run_vqvae.py

import argparse
import datetime
import numpy as np
import time
import torch
import torch.backends.cudnn as cudnn
import json
import os

from pathlib import Path
from collections import OrderedDict

from timm.models import create_model
from timm.utils import ModelEma
from optim_factory import create_optimizer, get_parameter_groups, LayerDecayValueAssigner

from datasets import build_shape_surface_occupancy_dataset
from engine_for_vqvae import train_one_epoch, evaluate
import utils
from utils import NativeScalerWithGradNormCount as NativeScaler

import modeling_vqvae

def get_args():
    parser = argparse.ArgumentParser('VQVAE', add_help=False)
    parser.add_argument('--batch_size', default=128, type=int)
    parser.add_argument('--epochs', default=800, type=int)
    parser.add_argument('--update_freq', default=1, type=int)
    parser.add_argument('--save_ckpt_freq', default=20, type=int)

    # Model parameters
    parser.add_argument('--model', default='vqvae_128', type=str,
                        metavar='MODEL', help='Name of model to train')
    parser.add_argument('--point_cloud_size', default=2048, type=int,
                        help='images point cloud size')

    parser.add_argument('--drop', type=float, default=0.0, metavar='PCT',
                        help='Dropout rate (default: 0.)')
    parser.add_argument('--attn_drop_rate', type=float, default=0.0, metavar='PCT',
                        help='Attention dropout rate (default: 0.)')
    parser.add_argument('--drop_path', type=float, default=0.1, metavar='PCT',
                        help='Drop path rate (default: 0.1)')

    parser.add_argument('--disable_eval', action='store_true', default=False)

    parser.add_argument('--model_ema', action='store_true', default=False)

    # Optimizer parameters
    parser.add_argument('--opt', default='adamw', type=str, metavar='OPTIMIZER',
                        help='Optimizer (default: "adamw"')
    parser.add_argument('--opt_eps', default=1e-8, type=float, metavar='EPSILON', # 1e-6
                        help='Optimizer Epsilon (default: 1e-8)')
    parser.add_argument('--opt_betas', default=None, type=float, nargs='+', metavar='BETA',
                        help='Optimizer Betas (default: None, use opt default)')
    parser.add_argument('--clip_grad', type=float, default=None, metavar='NORM', #
                        help='Clip gradient norm (default: None, no clipping)')
    parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
                        help='SGD momentum (default: 0.9)')
    parser.add_argument('--weight_decay', type=float, default=0.05,
                        help='weight decay (default: 0.05)')
    parser.add_argument('--weight_decay_end', type=float, default=None, help="""Final value of the
        weight decay. We use a cosine schedule for WD and using a larger decay by
        the end of training improves performance for ViTs.""")

    parser.add_argument('--lr', type=float, default=5e-4, metavar='LR',
                        help='learning rate (default: 1e-3)')
    parser.add_argument('--layer_decay', type=float, default=1.0)

    parser.add_argument('--warmup_lr', type=float, default=1e-6, metavar='LR',
                        help='warmup learning rate (default: 1e-6)')
    parser.add_argument('--min_lr', type=float, default=1e-6, metavar='LR',
                        help='lower lr bound for cyclic schedulers that hit 0 (1e-5)')

    parser.add_argument('--warmup_epochs', type=int, default=40, metavar='N',
                        help='epochs to warmup LR, if scheduler supports')
    parser.add_argument('--warmup_steps', type=int, default=-1, metavar='N',
                        help='num of steps to warmup LR, will overload warmup_epochs if set > 0')

    # Dataset parameters
    parser.add_argument('--data_path', default='/your/data/path', type=str,
                        help='dataset path')
    parser.add_argument('--output_dir', default='',
                        help='path where to save, empty for no saving')
    parser.add_argument('--log_dir', default=None,
                        help='path where to tensorboard log')
    parser.add_argument('--device', default='cuda',
                        help='device to use for training / testing')
    parser.add_argument('--seed', default=0, type=int)
    parser.add_argument('--resume', default='',
                        help='resume from checkpoint')
    parser.add_argument('--auto_resume', action='store_true')
    parser.add_argument('--no_auto_resume', action='store_false', dest='auto_resume')
    parser.set_defaults(auto_resume=True)
    
    parser.add_argument('--save_ckpt', action='store_true')
    parser.add_argument('--no_save_ckpt', action='store_false', dest='save_ckpt')
    parser.set_defaults(save_ckpt=True)

