# Copyright (c) 2021 PaddlePaddle Authors. 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. import os import sys __dir__ = os.path.dirname(os.path.abspath(__file__)) sys.path.append(__dir__) sys.path.append(os.path.abspath(os.path.join(__dir__, '../..'))) import random import time import copy import logging import argparse import paddle import numpy as np from seqeval.metrics import classification_report, f1_score, precision_score, recall_score from paddlenlp.transformers import LayoutXLMModel, LayoutXLMTokenizer, LayoutXLMForTokenClassification from paddlenlp.transformers import LayoutLMModel, LayoutLMTokenizer, LayoutLMForTokenClassification from xfun import XFUNDataset from vqa_utils import parse_args, get_bio_label_maps, print_arguments, set_seed from eval_ser import evaluate from losses import SERLoss from ppocr.utils.logging import get_logger MODELS = { 'LayoutXLM': (LayoutXLMTokenizer, LayoutXLMModel, LayoutXLMForTokenClassification), 'LayoutLM': (LayoutLMTokenizer, LayoutLMModel, LayoutLMForTokenClassification) } def train(args): os.makedirs(args.output_dir, exist_ok=True) rank = paddle.distributed.get_rank() distributed = paddle.distributed.get_world_size() > 1 logger = get_logger(log_file=os.path.join(args.output_dir, "train.log")) print_arguments(args, logger) label2id_map, id2label_map = get_bio_label_maps(args.label_map_path) loss_class = SERLoss(len(label2id_map)) pad_token_label_id = loss_class.ignore_index # dist mode if distributed: paddle.distributed.init_parallel_env() tokenizer_class, base_model_class, model_class = MODELS[args.ser_model_type] tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path) if not args.resume: base_model = base_model_class.from_pretrained(args.model_name_or_path) model = model_class( base_model, num_classes=len(label2id_map), dropout=None) logger.info('train from scratch') else: logger.info('resume from {}'.format(args.model_name_or_path)) model = model_class.from_pretrained(args.model_name_or_path) # dist mode if distributed: model = paddle.DataParallel(model) train_dataset = XFUNDataset( tokenizer, data_dir=args.train_data_dir, label_path=args.train_label_path, label2id_map=label2id_map, img_size=(224, 224), pad_token_label_id=pad_token_label_id, contains_re=False, add_special_ids=False, return_attention_mask=True, load_mode='all') eval_dataset = XFUNDataset( tokenizer, data_dir=args.eval_data_dir, label_path=args.eval_label_path, label2id_map=label2id_map, img_size=(224, 224), pad_token_label_id=pad_token_label_id, contains_re=False, add_special_ids=False, return_attention_mask=True, load_mode='all') train_sampler = paddle.io.DistributedBatchSampler( train_dataset, batch_size=args.per_gpu_train_batch_size, shuffle=True) train_dataloader = paddle.io.DataLoader( train_dataset, batch_sampler=train_sampler, num_workers=args.num_workers, use_shared_memory=True, collate_fn=None, ) eval_dataloader = paddle.io.DataLoader( eval_dataset, batch_size=args.per_gpu_eval_batch_size, num_workers=args.num_workers, use_shared_memory=True, collate_fn=None, ) t_total = len(train_dataloader) * args.num_train_epochs # build linear decay with warmup lr sch lr_scheduler = paddle.optimizer.lr.PolynomialDecay( learning_rate=args.learning_rate, decay_steps=t_total, end_lr=0.0, power=1.0) if args.warmup_steps > 0: lr_scheduler = paddle.optimizer.lr.LinearWarmup( lr_scheduler, args.warmup_steps, start_lr=0, end_lr=args.learning_rate, ) optimizer = paddle.optimizer.AdamW( learning_rate=lr_scheduler, parameters=model.parameters(), epsilon=args.adam_epsilon, weight_decay=args.weight_decay) # Train! logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_dataset)) logger.info(" Num Epochs = %d", args.num_train_epochs) logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size) logger.info( " Total train batch size (w. parallel, distributed) = %d", args.per_gpu_train_batch_size * paddle.distributed.get_world_size(), ) logger.info(" Total optimization steps = %d", t_total) global_step = 0 tr_loss = 0.0 set_seed(args.seed) best_metrics = None train_reader_cost = 0.0 train_run_cost = 0.0 total_samples = 0 reader_start = time.time() print_step = 1 model.train() for epoch_id in range(args.num_train_epochs): for step, batch in enumerate(train_dataloader): train_reader_cost += time.time() - reader_start if args.ser_model_type == 'LayoutLM': if 'image' in batch: batch.pop('image') labels = batch.pop('labels') train_start = time.time() outputs = model(**batch) train_run_cost += time.time() - train_start if args.ser_model_type == 'LayoutXLM': outputs = outputs[0] loss = loss_class(labels, outputs, batch['attention_mask']) # model outputs are always tuple in ppnlp (see doc) loss = loss.mean() loss.backward() tr_loss += loss.item() optimizer.step() lr_scheduler.step() # Update learning rate schedule optimizer.clear_grad() global_step += 1 total_samples += batch['input_ids'].shape[0] if rank == 0 and step % print_step == 0: logger.info( "epoch: [{}/{}], iter: [{}/{}], global_step:{}, train loss: {:.6f}, lr: {:.6f}, avg_reader_cost: {:.5f} sec, avg_batch_cost: {:.5f} sec, avg_samples: {:.5f}, ips: {:.5f} images/sec". format(epoch_id, args.num_train_epochs, step, len(train_dataloader), global_step, loss.numpy()[0], lr_scheduler.get_lr(), train_reader_cost / print_step, (train_reader_cost + train_run_cost) / print_step, total_samples / print_step, total_samples / (train_reader_cost + train_run_cost))) train_reader_cost = 0.0 train_run_cost = 0.0 total_samples = 0 if rank == 0 and args.eval_steps > 0 and global_step % args.eval_steps == 0 and args.evaluate_during_training: # Log metrics # Only evaluate when single GPU otherwise metrics may not average well results, _ = evaluate(args, model, tokenizer, loss_class, eval_dataloader, label2id_map, id2label_map, pad_token_label_id, logger) if best_metrics is None or results["f1"] >= best_metrics["f1"]: best_metrics = copy.deepcopy(results) output_dir = os.path.join(args.output_dir, "best_model") os.makedirs(output_dir, exist_ok=True) if distributed: model._layers.save_pretrained(output_dir) else: model.save_pretrained(output_dir) tokenizer.save_pretrained(output_dir) paddle.save(args, os.path.join(output_dir, "training_args.bin")) logger.info("Saving model checkpoint to {}".format( output_dir)) logger.info("[epoch {}/{}][iter: {}/{}] results: {}".format( epoch_id, args.num_train_epochs, step, len(train_dataloader), results)) if best_metrics is not None: logger.info("best metrics: {}".format(best_metrics)) reader_start = time.time() if rank == 0: # Save model checkpoint output_dir = os.path.join(args.output_dir, "latest_model") os.makedirs(output_dir, exist_ok=True) if distributed: model._layers.save_pretrained(output_dir) else: model.save_pretrained(output_dir) tokenizer.save_pretrained(output_dir) paddle.save(args, os.path.join(output_dir, "training_args.bin")) logger.info("Saving model checkpoint to {}".format(output_dir)) return global_step, tr_loss / global_step if __name__ == "__main__": args = parse_args() train(args)