alibaba / easycv

An all-in-one toolkit for computer vision

EasyCV

English | 简体中文

Introduction

EasyCV is an all-in-one computer vision toolbox based on PyTorch, mainly focuses on self-supervised learning, transformer based models, and major CV tasks including image classification, metric-learning, object detection, pose estimation, and so on.

Major features

• SOTA SSL Algorithms

  EasyCV provides state-of-the-art algorithms in self-supervised learning based on contrastive learning such as SimCLR, MoCO V2, Swav, DINO, and also MAE based on masked image modeling. We also provide standard benchmarking tools for ssl model evaluation.

• Vision Transformers

  EasyCV aims to provide an easy way to use the off-the-shelf SOTA transformer models trained either using supervised learning or self-supervised learning, such as ViT, Swin Transformer, and DETR Series. More models will be added in the future. In addition, we support all the pretrained models from timm.

• Functionality & Extensibility

  In addition to SSL, EasyCV also supports image classification, object detection, metric learning, and more areas will be supported in the future. Although covering different areas, EasyCV decomposes the framework into different components such as dataset, model and running hook, making it easy to add new components and combining it with existing modules.

  EasyCV provides simple and comprehensive interface for inference. Additionally, all models are supported on PAI-EAS, which can be easily deployed as online service and support automatic scaling and service monitoring.

• Efficiency

  EasyCV supports multi-gpu and multi-worker training. EasyCV uses DALI to accelerate data io and preprocessing process, and uses TorchAccelerator and fp16 to accelerate training process. For inference optimization, EasyCV exports model using jit script, which can be optimized by PAI-Blade

What's New

[🔥 2023.05.09]

• 09/05/2023 EasyCV v0.11.0 was released.

• Support EasyCV as a plug-in for [modelscope](https://github.com/modelscope/modelscope.

[🔥 2023.03.06]

• 06/03/2023 EasyCV v0.10.0 was released.

• Add segmentation model STDC
• Add skeleton based video recognition model STGCN
• Support ReID and Multi-len MOT

[🔥 2023.01.17]

• 17/01/2023 EasyCV v0.9.0 was released.

• Support Single-lens MOT
• Support video recognition (X3D, SWIN-video)

[🔥 2022.12.02]

• 02/12/2022 EasyCV v0.8.0 was released.

• bevformer-base NDS increased by 0.8 on nuscenes val, training speed increased by 10%, and inference speed increased by 40%.
• Support Objects365 pretrain and Adding the DINO++ model can achieve an accuracy of 63.4mAP at a model scale of 200M(Under the same scale, the accuracy is the best).

[🔥 2022.08.31] We have released our YOLOX-PAI that achieves SOTA results within 40~50 mAP (less than 1ms). And we also provide a convenient and fast export/predictor api for end2end object detection. To get a quick start of YOLOX-PAI, click here!

• 31/08/2022 EasyCV v0.6.0 was released.
  • Release YOLOX-PAI which achieves SOTA results within 40~50 mAP (less than 1ms)
  • Add detection algo DINO which achieves 58.5 mAP on COCO
  • Add mask2former algo
  • Releases imagenet1k, imagenet22k, coco, lvis, voc2012 data with BaiduDisk to accelerate downloading

Please refer to change_log.md for more details and history.

Technical Articles

We have a series of technical articles on the functionalities of EasyCV.

• EasyCV开源｜开箱即用的视觉自监督+Transformer算法库
• MAE自监督算法介绍和基于EasyCV的复现
• 基于EasyCV复现ViTDet：单层特征超越FPN
• 基于EasyCV复现DETR和DAB-DETR，Object Query的正确打开方式
• YOLOX-PAI: 加速YOLOX, 比YOLOv6更快更强
• EasyCV带你复现更好更快的自监督算法-FastConvMAE
• EasyCV DataHub 提供多领域视觉数据集下载，助力模型生产
• 使用EasyCV Mask2Former轻松实现图像分割

Installation

Please refer to the installation section in quick_start.md for installation.

Get Started

Please refer to quick_start.md for quick start. We also provides tutorials for more usages.

