_negative + + return result + + # Data collator will default to DataCollatorWithPadding when the tokenizer is passed to Trainer, so we change it if + # we already did the padding. + if data_args.pad_to_max_length: + data_collator = default_data_collator + elif training_args.fp16: + data_collator = DataCollatorWithPadding( + tokenizer, padding=True, max_length=max_seq_length, pad_to_multiple_of=8 + ) + else: + data_collator = None + + # Initialize our Trainer + trainer = Trainer( + model=model, + args=training_args, + train_dataset=train_dataset if training_args.do_train else None, + eval_dataset=eval_dataset if training_args.do_eval else None, + compute_metrics=compute_metrics, + tokenizer=tokenizer, + data_collator=data_collator, + ) + + # Training + if training_args.do_train: + checkpoint = None + if training_args.resume_from_checkpoint is not None: + checkpoint = training_args.resume_from_checkpoint + elif last_checkpoint is not None: + checkpoint = last_checkpoint + train_result = trainer.train(resume_from_checkpoint=checkpoint) + metrics = train_result.metrics + max_train_samples = ( + data_args.max_train_samples + if data_args.max_train_samples is not None + else len(train_dataset) + ) + metrics["train_samples"] = min(max_train_samples, len(train_dataset)) + trainer.save_model() # Saves the tokenizer too for easy upload + trainer.log_metrics("train", metrics) + trainer.save_metrics("train", metrics) + trainer.save_state() + + # Evaluation + if training_args.do_eval: + logger.info("*** Evaluate ***") + metrics = trainer.evaluate(eval_dataset=eval_dataset) + logger.info(f"metrics: {metrics}") + max_eval_samples = ( + data_args.max_eval_samples + if data_args.max_eval_samples is not None + else len(eval_dataset) + ) + metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset)) + trainer.log_metrics("eval", metrics) + trainer.save_metrics("eval", metrics) + + if training_args.do_predict: + logger.info("*** Predict ***") + # Removing the `label` columns if exists because it might contains -1 and Trainer won't like that. + if "label" in predict_dataset.features: + predict_dataset = predict_dataset.remove_columns("label") + predictions = trainer.predict( + predict_dataset, metric_key_prefix="predict" + ).predictions + if is_regression: + predictions = np.squeeze(predictions) + elif is_multi_label: + # Convert logits to multi-hot encoding. We compare the logits to 0 instead of 0.5, because the sigmoid is not applied. + # You can also pass `preprocess_logits_for_metrics=lambda logits, labels: nn.functional.sigmoid(logits)` to the Trainer + # and set p > 0.5 below (less efficient in this case) + predictions = np.array([np.where(p > 0, 1, 0) for p in predictions]) + else: + predictions = np.argmax(predictions, axis=1) + output_predict_file = os.path.join( + training_args.output_dir, "predict_results.txt" + ) + if trainer.is_world_process_zero(): + with open(output_predict_file, "w") as writer: + logger.info("***** Predict results *****") + writer.write("index\tprediction\n") + for index, item in enumerate(predictions): + if is_regression: + writer.write(f"{index}\t{item:3.3f}\n") + elif is_multi_label: + # recover from multi-hot encoding + item = [label_list[i] for i in range(len(item)) if item[i] == 1] + writer.write(f"{index}\t{item}\n") + else: + item = label_list[item] + writer.write(f"{index}\t{item}\n") + logger.info("Predict results saved at {}".format(output_predict_file)) + kwargs = { + "finetuned_from": model_args.model_name_or_path, + "tasks": "text-classification", + } + + if training_args.push_to_hub: + trainer.push_to_hub(**kwargs) + else: + trainer.create_model_card(**kwargs) + + +def _mp_fn(index): + # For xla_spawn (TPUs) + main() + + +if __name__ == "__main__": + main() diff --git a/examples/CCI3/run_classification_trainval.sh b/examples/CCI3/run_classification_trainval.sh new file mode 100644 index 00000000..4c2ae8f7 --- /dev/null +++ b/examples/CCI3/run_classification_trainval.sh @@ -0,0 +1,88 @@ +#!/bin/bash +# --------------------------------------------------------------- +# [Author] : shixiaofeng, ldwang +# [Descriptions] : + +set -x -e +RUN_ID=$(date +"%Y%m%d%H") + +PROJ_HOME=/share/project/ldwang/text-quality + +GOLDEN_MODEL=qwen2 + +FREEZE_BACKBONE=1 + +PRETRAINED_MODEL=bge-m3-xlmroberta-nodrop +MODEL_PATH=/share/project/ldwang/Aquila3/bge-m3-nodrop +MAX_SEQ_LEN=2048 + +DATA_VERSION=145k +EVAL_SAMPLES=15000 +train_file=/share/projset/ldwang/text-quality/datasets/$DATA_VERSION/train.jsonl +validation_file=/share/projset/ldwang/text-quality/datasets/$DATA_VERSION/val.jsonl + +# Final Settings +BSZ=256 +EPOCHNUM=20 +LR=1e-4 + +EXP=quality_scorer_${PRETRAINED_MODEL}-fb$FREEZE_BACKBONE +SUBEXP=from-${GOLDEN_MODEL}-ds$DATA_VERSION-ep$EPOCHNUM-bsz$BSZ-lr$LR-seq$MAX_SEQ_LEN +WANDB_RUN_ID=$EXP-$SUBEXP-$RUN_ID + +output_path=$PROJ_HOME/checkpoints/$EXP/$SUBEXP/$RUN_ID + +RUN_SCRIPT=run_classification.py +mkdir -p $output_path +cp $0 $output_path/ +cp $RUN_SCRIPT $output_path/ + +MASTER_PORT=20001 +export CUDA_VISIBLE_DEVICES="0,1,2,3" +GPU_NUM=4 +MICRO_BSZ=1 +GRAD_ACCUM=$(( $BSZ / $GPU_NUM / $MICRO_BSZ )) +export WANDB_MODE="offline" +export WANDB_PROJECT="data_quality_scorer" +export WANDB_DIR=$output_path +export WANDB_RUN_ID=$WANDB_RUN_ID + +torchrun \ + --nnodes=1 \ + --node_rank=0 \ + --nproc_per_node=$GPU_NUM \ + --master_addr=127.0.0.1 \ + --master_port=$MASTER_PORT \ + run_classification.py \ + --model_name_or_path $MODEL_PATH \ + --output_dir $output_path \ + --overwrite_output_dir \ + --train_file ${train_file} \ + --validation_file ${validation_file} \ + --shuffle_train_dataset \ + --metric_name mse \ + --text_column_name text \ + --label_column_name label \ + --do_regression true \ + --do_train \ + --do_eval \ + --freeze_backbone $FREEZE_BACKBONE \ + --model_name $PRETRAINED_MODEL \ + --max_eval_samples $EVAL_SAMPLES \ + --max_seq_length ${MAX_SEQ_LEN} \ + --per_device_train_batch_size $MICRO_BSZ \ + --per_device_eval_batch_size $MICRO_BSZ \ + --gradient_accumulation_steps $GRAD_ACCUM \ + --pad_to_max_length False \ + --save_strategy epoch \ + --save_only_model \ + --evaluation_strategy epoch \ + --seed 42 \ + --lr_scheduler_type cosine \ + --learning_rate $LR \ + --num_train_epochs $EPOCHNUM \ + --logging_steps 1 \ + --bf16 \ + --run_name $WANDB_RUN_ID \ + --report_to wandb 1>>$output_path/run.log 2>&1 & +