train.py

import os
import argparse
import time
import logging
import datetime
from copy import deepcopy

import cv2
import numpy as np
from tqdm import tqdm
import random
import sys
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.init as init
import torch.distributed as dist

import torch.utils.data
from torch.backends import cudnn
from torch.utils.tensorboard import SummaryWriter
import torchvision.utils as vutils
from fastai.vision import *

from Dino.utils.utils import Config, Logger, MyConcatDataset
from Dino.utils.util import Averager
from Dino.dataset.datasetsupervised_kmeans import ImageDatasetSelfSupervisedKmeans
from Dino.dataset.dataset import collate_fn_filter_none

from Dino.modules.vision_transformer import DINOHead
from Dino.modules import vision_transformer as vits
from Dino.modules.segmentor import SegHead
from Dino.model.dino_vision import ABIDINOModel
from Dino.modules import utils
from Dino.loss.Dino_loss import DINOLoss
from torchvision import models as torchvision_models
import warnings

warnings.filterwarnings("ignore")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
torchvision_archs = sorted(name for name in torchvision_models.__dict__
                           if name.islower() and not name.startswith("__")
                           and callable(torchvision_models.__dict__[name]))

def train(config):
    """parameter configuration"""
    utils.init_distributed_mode(config)
    utils.fix_random_seeds(config.seed)
    cudnn.benchmark = True

    """dataset preparation"""
    logging.info('Construct dataset.')
    train_dataloader = _get_databaunch(config)
    config.iter_num = len(train_dataloader)
    logging.info(f"each epoch iteration: {config.iter_num}")

    """ model configuration """
    # ============ building student and teacher networks ... ============
    # we changed the name DeiT-S for ViT-S to avoid confusions
    config.arch = config.arch.replace("deit", "vit")
    # if the network is a Vision Transformer (i.e. vit_tiny, vit_small, vit_base)
    if config.arch in vits.__dict__.keys():
        student = vits.__dict__[config.arch](
            patch_size=config.patch_size,
            drop_path_rate=config.drop_path_rate,  # stochastic depth
        )
        teacher = vits.__dict__[config.arch](patch_size=config.patch_size)
        embed_dim = student.embed_dim
    # if the network is a XCiT
    elif config.arch in torch.hub.list("facebookresearch/xcit:main"):
        student = torch.hub.load('facebookresearch/xcit:main', config.arch,
                                 pretrained=False, drop_path_rate=config.drop_path_rate)
        teacher = torch.hub.load('facebookresearch/xcit:main', config.arch, pretrained=False)
        embed_dim = student.embed_dim
    # otherwise, we check if the architecture is in torchvision models
    elif config.arch in torchvision_models.__dict__.keys():
        student = torchvision_models.__dict__[config.arch]()
        teacher = torchvision_models.__dict__[config.arch]()
        embed_dim = student.fc.weight.shape[1]
    else:
        print(f"Unknow architecture: {config.arch}")

    # multi-crop wrapper handles forward with inputs of different resolutions
    student = ABIDINOModel(
        student,
        SegHead(in_channels=config.model_seg_channel, mla_channels=128, mlahead_channels=64, num_classes=2),
        DINOHead(embed_dim, config.out_dim, use_bn=config.use_bn_in_head, norm_last_layer=config.norm_last_layer, ))
    teacher = ABIDINOModel(
        teacher,
        None,
        DINOHead(embed_dim, config.out_dim, config.use_bn_in_head), )

    # move networks to gpu
    student, teacher = student.cuda(), teacher.cuda()
    # synchronize batch norms (if any)
    if utils.has_batchnorms(student):
        student = nn.SyncBatchNorm.convert_sync_batchnorm(student)
        teacher = nn.SyncBatchNorm.convert_sync_batchnorm(teacher)

        # we need DDP wrapper to have synchro batch norms working...
        teacher = nn.parallel.DistributedDataParallel(teacher, device_ids=[config.gpu])
        teacher_without_ddp = teacher.module
    else:
        # teacher_without_ddp and teacher are the same thing
        teacher_without_ddp = teacher
    student = nn.parallel.DistributedDataParallel(student, device_ids=[config.gpu], find_unused_parameters=True)
    # teacher and student start with the same weights
    # teacher_without_ddp.load_state_dict(student.module.state_dict())
    teacher_without_ddp.backbone.load_state_dict(student.module.backbone.state_dict())
    teacher_without_ddp.head.load_state_dict(student.module.head.state_dict())

    # there is no backpropagation through the teacher, so no need for gradients
    for p in teacher.parameters():
        p.requires_grad = False
    print(f"Student and Teacher are built: they are both {config.arch} network.")

