main_dino.py

# Copyright (c) Facebook, Inc. and its affiliates.
# 
# 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 argparse
import os
import sys
import datetime
import time
import yaml
import json
from tqdm import tqdm
from pathlib import Path

import numpy as np
from PIL import Image
import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
import torch.nn.functional as F
from torchvision import datasets, transforms
from torchvision import models as torchvision_models
from torch.utils.tensorboard import SummaryWriter

import utils
import vision_transformer as vits
from vision_transformer import DINOHead

import math
import numbers
import random
import warnings
from collections.abc import Sequence
from typing import Tuple, List, Optional

from torch import Tensor

try:
    import accimage
except ImportError:
    accimage = None

from torchvision.transforms import functional as torchvision_F
from torchvision.transforms.functional import InterpolationMode, _interpolation_modes_from_int
from torchvision.transforms import RandomResizedCrop

torchvision_archs = sorted(name for name in torchvision_models.__dict__
                           if name.islower() and not name.startswith("__")
                           and callable(torchvision_models.__dict__[name]))


def get_args_parser():
    parser = argparse.ArgumentParser('DINO', add_help=False)

    # Model parameters
    parser.add_argument('--arch', default='vit_small', type=str,
                        choices=['vit_tiny', 'vit_small', 'vit_base', 'deit_tiny', 'deit_small'] + torchvision_archs,
                        help="""Name of architecture to train. For quick experiments with ViTs,
        we recommend using vit_tiny or vit_small.""")
    parser.add_argument('--pretrained_weights', default=None, type=str, help="Path to weights of pretrained models, "
                                                                             "will start training with them if selected")
    parser.add_argument('--patch_size', default=16, 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', default=64, type=int,
                        help='Per-GPU batch-size : number of distinct images loaded on one GPU.')
    parser.add_argument('--accum_iter', default=8, type=int)
    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.""")

    # 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=16, type=int, help='Number of data loading workers per GPU.')
    parser.add_argument("--gpus", default=('0', '1'), type=int, nargs='+', help='GPUs ids  to use for training')
    parser.add_argument("--local_rank", default=0, type=int)

    return parser


def train_dino(args):
    utils.fix_random_seeds(args.seed)
    print("\n".join("%s: %s" % (k, str(v)) for k, v in sorted(dict(vars(args)).items())))
    cudnn.benchmark = True

    with (Path(args.output_dir) / "config.yaml").open("w") as fp:
        yaml.dump(args, fp)
    writer = SummaryWriter(log_dir=args.output_dir)

    # ============ preparing data ... ============
    transform = DataAugmentationDINO(
        args.global_crops_scale,
        args.local_crops_scale,
        args.local_crops_number,
    )
    dataset = datasets.ImageFolder(args.data_path, transform=transform)
    sampler = torch.utils.data.distributed.DistributedSampler(dataset)
    data_loader = torch.utils.data.DataLoader(
        dataset,
        sampler=sampler,
        batch_size=args.batch_size,
        num_workers=args.num_workers,
        pin_memory=True,
        drop_last=True,
    )
    print(f"Data loaded: there are {len(dataset)} images.")

    # ============ building student and teacher networks ... ============
    # we changed the name DeiT-S for ViT-S to avoid confusions
    args.arch = args.arch.replace("deit", "vit")
    # if the network is a vision transformer (i.e. vit_tiny, vit_small, vit_base)
    if args.arch in vits.__dict__.keys():
        student = vits.__dict__[args.arch](
            patch_size=args.patch_size,
            drop_path_rate=0.1,  # stochastic depth
        )
        teacher = vits.__dict__[args.arch](patch_size=args.patch_size)
        embed_dim = student.embed_dim
    # otherwise, we check if the architecture is in torchvision models
    elif args.arch in torchvision_models.__dict__.keys():
        student = torchvision_models.__dict__[args.arch]()
        teacher = torchvision_models.__dict__[args.arch]()
        embed_dim = student.fc.weight.shape[1]
    else:
        print(f"Unknow architecture: {args.arch}")

