改动

1.数据处理

整理好数据集按照要求：

先运行

import os
import tarfile

TRAIN_SRC_DIR = '/root/autodl-pub/ImageNet/ILSVRC2012/ILSVRC2012_img_train.tar'
TRAIN_DEST_DIR = '/root/autodl-tmp/imagenet/train'
VAL_SRC_DIR = '/root/autodl-pub/ImageNet/ILSVRC2012/ILSVRC2012_img_val.tar'
VAL_DEST_DIR = '/root/autodl-tmp/imagenet/val'


def extract_train():
    with open(TRAIN_SRC_DIR, 'rb') as f:
        tar = tarfile.open(fileobj=f, mode='r:')
        for i, item in enumerate(tar):
            cls_name = item.name.strip(".tar")
            a = tar.extractfile(item)
            b = tarfile.open(fileobj=a, mode="r:")
            e_path = "{}/{}/".format(TRAIN_DEST_DIR, cls_name)
            if not os.path.isdir(e_path):
                os.makedirs(e_path)
            print("#", i, "extract train dateset to >>>", e_path)
            names = b.getnames()
            for name in names:
                b.extract(name, e_path)


def extract_val():
    with open(VAL_SRC_DIR, 'rb') as f:
        tar = tarfile.open(fileobj=f, mode='r:')
        if not os.path.isdir(VAL_DEST_DIR):
            os.makedirs(VAL_DEST_DIR)
        print("extract val dateset to >>>", VAL_DEST_DIR)
        names = tar.getnames()
        for name in names:
            tar.extract(name, VAL_DEST_DIR)


if __name__ == '__main__':
    extract_train()
    extract_val()


​

再运行

import os
import tarfile

# TRAIN_SRC_DIR = '/root/autodl-pub/ImageNet/ILSVRC2012/ILSVRC2012_img_train.tar'
# TRAIN_DEST_DIR = '/root/autodl-tmp/imagenet/train'
VAL_SRC_DIR = '/root/autodl-pub/ImageNet/ILSVRC2012/ILSVRC2012_img_test.tar'
VAL_DEST_DIR = '/root/autodl-tmp/imagenet/test'


# def extract_train():
#     with open(TRAIN_SRC_DIR, 'rb') as f:
#         tar = tarfile.open(fileobj=f, mode='r:')
#         for i, item in enumerate(tar):
#             cls_name = item.name.strip(".tar")
#             a = tar.extractfile(item)
#             b = tarfile.open(fileobj=a, mode="r:")
#             e_path = "{}/{}/".format(TRAIN_DEST_DIR, cls_name)
#             if not os.path.isdir(e_path):
#                 os.makedirs(e_path)
#             print("#", i, "extract train dateset to >>>", e_path)
#             names = b.getnames()
#             for name in names:
#                 b.extract(name, e_path)


def extract_val():
    with open(VAL_SRC_DIR, 'rb') as f:
        tar = tarfile.open(fileobj=f, mode='r:')
        if not os.path.isdir(VAL_DEST_DIR):
            os.makedirs(VAL_DEST_DIR)
        print("extract val dateset to >>>", VAL_DEST_DIR)
        names = tar.getnames()
        for name in names:
            tar.extract(name, VAL_DEST_DIR)


if __name__ == '__main__':
    # extract_train()
    extract_val()

​

处理好数据 差一个 label 文件

‍

我帮你处理好了 处理过程就不说了比较繁琐

2.使用这个 生成 extra 文件夹用于训练

from dinov2.data.datasets import ImageNet

for split in ImageNet.Split:
    dataset = ImageNet(split=split, root="/root/autodl-tmp/imagenet", extra="/root/autodl-tmp/extra")
    dataset.dump_extra()
  
​

过程中会报错 label

‍

在报错位置

​class_id, class_name = row​

修改为

​class_id, class_name,*_ = row​

‍

3.OK 环境已经配好

如果需要重新配

输入

conda env create -f conda.yaml
conda activate dinov2

即可

‍

运行过程中会报字符串错误 将报错位置为止修改为：

def remove_suffix(s, suffix):
if s.endswith(suffix):
return s[:-len(suffix)]
return s

args.arch = remove_suffix(args.arch, "_memeff")

‍

‍

4.运行

‍

github 给出的运行代码是在集群运行我们没法用

下面是单卡运行

配置我写好了：

首先

cd /root/dinov2

python setup.py install （已经做过了 不用重复做）

然后 cd 到/root/dinov2/dinov2/train

source activte base

python main.py

直接跑起来了就

‍

5.配置

vitl16_short.yaml 修改为

train:
  dataset_path: ImageNet:split=TRAIN:root=/root/autodl-tmp/imagenet:extra=/root/autodl-tmp/extra
  batch_size_per_gpu: 8
student:
  block_chunks: 1
​

train.py 修改为

‍

parser.add_argument("--config-file", default="/root/dinov2/dinov2/configs/train/vitl16_short.yaml", metavar="FILE", help="path to config file")`  
parser.add_argument(
        "--output-dir",
        "--output_dir",
        default="~/output",
        type=str,
        help="Output directory to save logs and checkpoints",
    )
​

‍