    """ setup loss """
    config.epochs = int(config.training_epochs * len(train_dataloader) * (
                config.batch_size_per_gpu * utils.get_world_size()) / config.imgnet_based) + 1
    print(f'training epochs is {config.epochs}')
    # ============ preparing loss ... ============
    dino_loss = DINOLoss(
        config.out_dim,
        config.crops_number,
        config.warmup_teacher_temp,
        config.teacher_temp,
        config.warmup_teacher_temp_epochs,
        config.epochs,
    ).cuda()
    # ============ preparing optimizer ... ============
    params_groups = utils.get_params_groups(student)
    if config.optimizer == "adamw":
        optimizer = torch.optim.AdamW(params_groups)  # to use with ViTs
    elif config.optimizer == "sgd":
        optimizer = torch.optim.SGD(params_groups, lr=0, momentum=0.9)  # lr is set by scheduler
    elif config.optimizer == "lars":
        optimizer = utils.LARS(params_groups)  # to use with convnet and large batches
    # for mixed precision training
    fp16_scaler = None
    if config.use_fp16:
        fp16_scaler = torch.cuda.amp.GradScaler()

    # ============ init schedulers ... ============
    lr_schedule = utils.cosine_iter_scheduler(
        config.lr * (config.batch_size_per_gpu * utils.get_world_size()) / 256.,  # linear scaling rule
        config.min_lr,
        config.training_epochs * len(train_dataloader),
        warmup_iters=int(
            (config.warmup_epoch * config.imgnet_based) / (config.batch_size_per_gpu * utils.get_world_size())),
    )
    wd_schedule = utils.cosine_iter_scheduler(
        config.weight_decay,
        config.weight_decay_end,
        config.training_epochs * len(train_dataloader),
    )
    # momentum parameter is increased to 1. during training with a cosine schedule
    momentum_schedule = utils.cosine_iter_scheduler(config.momentum_teacher, 1,
                                                    config.training_epochs * len(train_dataloader))
    print(f"Loss, optimizer and schedulers ready.")

    # ============ optionally resume training ... ============
    to_restore = {"epoch": 0, 'iteration': 0}
    utils.restart_from_checkpoint(
        os.path.join(config.output_dir, config.global_name, "checkpoint.pth"),
        run_variables=to_restore,
        student=student,
        teacher=teacher,
        optimizer=optimizer,
        fp16_scaler=fp16_scaler,
        dino_loss=dino_loss,
    )
    iteration = int(to_restore["iteration"])
    epoch = to_restore["epoch"]
    print(f'continue to train:{iteration}:{epoch}')
    start_time = time.time()
    min_loss = 100.
    best_eval_accuracy = 0.
    best_eval_char_accuracy = 0.
    global global_epoch
    global_epoch = 0

    print("Starting DINO training !")
    for train_epoch in range(config.training_epochs):
        train_dataloader.sampler.set_epoch(train_epoch)
        metric_logger = utils.MetricLogger(delimiter="  ")
        header = 'Epoch: [{}/{}]'.format(train_epoch, config.training_epochs)
        for (image_tensors, masks, metrics) in metric_logger.log_every(train_dataloader, 10, header):
            epoch = int((iteration + 1) * (config.batch_size_per_gpu * utils.get_world_size()) / config.imgnet_based)
            ### examine epoch updating state
            if epoch != global_epoch:
                global_epoch = deepcopy(epoch)
                # gather the stats from all processes
                metric_logger.synchronize_between_processes()
                print("Averaged stats:", metric_logger)
                train_stats = {k: meter.global_avg for k, meter in metric_logger.meters.items()}

                save_dict = {
                    'student': student.state_dict(),
                    'teacher': teacher.state_dict(),
                    'optimizer': optimizer.state_dict(),
                    'epoch': epoch,
                    'iteration': iteration,
                    # 'args': config,
                    'dino_loss': dino_loss.state_dict(),
                }
                if fp16_scaler is not None:
                    save_dict['fp16_scaler'] = fp16_scaler.state_dict()
                utils.save_on_master(save_dict, os.path.join(config.output_dir, config.global_name, 'checkpoint.pth'))
                if config.saveckp_freq and epoch % config.saveckp_freq == 0:
                    utils.save_on_master(save_dict, os.path.join(config.output_dir, config.global_name,
                                                                 f'checkpoint{epoch:04}.pth'))
                log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
                             'epoch': epoch}
                if utils.is_main_process():
                    with (Path(config.output_dir) / f"{config.global_name}/log.txt").open("a") as f:
                        f.write(json.dumps(log_stats) + "\n")