    # multi-crop wrapper handles forward with inputs of different resolutions
    student = utils.MultiCropWrapper(student, DINOHead(
        embed_dim,
        args.out_dim,
        use_bn=args.use_bn_in_head,
        norm_last_layer=args.norm_last_layer,
    ))
    teacher = utils.MultiCropWrapper(
        teacher,
        DINOHead(embed_dim, args.out_dim, args.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)

    if len(args.gpus) > 1:
        teacher = nn.parallel.DistributedDataParallel(teacher, device_ids=[args.gpus])
        student = nn.parallel.DistributedDataParallel(student, device_ids=[args.gpus])

    # 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 {args.arch} network.")

    # ============ preparing loss ... ============
    dino_loss = DINOLoss(
        args.out_dim,
        args.local_crops_number + 2,  # total number of crops = 2 global crops + local_crops_number
        args.warmup_teacher_temp,
        args.teacher_temp,
        args.warmup_teacher_temp_epochs,
        args.epochs,
    ).cuda()

    # ============ preparing optimizer ... ============
    params_groups = utils.get_params_groups(student)
    if args.optimizer == "adamw":
        optimizer = torch.optim.AdamW(params_groups)  # to use with ViTs
    elif args.optimizer == "sgd":
        optimizer = torch.optim.SGD(params_groups, lr=0, momentum=0.9)  # lr is set by scheduler
    elif args.optimizer == "lars":
        optimizer = utils.LARS(params_groups)  # to use with convnet and large batches
    # for mixed precision training
    fp16_scaler = None
    if args.use_fp16:
        fp16_scaler = torch.cuda.amp.GradScaler()

    # ============ init schedulers ... ============
    lr_schedule = utils.cosine_scheduler(
        args.lr,  # linear scaling rule
        args.min_lr,
        args.epochs, len(data_loader),
        warmup_epochs=args.warmup_epochs,
    )
    wd_schedule = utils.cosine_scheduler(
        args.weight_decay,
        args.weight_decay_end,
        args.epochs, len(data_loader),
    )
    # momentum parameter is increased to 1. during training with a cosine schedule
    momentum_schedule = utils.cosine_scheduler(args.momentum_teacher, 1,
                                               args.epochs, len(data_loader))
    print(f"Loss, optimizer and schedulers ready.")

    # ============ optionally resume training ... ============
    to_restore = {"epoch": 0}
    utils.restart_from_checkpoint(
        os.path.join(args.output_dir, "checkpoint.pth"),
        run_variables=to_restore,
        student=student,
        teacher=teacher,
        optimizer=optimizer,
        fp16_scaler=fp16_scaler,
        dino_loss=dino_loss,
    )
    start_epoch = to_restore["epoch"]

    if args.pretrained_weights:
        state_dict = torch.load(args.pretrained_weights, map_location=primary_gpu)
        state_dict_student, state_dict_teacher = state_dict['student'], state_dict['teacher']

        # remove `module.` prefix
        state_dict_student = {k.replace("module.", ""): v for k, v in state_dict_student.items()}
        state_dict_teacher = {k.replace("module.", ""): v for k, v in state_dict_teacher.items()}

        # remove `backbone.` prefix induced by multicrop wrapper
        state_dict_teacher = {k.replace("backbone.", ""): v for k, v in state_dict_teacher.items()}
        state_dict_student = {k.replace("backbone.", ""): v for k, v in state_dict_student.items()}

        student.load_state_dict(state_dict_student, strict=False)
        teacher.load_state_dict(state_dict_teacher, strict=False)
        del state_dict
        del state_dict_student
        del state_dict_teacher
        print('Pretrained weights found and loaded  for student and for teacher')

    start_time = time.time()
    print("Starting DINO training !")
    for epoch in range(start_epoch, args.epochs):

        # ============ training one epoch of DINO ... ============
        train_stats = train_one_epoch(student, teacher, dino_loss,
                                      data_loader, optimizer, lr_schedule, wd_schedule, momentum_schedule,
                                      epoch, fp16_scaler, args, writer)