• self-supervised learning
• image classification
• metric learning
• object detection with yolox-pai
• model compression with yolox
• using torchacc
• file io for local and oss files
• using mmdetection model in EasyCV
• batch prediction tools

notebook

• self-supervised learning
• image classification
• object detection with yolox-pai
• metric learning

Model Zoo

Architectures

Self-Supervised Learning	Image Classification	Object Detection	Segmentation	Object Detection 3D
BYOL (NeurIPS'2020) DINO (ICCV'2021) MiXCo (NeurIPS'2020) MoBY (ArXiv'2021) MoCov2 (ArXiv'2020) SimCLR (ICML'2020) SwAV (NeurIPS'2020) MAE (CVPR'2022) FastConvMAE (ArXiv'2022)	ResNet (CVPR'2016) ResNeXt (CVPR'2017) HRNet (CVPR'2019) ViT (ICLR'2021) SwinT (ICCV'2021) EfficientFormer (ArXiv'2022) DeiT (ICML'2021) XCiT (ArXiv'2021) TNT (NeurIPS'2021) ConViT (ArXiv'2021) CaiT (ICCV'2021) LeViT (ICCV'2021) ConvNeXt (CVPR'2022) ResMLP (ArXiv'2021) CoaT (ICCV'2021) ConvMixer (ICLR'2022) MLP-Mixer (ArXiv'2021) NesT (AAAI'2022) PiT (ArXiv'2021) Twins (NeurIPS'2021) Shuffle Transformer (ArXiv'2021) DeiT III (ECCV'2022) Hydra Attention (2022)	FCOS (ICCV'2019) YOLOX (ArXiv'2021) YOLOX-PAI (ArXiv'2022) DETR (ECCV'2020) DAB-DETR (ICLR'2022) DN-DETR (CVPR'2022) DINO (ArXiv'2022)	Instance Segmentation Mask R-CNN (ICCV'2017) ViTDet (ArXiv'2022) Mask2Former (CVPR'2022) Semantic Segmentation FCN (CVPR'2015) UperNet (ECCV'2018) Panoptic Segmentation Mask2Former (CVPR'2022)	BEVFormer (ECCV'2022)

Please refer to the following model zoo for more details.

• self-supervised learning model zoo
• classification model zoo
• detection model zoo
• detection3d model zoo
• segmentation model zoo
• pose model zoo

Data Hub

EasyCV have collected dataset info for different scenarios, making it easy for users to finetune or evaluate models in EasyCV model zoo.

Please refer to data_hub.md.

License

This project is licensed under the Apache License (Version 2.0). This toolkit also contains various third-party components and some code modified from other repos under other open source licenses. See the NOTICE file for more information.

Contact

This repo is currently maintained by PAI-CV team, you can contact us by

• Dingding group number: 41783266
• Email: [email protected]

Enterprise Service

If you need EasyCV enterprise service support, or purchase cloud product services, you can contact us by DingDing Group.

easycv's Issues

Accelerate data preprocessing

Disable the multi-processsing of cv2 manually, just involve the cv2.setNumThreads(0) anywhere before involve the cv2.
Refer to: open-mmlab/mmdetection#6867

invalid literal for int() with base 10: 'ok'

https://github.com/alibaba/EasyCV/blob/master/easycv/datasets/classification/data_sources/image_list.py#L93

imagelist should support class_names rather than class ids,

/path/to/a.jpg cls1
/apth/to/img.jpg  cls2

[Algo] ViTDet algo development

reimplement the work described in https://arxiv.org/abs/2203.16527

TorchViTDetPredictor parse input error

When i use TorchViTDetPredictor to predict my images, the inputs cannot be parsed.
I tried two input formats.

Format 1: inputs is np.ndarray:

Raise KeyError: 'LoadImageFromWebcam is not in the pipeline registry'

Format 1: inputs is local file path:

Raise ValueError: unknown url type: '/000000037777.jpg'

My full code is as follows:

import os
import numpy as np
from PIL import Image
from easycv.predictors.detector import TorchViTDetPredictor

model_path='./vitdet_maskrcnn.pt'
predictor = TorchViTDetPredictor(model_path = model_path)
img = './000000037777.jpg'

# format 1
input_data_list = [img]  
# format 2
input_data_list = [np.asarray(Image.open(img))]

output_list = predictor.predict(input_data_list)
print(output_list)

how tot install torchacc

hi, seems torchacc can accelerate training, but how can we train with torchacc, how to install this lib,only through docker image?