                metric_logger = utils.MetricLogger(delimiter="  ")
                header = 'Epoch: [{}/{}]'.format(train_epoch, config.training_epochs)

            image_tensors = image_tensors.cuda(non_blocking=True)
            masks = masks.cuda(non_blocking=True)

            for i, param_group in enumerate(optimizer.param_groups):
                param_group["lr"] = lr_schedule[iteration]
                if i == 0:  # only the first group is regularized
                    param_group["weight_decay"] = wd_schedule[iteration]

            # teacher and student forward passes + compute dino loss
            with torch.cuda.amp.autocast(fp16_scaler is not None):
                metrics = metrics.float()
                student_output = student(image_tensors, metrics, masks, epoch, clusters=None)
                teacher_output = teacher(image_tensors, metrics, None, None, clusters=student_output['zero'], index=student_output['index'])  # only the 2 global views pass through the teacher
                affine_grid = F.affine_grid(metrics[:, :2, :], size=(masks.shape[0], 1, masks.shape[1], masks.shape[2]))
                masks_image = F.grid_sample(masks.unsqueeze(1), affine_grid.to(masks.device))
                masks_image = (masks_image > 0.1).float().squeeze()
                student_output['gt'] = [masks, masks_image]
                loss = dino_loss(student_output, teacher_output, epoch)
            if not math.isfinite(loss.item()):
                print("Loss is {}, stopping training".format(loss.item()), force=True)
                sys.exit(1)

            # student update
            optimizer.zero_grad()
            param_norms = None
            if fp16_scaler is None:
                loss.backward()
                if config.clip_grad:
                    param_norms = utils.clip_gradients(student, config.clip_grad)
                utils.cancel_gradients_last_layer(epoch, student,
                                                  config.freeze_last_layer)
                optimizer.step()
            else:
                fp16_scaler.scale(loss).backward()
                if config.clip_grad:
                    fp16_scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place
                    param_norms = utils.clip_gradients(student, config.clip_grad)
                utils.cancel_gradients_last_layer(epoch, student,
                                                  config.freeze_last_layer)
                fp16_scaler.step(optimizer)
                fp16_scaler.update()

            # EMA update for the teacher
            with torch.no_grad():
                m = momentum_schedule[iteration]  # momentum parameter
                # for param_q, param_k in zip(student.module.parameters(), teacher_without_ddp.parameters()):
                #     param_k.data.mul_(m).add_((1 - m) * param_q.detach().data)
                for param_q, param_k in zip(student.module.backbone.parameters(),
                                            teacher_without_ddp.backbone.parameters()):
                    param_k.data.mul_(m).add_((1 - m) * param_q.detach().data)
                for param_q, param_k in zip(student.module.head.parameters(), teacher_without_ddp.head.parameters()):
                    param_k.data.mul_(m).add_((1 - m) * param_q.detach().data)

            # logging
            torch.cuda.synchronize()
            metric_logger.update(loss=loss.item())
            metric_logger.update(lr=optimizer.param_groups[0]["lr"])
            metric_logger.update(wd=optimizer.param_groups[0]["weight_decay"])

            if iteration % config.training_show_iters == 0:
                i = random.randint(0, config.batch_size_per_gpu - 1)
                last_losses = dino_loss.last_losses
                for name, loss in last_losses.items():
                    scalar_value = loss.data.cpu().numpy()
                    tag = 'metric/' + name
                    config.writer.add_scalar(tag=tag, scalar_value=scalar_value, global_step=iteration)
                lr = optimizer.param_groups[0]["lr"]
                config.writer.add_scalar(tag='metric/' + 'lr', scalar_value=lr, global_step=iteration)
                wd = optimizer.param_groups[0]["weight_decay"]
                config.writer.add_scalar(tag='metric/' + 'wd', scalar_value=wd, global_step=iteration)

            if iteration > config.training_epochs * len(train_dataloader):
                break

            iteration += 1

        if iteration > config.training_epochs * len(train_dataloader):
            break
    total_time = time.time() - start_time
    total_time_str = str(datetime.timedelta(seconds=int(total_time)))
    print('Training time {}'.format(total_time_str))