    parser.add_argument('--start_epoch', default=0, type=int, metavar='N',
                        help='start epoch')
    parser.add_argument('--eval', action='store_true',
                        help='Perform evaluation only')
    parser.add_argument('--dist_eval', action='store_true', default=False,
                        help='Enabling distributed evaluation')
    parser.add_argument('--num_workers', default=16, type=int)
    parser.add_argument('--pin_mem', action='store_true',
                        help='Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU.')
    parser.add_argument('--no_pin_mem', action='store_false', dest='pin_mem')
    parser.set_defaults(pin_mem=True)

    # distributed training parameters
    parser.add_argument('--world_size', default=1, type=int,
                        help='number of distributed processes')
    parser.add_argument('--local_rank', default=-1, type=int)
    parser.add_argument('--dist_on_itp', action='store_true')
    parser.add_argument('--dist_url', default='env://',
                        help='url used to set up distributed training')
    
    ds_init = None
    return parser.parse_args(), ds_init

def main(args, ds_init):
    utils.init_distributed_mode(args)

    print(args)

    device = torch.device(args.device)

    # fix the seed for reproducibility
    seed = args.seed + utils.get_rank()
    torch.manual_seed(seed)
    np.random.seed(seed)

    cudnn.benchmark = True

    dataset_train  = build_shape_surface_occupancy_dataset('train', args=args)
    if args.disable_eval:
        dataset_val = None
    else:
        dataset_val = build_shape_surface_occupancy_dataset('val', args=args)

    if True:  # args.distributed:
        num_tasks = utils.get_world_size()
        global_rank = utils.get_rank()
        sampler_train = torch.utils.data.DistributedSampler(
            dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True
        )
        print("Sampler_train = %s" % str(sampler_train))
        if args.dist_eval:
            if len(dataset_val) % num_tasks != 0:
                print('Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. '
                      'This will slightly alter validation results as extra duplicate entries are added to achieve '
                      'equal num of samples per-process.')
            sampler_val = torch.utils.data.DistributedSampler(
                dataset_val, num_replicas=num_tasks, rank=global_rank, shuffle=False)
        else:
            sampler_val = torch.utils.data.SequentialSampler(dataset_val)
    else:
        sampler_train = torch.utils.data.RandomSampler(dataset_train)
        sampler_val = torch.utils.data.SequentialSampler(dataset_val)

    if global_rank == 0 and args.log_dir is not None:
        os.makedirs(args.log_dir, exist_ok=True)
        log_writer = utils.TensorboardLogger(log_dir=args.log_dir)
    else:
        log_writer = None

    data_loader_train = torch.utils.data.DataLoader(
        dataset_train, sampler=sampler_train,
        batch_size=args.batch_size,
        num_workers=args.num_workers,
        pin_memory=args.pin_mem,
        drop_last=True,
        prefetch_factor=1,
    )

    if dataset_val is not None:
        data_loader_val = torch.utils.data.DataLoader(
            dataset_val, sampler=sampler_val,
            batch_size=1,
            num_workers=args.num_workers,
            pin_memory=args.pin_mem,
            drop_last=False,
        )
    else:
        data_loader_val = None

    model = create_model(
        args.model,
    )

    model.to(device)

    model_ema = None
    if args.model_ema:
        # Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
        model_ema = ModelEma(
            model,
            decay=args.model_ema_decay,
            device='cpu' if args.model_ema_force_cpu else '',
            resume='')
        print("Using EMA with decay = %.8f" % args.model_ema_decay)

    model_without_ddp = model
    n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad)

    print("Model = %s" % str(model_without_ddp))
    print('number of params:', n_parameters)

    total_batch_size = args.batch_size * args.update_freq * utils.get_world_size()
    num_training_steps_per_epoch = len(dataset_train) // total_batch_size
    args.lr = args.lr * total_batch_size / 256
    print("LR = %.8f" % args.lr)
    print("Batch size = %d" % total_batch_size)
    print("Update frequent = %d" % args.update_freq)
    print("Number of training examples = %d" % len(dataset_train))
    print("Number of training training per epoch = %d" % num_training_steps_per_epoch)

