add swinv2_loader into libai

[shaoshitong]

add swinv2_loader into libai, make it can load huggingface's weight into libai

[shaoshitong]

测试代码：

import logging
import os
import sys

sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), os.path.pardir)))
from libai.config import LazyConfig, default_argument_parser, try_get_key
from configs.common.models.swinv2.swinv2_tiny_patch4_window8_256 import cfg as model_cfg
from libai.engine import DefaultTrainer, default_setup
from libai.utils.checkpoint import Checkpointer
from libai.models.utils import SwinV2LoaderHuggerFace
from libai.models import SwinTransformerV2
logger = logging.getLogger("libai." + __name__)


def main(args):
    cfg = LazyConfig.load(args.config_file)
    cfg = LazyConfig.apply_overrides(cfg, args.opts)
    default_setup(cfg, args)

    if args.fast_dev_run:
        cfg.train.train_epoch = 0
        cfg.train.train_iter = 20
        cfg.train.evaluation.eval_period = 10
        cfg.train.log_period = 1

    if args.eval_only:
        tokenizer = None
        if try_get_key(cfg, "tokenization") is not None:
            tokenizer = DefaultTrainer.build_tokenizer(cfg)
        loader = SwinV2LoaderHuggerFace(
        SwinTransformerV2,
        model_cfg,
        "/root/.cache/huggingface/hub/models--microsoft--swinv2-tiny-patch4-window8-256/snapshots/2b979ac403df19f72443cd151e9e957842eb9645",
         )
        model = loader.load()
        if try_get_key(cfg, "train.graph.enabled", default=False):
            model = DefaultTrainer.build_graph(cfg, model, is_train=False)
        test_loader = DefaultTrainer.build_test_loader(cfg, tokenizer)
        if len(test_loader) == 0:
            logger.info("No dataset in dataloader.test, please set dataset for dataloader.test")
        _ = DefaultTrainer.test(cfg, test_loader, model)
        return

    trainer = DefaultTrainer(cfg)
    return trainer.train()


if __name__ == "__main__":
    args = default_argument_parser().parse_args()
    main(args)

脚本命令：
bash tools/train.sh test_swinv2.py configs/swinv2_imagenet.py 1 --eval-only
测试结果：