def _parse_arguments():
    parser = argparse.ArgumentParser()
    parser.add_argument('-c', '--config', type=str, required=True,
                        help='path to config file')
    # Model parameters
    parser.add_argument('--arch', default='vit_small', type=str,
                        # choices=['vit_tiny', 'vit_small', 'vit_base', 'xcit', 'deit_tiny', 'deit_small'] \
                        #         + torchvision_archs + torch.hub.list("facebookresearch/xcit:main"),
                        help="""Name of architecture to train. For quick experiments with ViTs,
            we recommend using vit_tiny or vit_small.""")
    parser.add_argument('--patch_size', default=4, type=int, help="""Size in pixels
            of input square patches - default 16 (for 16x16 patches). Using smaller
            values leads to better performance but requires more memory. Applies only
            for ViTs (vit_tiny, vit_small and vit_base). If <16, we recommend disabling
            mixed precision training (--use_fp16 false) to avoid unstabilities.""")
    parser.add_argument('--out_dim', default=65536, type=int, help="""Dimensionality of
            the DINO head output. For complex and large datasets large values (like 65k) work well.""")
    parser.add_argument('--norm_last_layer', default=True, type=utils.bool_flag,
                        help="""Whether or not to weight normalize the last layer of the DINO head.
            Not normalizing leads to better performance but can make the training unstable.
            In our experiments, we typically set this paramater to False with vit_small and True with vit_base.""")
    parser.add_argument('--momentum_teacher', default=0.996, type=float, help="""Base EMA
            parameter for teacher update. The value is increased to 1 during training with cosine schedule.
            We recommend setting a higher value with small batches: for example use 0.9995 with batch size of 256.""")
    parser.add_argument('--use_bn_in_head', default=False, type=utils.bool_flag,
                        help="Whether to use batch normalizations in projection head (Default: False)")

    # Temperature teacher parameters
    parser.add_argument('--warmup_teacher_temp', default=0.04, type=float,
                        help="""Initial value for the teacher temperature: 0.04 works well in most cases.
            Try decreasing it if the training loss does not decrease.""")
    parser.add_argument('--teacher_temp', default=0.04, type=float, help="""Final value (after linear warmup)
            of the teacher temperature. For most experiments, anything above 0.07 is unstable. We recommend
            starting with the default value of 0.04 and increase this slightly if needed.""")
    parser.add_argument('--warmup_teacher_temp_epochs', default=0, type=int,
                        help='Number of warmup epochs for the teacher temperature (Default: 30).')

    # Training/Optimization parameters
    parser.add_argument('--use_fp16', type=utils.bool_flag, default=True, help="""Whether or not
            to use half precision for training. Improves training time and memory requirements,
            but can provoke instability and slight decay of performance. We recommend disabling
            mixed precision if the loss is unstable, if reducing the patch size or if training with bigger ViTs.""")
    parser.add_argument('--weight_decay', type=float, default=0.04, help="""Initial value of the
            weight decay. With ViT, a smaller value at the beginning of training works well.""")
    parser.add_argument('--weight_decay_end', type=float, default=0.4, 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('--clip_grad', type=float, default=3.0, help="""Maximal parameter
            gradient norm if using gradient clipping. Clipping with norm .3 ~ 1.0 can
            help optimization for larger ViT architectures. 0 for disabling.""")
    parser.add_argument('--batch_size_per_gpu', default=64, type=int,
                        help='Per-GPU batch-size : number of distinct images loaded on one GPU.')
    parser.add_argument('--epochs', default=100, type=int, help='Number of epochs of training.')
    parser.add_argument('--freeze_last_layer', default=1, type=int, help="""Number of epochs
            during which we keep the output layer fixed. Typically doing so during
            the first epoch helps training. Try increasing this value if the loss does not decrease.""")
    parser.add_argument("--lr", default=0.0005, type=float, help="""Learning rate at the end of
            linear warmup (highest LR used during training). The learning rate is linearly scaled
            with the batch size, and specified here for a reference batch size of 256.""")
    parser.add_argument("--warmup_epochs", default=10, type=int,
                        help="Number of epochs for the linear learning-rate warm up.")
    parser.add_argument('--min_lr', type=float, default=1e-6, help="""Target LR at the
            end of optimization. We use a cosine LR schedule with linear warmup.""")
    parser.add_argument('--optimizer', default='adamw', type=str,
                        choices=['adamw', 'sgd', 'lars'],
                        help="""Type of optimizer. We recommend using adamw with ViTs.""")
    parser.add_argument('--drop_path_rate', type=float, default=0.1, help="stochastic depth rate")