    # num_layers = model_without_ddp.get_num_layers()
    num_layers = 12
    if args.layer_decay < 1.0:
        assigner = LayerDecayValueAssigner(list(args.layer_decay ** (num_layers + 1 - i) for i in range(num_layers + 2)))
    else:
        assigner = None

    if assigner is not None:
        print("Assigned values = %s" % str(assigner.values))

    skip_weight_decay_list = model.no_weight_decay()
    print("Skip weight decay list: ", skip_weight_decay_list)

    if args.distributed:
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu], find_unused_parameters=False)
        model_without_ddp = model.module

    optimizer = create_optimizer(
        args, model_without_ddp, skip_list=skip_weight_decay_list,
        get_num_layer=assigner.get_layer_id if assigner is not None else None, 
        get_layer_scale=assigner.get_scale if assigner is not None else None,
        )
    loss_scaler = NativeScaler()

    print("Use step level LR scheduler!")
    lr_schedule_values = utils.cosine_scheduler(
        args.lr, args.min_lr, args.epochs, num_training_steps_per_epoch,
        warmup_epochs=args.warmup_epochs, warmup_steps=args.warmup_steps,
    )
    if args.weight_decay_end is None:
        args.weight_decay_end = args.weight_decay
    wd_schedule_values = utils.cosine_scheduler(
        args.weight_decay, args.weight_decay_end, args.epochs, num_training_steps_per_epoch)
    print("Max WD = %.7f, Min WD = %.7f" % (max(wd_schedule_values), min(wd_schedule_values)))

    criterion = torch.nn.BCEWithLogitsLoss()
    print("criterion = %s" % str(criterion))

    utils.auto_load_model(
        args=args, model=model, model_without_ddp=model_without_ddp,
        optimizer=optimizer, loss_scaler=loss_scaler, model_ema=model_ema)

    print(f"Start training for {args.epochs} epochs")
    start_time = time.time()
    max_accuracy = 0.0
    for epoch in range(args.start_epoch, args.epochs):
        if args.distributed:
            data_loader_train.sampler.set_epoch(epoch)

        train_stats = train_one_epoch(
            model, criterion, data_loader_train, optimizer,
            device, epoch, loss_scaler, args.clip_grad, model_ema,
            log_writer=log_writer, start_steps=epoch * num_training_steps_per_epoch,
            lr_schedule_values=lr_schedule_values, wd_schedule_values=wd_schedule_values,
            num_training_steps_per_epoch=num_training_steps_per_epoch, update_freq=args.update_freq,
        )
        # print(train_stats)

        if args.output_dir and args.save_ckpt:
            if (epoch + 1) % args.save_ckpt_freq == 0 or epoch + 1 == args.epochs:
                utils.save_model(
                    args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer,
                    loss_scaler=loss_scaler, epoch=epoch, model_ema=model_ema)
        if data_loader_val is not None and (epoch % 10 == 0 or epoch + 1 == args.epochs):
            test_stats = evaluate(data_loader_val, model, device)
            print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['iou']:.1f}%")
            if max_accuracy < test_stats["iou"]:
                max_accuracy = test_stats["iou"]
                if args.output_dir and args.save_ckpt:
                    utils.save_model(
                        args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer,
                        loss_scaler=loss_scaler, epoch="best", model_ema=model_ema)

            print(f'Max accuracy: {max_accuracy:.2f}%')
            if log_writer is not None:
                log_writer.update(test_iou=test_stats['iou'], head="perf", step=epoch)
                log_writer.update(test_loss=test_stats['loss'], head="perf", step=epoch)

            log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
                        **{f'test_{k}': v for k, v in test_stats.items()},
                        'epoch': epoch,
                        'n_parameters': n_parameters}
        else:
            log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
                         # **{f'test_{k}': v for k, v in test_stats.items()},
                         'epoch': epoch,
                         'n_parameters': n_parameters}

        if args.output_dir and utils.is_main_process():
            if log_writer is not None:
                log_writer.flush()
            with open(os.path.join(args.output_dir, "log.txt"), mode="a", encoding="utf-8") as f:
                f.write(json.dumps(log_stats) + "\n")

    total_time = time.time() - start_time
    total_time_str = str(datetime.timedelta(seconds=int(total_time)))
    print('Training time {}'.format(total_time_str))

if __name__ == '__main__':
    opts, ds_init = get_args()
    if opts.output_dir:
        Path(opts.output_dir).mkdir(parents=True, exist_ok=True)
    main(opts, ds_init)