# coding=utf-8 # copied from hugginface github # Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. # team. # Copyright (c) 2018, NVIDIA CORPORATION. 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. """BERT for Referring Expression Comprehension""" import argparse import json import os from os.path import exists, join import random from time import time import torch from torch.nn.utils import clip_grad_norm_ from torch.optim import Adam, Adamax from torch.utils.data import DataLoader # to be deprecated once upgraded to 1.2 # from torch.utils.data.distributed import DistributedSampler from data import DistributedSampler from apex import amp from horovod import torch as hvd import numpy as np from tqdm import tqdm from data import (ReImageFeatDir, ReferringExpressionDataset, ReferringExpressionEvalDataset, re_collate, re_eval_collate, PrefetchLoader) from model import BertForReferringExpressionComprehension from optim import warmup_linear, noam_schedule, AdamW from utils.logger import LOGGER, TB_LOGGER, RunningMeter, add_log_to_file from utils.distributed import (all_reduce_and_rescale_tensors, all_gather_list, broadcast_tensors) from utils.save import ModelSaver, save_training_meta from utils.misc import NoOp, parse_with_config def main(opts): hvd.init() n_gpu = hvd.size() device = torch.device("cuda", hvd.local_rank()) torch.cuda.set_device(hvd.local_rank()) rank = hvd.rank() opts.rank = rank LOGGER.info(f"device: {device}, n_gpu: {n_gpu}, rank: {hvd.rank()}, " f"16-bits training: {opts.fp16}") if opts.gradient_accumulation_steps < 1: raise ValueError("Invalid gradient_accumulation_steps parameter: {}, " "should be >= 1".format( opts.gradient_accumulation_steps)) random.seed(opts.seed) np.random.seed(opts.seed) torch.manual_seed(opts.seed) if n_gpu > 0: torch.cuda.manual_seed_all(opts.seed) # train_samples = None LOGGER.info(f"Loading Train Dataset {opts.train_txt_db}, " f"{opts.train_img_dir}") # load DBs and image dirs train_img_dir = ReImageFeatDir(opts.train_img_dir) train_dataset = ReferringExpressionDataset( opts.train_txt_db, train_img_dir, max_txt_len=opts.max_txt_len) val_img_dir = ReImageFeatDir(opts.val_img_dir) val_dataset = ReferringExpressionEvalDataset( opts.val_txt_db, val_img_dir, max_txt_len=opts.max_txt_len) # Prepro model if opts.checkpoint and opts.checkpoint != 'scratch': if opts.checkpoint == 'google-bert': # from google-bert checkpoint = None else: checkpoint = torch.load(opts.checkpoint) else: # from scratch checkpoint = {} bert_model = json.load(open(f'{opts.train_txt_db}/meta.json'))['bert'] model = BertForReferringExpressionComprehension.from_pretrained( bert_model, img_dim=2048, loss=opts.train_loss, margin=opts.margin, hard_ratio=opts.hard_ratio, mlp=opts.mlp, state_dict=checkpoint ) if opts.cut_bert != -1: # cut some layers of BERT model.bert.encoder.layer = torch.nn.ModuleList( model.bert.encoder.layer[:opts.cut_bert] ) del checkpoint for name, module in model.named_modules(): # we may want to tune dropout for smaller dataset if isinstance(module, torch.nn.Dropout): if module.p != opts.dropout: module.p = opts.dropout LOGGER.info(f'{name} set to {opts.dropout}') model.to(device) # make sure every process has same model params in the beginning broadcast_tensors([p.data for p in model.parameters()], 0) # Prepare optimizer param_optimizer = list(model.named_parameters()) no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight'] optimizer_grouped_parameters = [ {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': opts.weight_decay}, {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0} ] # currently Adam only if opts.optim == 'adam': OptimCls = Adam elif opts.optim == 'adamax': OptimCls = Adamax elif opts.optim == 'adamw': OptimCls = AdamW else: raise ValueError('invalid optimizer') optimizer = OptimCls(optimizer_grouped_parameters, lr=opts.learning_rate, betas=opts.betas) model, optimizer = amp.initialize(model, optimizer, enabled=opts.fp16, opt_level='O2') global_step = 0 LOGGER.info("***** Running training *****") LOGGER.info(" Num examples = %d", len(train_dataset)) LOGGER.info(" Batch size = %d", opts.train_batch_size) LOGGER.info(" Accumulate steps = %d", opts.gradient_accumulation_steps) LOGGER.info(" Num steps = %d", opts.num_train_steps) train_sampler = DistributedSampler( train_dataset, num_replicas=n_gpu, rank=rank, shuffle=False) train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=opts.train_batch_size, num_workers=opts.n_workers, pin_memory=opts.pin_mem, collate_fn=re_collate) train_dataloader = PrefetchLoader(train_dataloader) val_sampler = DistributedSampler( val_dataset, num_replicas=n_gpu, rank=rank, shuffle=False) val_dataloader = DataLoader(val_dataset, sampler=val_sampler, batch_size=opts.val_batch_size, num_workers=opts.n_workers, pin_memory=opts.pin_mem, collate_fn=re_eval_collate) val_dataloader = PrefetchLoader(val_dataloader) if rank == 0: save_training_meta(opts) TB_LOGGER.create(join(opts.output_dir, 'log')) pbar = tqdm(total=opts.num_train_steps) model_saver = ModelSaver(join(opts.output_dir, 'ckpt'), 'model_epoch') os.makedirs(join(opts.output_dir, 'results')) # store ITM predictions add_log_to_file(join(opts.output_dir, 'log', 'log.txt')) else: LOGGER.disabled = True pbar = NoOp() model_saver = NoOp() running_loss = RunningMeter(opts.train_loss) n_examples = 0 n_epoch = 0 best_val_acc, best_epoch = None, None start = time() # quick hack for amp delay_unscale bug optimizer.zero_grad() optimizer.step() while True: model.train() for step, batch in enumerate(train_dataloader): if global_step >= opts.num_train_steps: break *_, targets = batch n_examples += targets.size(0) loss = model(*batch, compute_loss=True) loss = loss.sum() # sum over vectorized loss TODO: investigate delay_unscale = (step+1) % opts.gradient_accumulation_steps != 0 with amp.scale_loss(loss, optimizer, delay_unscale=delay_unscale ) as scaled_loss: scaled_loss.backward() if not delay_unscale: # gather gradients from every processes # do this before unscaling to make sure every process uses # the same gradient scale grads = [p.grad.data for p in model.parameters() if p.requires_grad and p.grad is not None] all_reduce_and_rescale_tensors(grads, float(1)) running_loss(loss.item()) if (step + 1) % opts.gradient_accumulation_steps == 0: global_step += 1 # learning rate scheduling if opts.decay == 'linear': lr_this_step = opts.learning_rate * warmup_linear( global_step, opts.warmup_steps, opts.num_train_steps) elif opts.decay == 'invsqrt': lr_this_step = opts.learning_rate * noam_schedule( global_step, opts.warmup_steps) elif opts.decay == 'constant': lr_this_step = opts.learning_rate if lr_this_step < 0: # save guard for possible miscalculation of train steps lr_this_step = 1e-8 for param_group in optimizer.param_groups: param_group['lr'] = lr_this_step TB_LOGGER.add_scalar('lr', lr_this_step, global_step) # log loss losses = all_gather_list(running_loss) running_loss = RunningMeter( opts.train_loss, sum(l.val for l in losses)/len(losses)) TB_LOGGER.add_scalar('loss_'+opts.train_loss, running_loss.val, global_step) TB_LOGGER.step() # update model params if opts.grad_norm != -1: grad_norm = clip_grad_norm_(amp.master_params(optimizer), opts.grad_norm) TB_LOGGER.add_scalar('grad_norm', grad_norm, global_step) optimizer.step() optimizer.zero_grad() pbar.update(1) if global_step % 5 == 0: torch.cuda.empty_cache() if global_step % 100 == 0: # monitor training throughput tot_ex = sum(all_gather_list(n_examples)) ex_per_sec = int(tot_ex / (time()-start)) LOGGER.info(f'{tot_ex} examples trained at ' f'{ex_per_sec} ex/s') TB_LOGGER.add_scalar('perf/ex_per_s', ex_per_sec, global_step) # evaluate after each epoch val_log, _ = validate(model, val_dataloader) TB_LOGGER.log_scaler_dict(val_log) # save model n_epoch += 1 model_saver.save(model, n_epoch) LOGGER.info(f"finished {n_epoch} epochs") # save best model if best_val_acc is None or val_log['valid/acc'] > best_val_acc: best_val_acc = val_log['valid/acc'] best_epoch = n_epoch model_saver.save(model, 'best') # shuffle training data for the next epoch train_dataloader.loader.dataset.shuffle() # is training finished? if global_step >= opts.num_train_steps: break val_log, results = validate(model, val_dataloader) with open(f'{opts.output_dir}/results/' f'results_{global_step}_' f'rank{rank}_final.json', 'w') as f: json.dump(results, f) TB_LOGGER.log_scaler_dict(val_log) model_saver.save(model, f'{global_step}_final') # print best model LOGGER.info(f'best_val_acc = {best_val_acc*100:.2f}% ' f'at epoch {best_epoch}.') @torch.no_grad() def validate(model, val_dataloader): LOGGER.info(f"start running evaluation.") model.eval() tot_score = 0 n_ex = 0 st = time() predictions = {} for i, batch in enumerate(val_dataloader): # inputs (*batch_inputs, tgt_box_list, obj_boxes_list, sent_ids) = batch # scores (n, max_num_bb) scores = model(*batch_inputs, targets=None, compute_loss=False) ixs = torch.argmax(scores, 1).cpu().detach().numpy() # (n, ) # pred_boxes for ix, obj_boxes, tgt_box, sent_id in \ zip(ixs, obj_boxes_list, tgt_box_list, sent_ids): pred_box = obj_boxes[ix] predictions['sent_id'] = {'pred_box': pred_box.tolist(), 'tgt_box': tgt_box.tolist()} if (val_dataloader.loader.dataset.computeIoU(pred_box, tgt_box) > .5): tot_score += 1 n_ex += 1 tot_time = time()-st tot_score = sum(all_gather_list(tot_score)) n_ex = sum(all_gather_list(n_ex)) val_acc = tot_score / n_ex val_log = {'valid/acc': val_acc, 'valid/ex_per_s': n_ex/tot_time} model.train() LOGGER.info(f"validation ({n_ex} sents) finished in " f"{int(tot_time)} seconds" f", accuracy: {val_acc*100:.2f}%") return val_log, predictions if __name__ == '__main__': parser = argparse.ArgumentParser() # Required parameters parser.add_argument("--train_txt_db", default=None, type=str, help="The input train corpus. (LMDB)") parser.add_argument("--train_img_dir", default=None, type=str, help="The input train images.") parser.add_argument("--val_txt_db", default=None, type=str, help="The input validation corpus. (LMDB)") parser.add_argument("--val_img_dir", default=None, type=str, help="The input validation images.") parser.add_argument('--img_format', default='npz', choices=['npz', 'lmdb', 'lmdb-compress'], help='format of image feature') parser.add_argument("--checkpoint", default=None, type=str, help="pretrained model (can take 'google-bert') ") parser.add_argument("--cut_bert", default=-1, type=int, help="reduce BERT layers (-1 for original depth)") parser.add_argument("--mlp", default=1, type=int, help="number of MLP layers for RE output") parser.add_argument( "--output_dir", default=None, type=str, help="The output directory where the model checkpoints will be " "written.") # Prepro parameters parser.add_argument('--max_txt_len', type=int, default=60, help='max number of tokens in text (BERT BPE)') # training parameters parser.add_argument("--train_batch_size", default=128, type=int, help="Total batch size for training. " "(batch by examples)") parser.add_argument("--val_batch_size", default=256, type=int, help="Total batch size for validation. " "(batch by tokens)") parser.add_argument("--train_loss", default="cls", type=str, choices=['cls', 'rank'], help="loss to used during training") parser.add_argument("--margin", default=0.2, type=float, help="margin of ranking loss") parser.add_argument("--hard_ratio", default=0.3, type=float, help="sampling ratio of hard negatives") parser.add_argument('--gradient_accumulation_steps', type=int, default=16, help="Number of updates steps to accumualte before " "performing a backward/update pass.") parser.add_argument("--learning_rate", default=3e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--num_train_steps", default=32000, type=int, help="Total number of training updates to perform.") parser.add_argument("--optim", default='adam', choices=['adam', 'adamax', 'adamw'], help="optimizer") parser.add_argument("--betas", default=[0.9, 0.98], nargs='+', type=float, help="beta for adam optimizer") parser.add_argument("--decay", default='linear', choices=['linear', 'invsqrt', 'constant'], help="learning rate decay method") parser.add_argument("--dropout", default=0.1, type=float, help="tune dropout regularization") parser.add_argument("--weight_decay", default=0.0, type=float, help="weight decay (L2) regularization") parser.add_argument("--grad_norm", default=0.25, type=float, help="gradient clipping (-1 for no clipping)") parser.add_argument("--warmup_steps", default=4000, type=int, help="Number of training steps to perform linear " "learning rate warmup for. (invsqrt decay)") # device parameters parser.add_argument('--seed', type=int, default=24, help="random seed for initialization") parser.add_argument('--fp16', action='store_true', help="Whether to use 16-bit float precision instead " "of 32-bit") parser.add_argument('--n_workers', type=int, default=4, help="number of data workers") parser.add_argument('--pin_mem', action='store_true', help="pin memory") # can use config files parser.add_argument('--config', help='JSON config files') args = parse_with_config(parser) if exists(args.output_dir) and os.listdir(args.output_dir): raise ValueError("Output directory ({}) already exists and is not " "empty.".format(args.output_dir)) # options safe guard main(args)