        # ============ writing logs ... ============
        save_dict = {
            'student': student.state_dict(),
            'teacher': teacher.state_dict(),
            'optimizer': optimizer.state_dict(),
            'epoch': epoch + 1,
            'args': args,
            'dino_loss': dino_loss.state_dict(),
        }
        writer.add_scalar("Loss/train", train_stats['loss_total'], epoch)
        writer.add_scalar("Learning Rate", train_stats['lr'], epoch)

        if fp16_scaler is not None:
            save_dict['fp16_scaler'] = fp16_scaler.state_dict()
        torch.save(save_dict, os.path.join(args.output_dir, 'checkpoint.pth'))
        if args.saveckp_freq and epoch % args.saveckp_freq == 0:
            torch.save(save_dict, os.path.join(args.output_dir, f'checkpoint{epoch:04}.pth'))
        log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
                     'epoch': epoch}
        with (Path(args.output_dir) / "log.txt").open("a") 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))
    writer.flush()
    writer.close()


def train_one_epoch(student, teacher, dino_loss, data_loader,
                    optimizer, lr_schedule, wd_schedule, momentum_schedule, epoch,
                    fp16_scaler, args, logger):
    header = 'Epoch: [{}/{}]'.format(epoch, args.epochs)
    loss_history = []
    with tqdm(data_loader, unit="batch") as tepoch:
        for it, (images, _) in enumerate(tepoch):

            # update weight decay and learning rate according to their schedule
            it = len(data_loader) * epoch + it  # global training iteration
            for i, param_group in enumerate(optimizer.param_groups):
                param_group["lr"] = lr_schedule[it]
                if i == 0:  # only the first group is regularized
                    param_group["weight_decay"] = wd_schedule[it]

            if it % 200 == 0:
                mean, std = 0.5, 0.25
                rnd_image_id = random.randint(0, args.batch_size - 1)
                sample_images_global = np.array([np.array(i) for i in images[:2]]
                                                )[:, rnd_image_id, :, :, :]
                sample_images_local = np.array([np.array(i) for i in images[2:]]
                                               )[:, rnd_image_id, :, :, :]
                logger.add_images('Global views', sample_images_global * std + mean, it)
                logger.add_images('Local views', sample_images_local * std + mean, it)

            # move images to gpu
            images = [im.cuda(non_blocking=True) for im in images]
            # teacher and student forward passes + compute dino loss
            with torch.cuda.amp.autocast(fp16_scaler is not None):
                teacher_output = teacher(images[:2])  # only the 2 global views pass through the teacher
                student_output = student(images)
                loss = dino_loss(student_output, teacher_output, epoch)
                loss_history.append(loss.item())

            if not math.isfinite(loss.item()):
                print("Loss is {}, stopping training".format(loss.item()), force=True)
                sys.exit(1)

            if fp16_scaler is None:
                loss.backward()
            else:
                fp16_scaler.scale(loss).backward()

            utils.cancel_gradients_last_layer(epoch, student,
                                              args.freeze_last_layer)

            if (it + 1) % args.accum_iter == 0:

                if args.clip_grad:
                    if fp16_scaler is not None:
                        fp16_scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place
                    param_norms = utils.clip_gradients(student, args.clip_grad)

                # utils.cancel_gradients_last_layer(epoch, student,
                #                                   args.freeze_last_layer)

                if fp16_scaler is not None:
                    fp16_scaler.step(optimizer)
                    fp16_scaler.update()
                else:
                    optimizer.step()

                # optimizer.zero_grad()

                # EMA update for the teacher
                with torch.no_grad():
                    m = momentum_schedule[it]  # momentum parameter
                    stud_params = student.parameters() if len(args.gpus) == 1 else student.module.parameters()
                    for param_q, param_k in zip(stud_params, teacher.parameters()):
                        param_k.data.mul_(m).add_((1 - m) * param_q.detach().data)

                optimizer.zero_grad()