[08/12 23:44:36 lb.engine.default]: Prepare testing set
[08/12 23:44:36 lb.evaluation.evaluator]: with eval_iter 100000.0, reset total samples 50000 to 50000
[08/12 23:44:36 lb.evaluation.evaluator]: Start inference on 50000 samples
[08/12 23:44:43 lb.evaluation.evaluator]: Inference done 1408/50000. Dataloading: 0.0007 s/iter. Inference: 0.0187 s/iter. Eval: 0.2582 s/iter. Total: 0.2776 s/iter. ETA=0:01:45
[08/12 23:44:48 lb.evaluation.evaluator]: Inference done 3712/50000. Dataloading: 0.0008 s/iter. Inference: 0.0159 s/iter. Eval: 0.2626 s/iter. Total: 0.2794 s/iter. ETA=0:01:40
[08/12 23:44:53 lb.evaluation.evaluator]: Inference done 6016/50000. Dataloading: 0.0008 s/iter. Inference: 0.0171 s/iter. Eval: 0.2630 s/iter. Total: 0.2810 s/iter. ETA=0:01:36
[08/12 23:44:58 lb.evaluation.evaluator]: Inference done 8320/50000. Dataloading: 0.0008 s/iter. Inference: 0.0165 s/iter. Eval: 0.2642 s/iter. Total: 0.2815 s/iter. ETA=0:01:31
[08/12 23:45:04 lb.evaluation.evaluator]: Inference done 10624/50000. Dataloading: 0.0009 s/iter. Inference: 0.0170 s/iter. Eval: 0.2644 s/iter. Total: 0.2824 s/iter. ETA=0:01:26
[08/12 23:45:09 lb.evaluation.evaluator]: Inference done 12928/50000. Dataloading: 0.0009 s/iter. Inference: 0.0171 s/iter. Eval: 0.2647 s/iter. Total: 0.2827 s/iter. ETA=0:01:21
[08/12 23:45:14 lb.evaluation.evaluator]: Inference done 15232/50000. Dataloading: 0.0009 s/iter. Inference: 0.0170 s/iter. Eval: 0.2650 s/iter. Total: 0.2829 s/iter. ETA=0:01:16
[08/12 23:45:19 lb.evaluation.evaluator]: Inference done 17536/50000. Dataloading: 0.0009 s/iter. Inference: 0.0176 s/iter. Eval: 0.2648 s/iter. Total: 0.2834 s/iter. ETA=0:01:11
[08/12 23:45:24 lb.evaluation.evaluator]: Inference done 19840/50000. Dataloading: 0.0009 s/iter. Inference: 0.0182 s/iter. Eval: 0.2645 s/iter. Total: 0.2837 s/iter. ETA=0:01:06
[08/12 23:45:29 lb.evaluation.evaluator]: Inference done 22144/50000. Dataloading: 0.0009 s/iter. Inference: 0.0185 s/iter. Eval: 0.2644 s/iter. Total: 0.2839 s/iter. ETA=0:01:01
[08/12 23:45:34 lb.evaluation.evaluator]: Inference done 24448/50000. Dataloading: 0.0010 s/iter. Inference: 0.0188 s/iter. Eval: 0.2646 s/iter. Total: 0.2844 s/iter. ETA=0:00:56
[08/12 23:45:40 lb.evaluation.evaluator]: Inference done 26752/50000. Dataloading: 0.0009 s/iter. Inference: 0.0185 s/iter. Eval: 0.2650 s/iter. Total: 0.2846 s/iter. ETA=0:00:51
[08/12 23:45:45 lb.evaluation.evaluator]: Inference done 29056/50000. Dataloading: 0.0009 s/iter. Inference: 0.0183 s/iter. Eval: 0.2654 s/iter. Total: 0.2847 s/iter. ETA=0:00:46
[08/12 23:45:50 lb.evaluation.evaluator]: Inference done 31360/50000. Dataloading: 0.0010 s/iter. Inference: 0.0184 s/iter. Eval: 0.2655 s/iter. Total: 0.2850 s/iter. ETA=0:00:41
[08/12 23:45:55 lb.evaluation.evaluator]: Inference done 33664/50000. Dataloading: 0.0010 s/iter. Inference: 0.0187 s/iter. Eval: 0.2654 s/iter. Total: 0.2851 s/iter. ETA=0:00:36
[08/12 23:46:00 lb.evaluation.evaluator]: Inference done 35968/50000. Dataloading: 0.0010 s/iter. Inference: 0.0186 s/iter. Eval: 0.2655 s/iter. Total: 0.2852 s/iter. ETA=0:00:31
[08/12 23:46:05 lb.evaluation.evaluator]: Inference done 38272/50000. Dataloading: 0.0010 s/iter. Inference: 0.0188 s/iter. Eval: 0.2656 s/iter. Total: 0.2854 s/iter. ETA=0:00:25
[08/12 23:46:11 lb.evaluation.evaluator]: Inference done 40576/50000. Dataloading: 0.0010 s/iter. Inference: 0.0189 s/iter. Eval: 0.2657 s/iter. Total: 0.2857 s/iter. ETA=0:00:20
[08/12 23:46:16 lb.evaluation.evaluator]: Inference done 42880/50000. Dataloading: 0.0010 s/iter. Inference: 0.0188 s/iter. Eval: 0.2659 s/iter. Total: 0.2858 s/iter. ETA=0:00:15
[08/12 23:46:21 lb.evaluation.evaluator]: Inference done 45184/50000. Dataloading: 0.0010 s/iter. Inference: 0.0189 s/iter. Eval: 0.2660 s/iter. Total: 0.2859 s/iter. ETA=0:00:10
[08/12 23:46:26 lb.evaluation.evaluator]: Inference done 47488/50000. Dataloading: 0.0010 s/iter. Inference: 0.0192 s/iter. Eval: 0.2659 s/iter. Total: 0.2862 s/iter. ETA=0:00:05
[08/12 23:46:31 lb.evaluation.evaluator]: Inference done 49664/50000. Dataloading: 0.0010 s/iter. Inference: 0.0192 s/iter. Eval: 0.2664 s/iter. Total: 0.2866 s/iter. ETA=0:00:00
[08/12 23:46:33 lb.evaluation.evaluator]: Total valid samples: 50000
[08/12 23:46:33 lb.evaluation.evaluator]: Total inference time: 0:01:50.978096 (0.002220 s / iter per device, on 1 devices)
[08/12 23:46:33 lb.evaluation.evaluator]: Total inference pure compute time: 0:00:07 (0.000148 s / iter per device, on 1 devices)
[08/12 23:46:33 lb.engine.default]: Evaluation results for ImageNetDataset in csv format:
[08/12 23:46:33 lb.evaluation.utils]: copypaste: Acc@1=81.854
[08/12 23:46:33 lb.evaluation.utils]: copypaste: Acc@5=95.944

标准库的推理精度：

因此对齐了

[xiezipeng-ML]

添加一下测试，类似于swin的这个，其中allclose的数值需要由huggingface模型得到
https://github.com/Oneflow-Inc/libai/blob/main/tests/model_utils/test_swin_loader.py

[shaoshitong]

添加一下测试，类似于swin的这个，其中allclose的数值需要由huggingface模型得到 https://github.com/Oneflow-Inc/libai/blob/main/tests/model_utils/test_swin_loader.py

已经添加