    # Multi-crop parameters
    parser.add_argument('--global_crops_scale', type=float, nargs='+', default=(0.4, 1.),
                        help="""Scale range of the cropped image before resizing, relatively to the origin image.
            Used for large global view cropping. When disabling multi-crop (--local_crops_number 0), we
            recommand using a wider range of scale ("--global_crops_scale 0.14 1." for example)""")
    parser.add_argument('--local_crops_number', type=int, default=8, help="""Number of small
            local views to generate. Set this parameter to 0 to disable multi-crop training.
            When disabling multi-crop we recommend to use "--global_crops_scale 0.14 1." """)
    parser.add_argument('--local_crops_scale', type=float, nargs='+', default=(0.05, 0.4),
                        help="""Scale range of the cropped image before resizing, relatively to the origin image.
            Used for small local view cropping of multi-crop.""")

    # Misc
    parser.add_argument('--data_path', default='/path/to/imagenet/train/', type=str,
                        help='Please specify path to the ImageNet training data.')
    parser.add_argument('--output_dir', default=".", type=str, help='Path to save logs and checkpoints.')
    parser.add_argument('--saveckp_freq', default=20, type=int, help='Save checkpoint every x epochs.')
    parser.add_argument('--seed', default=0, type=int, help='Random seed.')
    parser.add_argument('--num_workers', default=10, type=int, help='Number of data loading workers per GPU.')
    parser.add_argument("--dist_url", default="env://", type=str, help="""url used to set up
            distributed training; see https://pytorch.org/docs/stable/distributed.html""")
    parser.add_argument("--local_rank", default=0, type=int, help="Please ignore and do not set this argument.")
    args = parser.parse_args()
    config = Config(args.config)
    return config


def _get_databaunch(config):
    def _get_dataset(ds_type, paths, is_training, config, **kwargs):
        kwargs.update({
            'img_h': config.dataset_image_height,
            'img_w': config.dataset_image_width,
            'max_length': config.dataset_max_length,
            'case_sensitive': config.dataset_case_sensitive,
            'charset_path': config.dataset_charset_path,
            'data_aug': config.dataset_data_aug,
            'deteriorate_ratio': config.dataset_deteriorate_ratio,
            'multiscales': config.dataset_multiscales,
            'data_portion': config.dataset_portion,
            'filter_single_punctuation': config.dataset_filter_single_punctuation,
            'mask': config.dataset_mask,
            'mask_path': config.dataset_mask_path,
        })
        datasets = []
        for p in paths:
            subfolders = [f.path for f in os.scandir(p) if f.is_dir()]
            if subfolders:  # Concat all subfolders
                datasets.append(_get_dataset(ds_type, subfolders, is_training, config, **kwargs))
            else:
                datasets.append(ds_type(path=p, is_training=is_training, **kwargs))
        if len(datasets) > 1:
            return MyConcatDataset(datasets)
        else:
            return datasets[0]

    bunch_kwargs = {}
    ds_kwargs = {}
    bunch_kwargs['collate_fn'] = collate_fn_filter_none
    dataset_class = ImageDatasetSelfSupervisedKmeans
    if config.dataset_augmentation_severity is not None:
        ds_kwargs['augmentation_severity'] = config.dataset_augmentation_severity
    ds_kwargs['supervised_flag'] = ifnone(config.model_contrastive_supervised_flag, False)
    train_ds = _get_dataset(dataset_class, config.dataset_train_roots, True, config, **ds_kwargs)
    sampler = torch.utils.data.DistributedSampler(train_ds, shuffle=True)
    train_dataloader = torch.utils.data.DataLoader(
        train_ds,
        sampler=sampler,
        batch_size=config.batch_size_per_gpu,
        num_workers=config.dataset_num_workers,
        collate_fn=collate_fn_filter_none,
        pin_memory=config.dataset_pin_memory,
        drop_last=True,
    )
    return train_dataloader


if __name__ == "__main__":
    config = _parse_arguments()
    Logger.init(config.global_workdir, config.global_name, config.global_phase)
    Logger.enable_file()
    # _set_random_seed(config.global_seed)
    logging.info(config)
    os.makedirs(f"./saved_models/{config.global_name}", exist_ok=True)
    os.makedirs(f"./tensorboard", exist_ok=True)
    config.writer = SummaryWriter(log_dir=f"./tensorboard/{config.global_name}")
    train(config)