            # logging
            tepoch.set_description(header)
            tepoch.set_postfix(loss=loss.item())

    return {
        'loss': loss.item(),
        'loss_total': np.mean(loss_history),
        'lr': optimizer.param_groups[0]["lr"],
        'wd': optimizer.param_groups[0]["weight_decay"]
    }


class DINOLoss(nn.Module):
    def __init__(self, out_dim, ncrops, warmup_teacher_temp, teacher_temp,
                 warmup_teacher_temp_epochs, nepochs, student_temp=0.1,
                 center_momentum=0.9):
        super().__init__()
        self.student_temp = student_temp
        self.center_momentum = center_momentum
        self.ncrops = ncrops
        self.register_buffer("center", torch.zeros(1, out_dim))
        # we apply a warm up for the teacher temperature because
        # a too high temperature makes the training instable at the beginning
        self.teacher_temp_schedule = np.concatenate((
            np.linspace(warmup_teacher_temp,
                        teacher_temp, warmup_teacher_temp_epochs),
            np.ones(nepochs - warmup_teacher_temp_epochs) * teacher_temp
        ))

    def forward(self, student_output, teacher_output, epoch):
        """
        Cross-entropy between softmax outputs of the teacher and student networks.
        """
        student_out = student_output / self.student_temp
        student_out = student_out.chunk(self.ncrops)

        # teacher centering and sharpening
        temp = self.teacher_temp_schedule[epoch]
        teacher_out = F.softmax((teacher_output - self.center) / temp, dim=-1)
        teacher_out = teacher_out.detach().chunk(2)

        total_loss = 0
        n_loss_terms = 0
        for iq, q in enumerate(teacher_out):
            for v in range(len(student_out)):
                if v == iq:
                    # we skip cases where student and teacher operate on the same view
                    continue
                loss = torch.sum(-q * F.log_softmax(student_out[v], dim=-1), dim=-1)
                total_loss += loss.mean()
                n_loss_terms += 1
        del teacher_out
        del student_out
        del loss
        total_loss /= n_loss_terms
        self.update_center(teacher_output)
        return total_loss

    @torch.no_grad()
    def update_center(self, teacher_output):
        """
        Update center used for teacher output.
        """
        batch_center = torch.sum(teacher_output, dim=0, keepdim=True)
        batch_center = batch_center / (len(teacher_output))

        # ema update
        self.center = self.center * self.center_momentum + batch_center * (1 - self.center_momentum)


def _setup_size(size, error_msg):
    if isinstance(size, numbers.Number):
        return int(size), int(size)

    if isinstance(size, Sequence) and len(size) == 1:
        return size[0], size[0]

    if len(size) != 2:
        raise ValueError(error_msg)

    return size


class QuarterRandomResizedCrop(RandomResizedCrop):

    def __init__(self, size, quarter=-1, scale=(0.08, 1.0), ratio=(3. / 4., 4. / 3.),
                 interpolation=InterpolationMode.BILINEAR):
        super().__init__(size, scale, ratio, interpolation)
        self.size = _setup_size(size, error_msg="Please provide only two dimensions (h, w) for size.")

        if not isinstance(scale, Sequence):
            raise TypeError("Scale should be a sequence")
        if not isinstance(ratio, Sequence):
            raise TypeError("Ratio should be a sequence")
        if (scale[0] > scale[1]) or (ratio[0] > ratio[1]):
            warnings.warn("Scale and ratio should be of kind (min, max)")

        # Backward compatibility with integer value
        if isinstance(interpolation, int):
            warnings.warn(
                "Argument interpolation should be of type InterpolationMode instead of int. "
                "Please, use InterpolationMode enum."
            )
            interpolation = _interpolation_modes_from_int(interpolation)

        self.interpolation = interpolation
        self.scale = scale
        self.ratio = ratio
        self.quarter = quarter

    @staticmethod
    def get_bounds_for_quarter(width, height, quarter):
        if quarter == 1:
            return 0,  width // 4, 0, height // 2
        if quarter == 2:
            return (3 * width) // 4, width, 0, height // 2
        if quarter == 3:
            return 0, width // 4, height // 2, height
        if quarter == 0:
            return (3 * width) // 4, width, height // 2, height
        else:
            return width, height

    @staticmethod
    def get_params(
            img: Tensor, scale: List[float], ratio: List[float], quarter: int,
    ) -> Tuple[int, int, int, int]:
        """Get parameters for ``crop`` for a random sized crop.

        Args:
            img (PIL Image or Tensor): Input image.
            scale (list): range of scale of the origin size cropped
            ratio (list): range of aspect ratio of the origin aspect ratio cropped

        Returns:
            tuple: params (i, j, h, w) to be passed to ``crop`` for a random
            sized crop.
            :param quarter:
        """
        width, height = torchvision_F._get_image_size(img)
        area = height * width

        log_ratio = torch.log(torch.tensor(ratio))
        for _ in range(10):
            target_area = area * torch.empty(1).uniform_(scale[0], scale[1]).item()
            aspect_ratio = torch.exp(
                torch.empty(1).uniform_(log_ratio[0], log_ratio[1])
            ).item()

            w = int(round(math.sqrt(target_area * aspect_ratio)))
            h = int(round(math.sqrt(target_area / aspect_ratio)))

            width_start, width_finish, height_start, height_finish = \
                QuarterRandomResizedCrop.get_bounds_for_quarter(w, h, quarter)

            if 0 < w <= width and 0 < h <= height:
                #                 i = torch.randint(0, height - h + 1, size=(1,)).item()
                #                 j = torch.randint(0, width - w + 1, size=(1,)).item()
                i = torch.randint(height_start, height_finish + 1, size=(1,)).item()
                j = torch.randint(width_start, width_finish + 1, size=(1,)).item()
                return i, j, h, w

        # Fallback to central crop
        in_ratio = float(width) / float(height)
        if in_ratio < min(ratio):
            w = width
            h = int(round(w / min(ratio)))
        elif in_ratio > max(ratio):
            h = height
            w = int(round(h * max(ratio)))
        else:  # whole image
            w = width
            h = height
        i = (height - h) // 2
        j = (width - w) // 2
        return i, j, h, w

    def forward(self, img):
        """
        Args:
            img (PIL Image or Tensor): Image to be cropped and resized.

        Returns:
            PIL Image or Tensor: Randomly cropped and resized image.
        """
        i, j, h, w = QuarterRandomResizedCrop.get_params(img, self.scale, self.ratio, self.quarter)
        return torchvision_F.resized_crop(img, i, j, h, w, self.size, self.interpolation)


class DataAugmentationDINO(object):
    def __init__(self, global_crops_scale, local_crops_scale, local_crops_number):
        flip_and_color_jitter = transforms.Compose([
            transforms.RandomHorizontalFlip(p=0.5),
            transforms.RandomApply(
                [transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.2, hue=0.1)],
                p=0.8
            ),
            transforms.RandomGrayscale(p=0.2),
        ])
        normalize = transforms.Compose([
            transforms.ToTensor(),
            # transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
            transforms.Normalize((0.5), (0.25)),
        ])

        # first global crop
        self.global_transfo1 = transforms.Compose([
            transforms.RandomResizedCrop(224, scale=global_crops_scale, interpolation=Image.BICUBIC),
            flip_and_color_jitter,
            utils.GaussianBlur(1.0),
            normalize,
        ])
        # second global crop
        self.global_transfo2 = transforms.Compose([
            transforms.RandomResizedCrop(224, scale=global_crops_scale, interpolation=Image.BICUBIC),
            flip_and_color_jitter,
            utils.GaussianBlur(0.1),
            utils.Solarization(0.2),
            normalize,
        ])
        # transformation for the local small crops
        self.local_crops_number = local_crops_number
        self.local_transfo = [transforms.Compose([
            RandomResizedCrop(96, scale=local_crops_scale, interpolation=Image.BICUBIC),
            flip_and_color_jitter,
            utils.GaussianBlur(p=0.5),
            normalize,
        ]) for i in range(self.local_crops_number)]

    def __call__(self, image):
        crops = []
        crops.append(self.global_transfo1(image))
        crops.append(self.global_transfo2(image))
        for transform in self.local_transfo:
            crops.append(transform(image))
        return crops


if __name__ == '__main__':
    parser = argparse.ArgumentParser('DINO', parents=[get_args_parser()])
    args = parser.parse_args()
    os.environ['CUDA_VISIBLE_DEVICES'] = ','.join([str(x) for x in args.gpus])
    Path(args.output_dir).mkdir(parents=True, exist_ok=True)
    torch.distributed.init_process_group(backend='nccl')
    torch.cuda.set_device(args.local_rank)
    train_dino(args)