[feat]: cpu env support

some user may want to use this library in cpu-only environment due to resource limitation or for better development debugging

'DistEvalHook' object has no attribute 'vis_config'

there is an error while fintune mae model in notebook。

Link error of yoloxs_pai_asff_tood3_8xb16_300e_coco config

Something wrong with the config link, yoloxs_pai_asff_tood3_8xb16_300e_coco in https://github.com/alibaba/EasyCV/blob/master/docs/source/model_zoo_det.md

it is very slow to scan samples of detection dataset

Passing the class function to the multi-process pool causes slow operation. The class function should be extract from class into independent function.

modify load pretrained

raise '%s and %s not Found' % (ori_filename, filename)

Some scripts written in docs are not found

Hi developers:

Thanks for your excellent project sharing to the community.

With a little pity, I found some faults.
For example, dist_train.sh in the following command is not found.
bash tools/dist_train.sh configs/classification/cifar10/swintiny_b64_5e_jpg.py $GPUS --fp16

Would you check and commit it anyway?

Thanks again.

Died with <Signals.SIGKILL: 9>. When first epoch ends, the program is killed

Using train on PAI and smart cache, config is like metric_learning/imagenet_resnet50_1000kid_jpg.py, then the program is always killed when fist epoch ends.
The error info is like below:
Traceback (most recent call last):
File "/home/pai/lib/python3.6/runpy.py", line 193, in _run_module_as_main
"main", mod_spec)
File "/home/pai/lib/python3.6/runpy.py", line 85, in _run_code
exec(code, run_globals)
File "/home/pai/lib/python3.6/site-packages/easypai/torch/launch.py", line 428, in
main()
File "/home/pai/lib/python3.6/site-packages/easypai/torch/launch.py", line 414, in main
sigkill_handler(signal.SIGTERM, None) # not coming back
File "/home/pai/lib/python3.6/site-packages/easypai/torch/launch.py", line 389, in sigkill_handler
raise subprocess.CalledProcessError(returncode=last_return_code, cmd=cmd)
subprocess.CalledProcessError: Command '['/apsara/TempRoot/xxxxx/workspace/python_bin', '-u', 'tools/train.py', '--local_rank=7', 'configs/metric_learning/xxxxx.py', '--work_dir', 'oss://xxxxx/', '--load_from', '/data/oss_bucket_0/xxxxx/r50_imagenet_epoch_100.pth', '--launcher', 'pytorch', '--fp16']' died with <Signals.SIGKILL: 9>.

YOLOX mAP baseline is lower than Megvii/YOLOX and mmdet/YOLOX？

could cpu machines train the models?

Bug: KeyError: 'YOLOXLrUpdaterHook is not in the hook registry'

exported classification model run failed during inference using easycv.predictos.Classifier

Traceback is as follows:

---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
<ipython-input-16-7181aabb565a> in <module>
      8 # input image should be RGB order
      9 img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
---> 10 output = tcls.predict([img])
     11 print(output)

/home/pai/lib/python3.6/site-packages/easycv/predictors/classifier.py in predict(self, input_data_list, batch_size)
     99             batch_image_list = image_list[batch_idx * batch_size:min(
    100                 (batch_idx + 1) * batch_size, len(image_list))]
--> 101             image_tensor_list = self.predictor.preprocess(batch_image_list)
    102             input_data = self.batch(image_tensor_list)
    103             output_prob = self.predictor.predict_batch(

/home/pai/lib/python3.6/site-packages/easycv/predictors/base.py in preprocess(self, image_list)
     74         for img in image_list:
     75             tmp_input = {'img': img}
---> 76             tmp_results = self.pipeline(tmp_input)
     77             output_imgs_list.append(tmp_results['img'])
     78 

/home/pai/lib/python3.6/site-packages/torchvision/transforms/transforms.py in __call__(self, img)
     58     def __call__(self, img):
     59         for t in self.transforms:
---> 60             img = t(img)
     61         return img
     62 

/home/pai/lib/python3.6/site-packages/easycv/datasets/shared/pipelines/format.py in __call__(self, results)
    138         data['img_metas'] = DC(img_meta, cpu_only=True)
    139         for key in self.keys:
--> 140             data[key] = results[key]
    141         return data
    142 

KeyError: 'gt_labels'

the main reason is the Collection operation configured in test_pipeline, and gt_labels is missing during inference.

test_pipeline = [
    dict(type='Resize', size=256),
    dict(type='CenterCrop', size=224),
    dict(type='ToTensor'),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='Collect', keys=['img', 'gt_labels'])
]

[Doc] config file link error

The link of config files in tutorials failed to track the real file in github

error link:

https://github.com/alibaba/EasyCV/blob/master/docs/configs/selfsup/mocov2/mocov2_rn50_8xb32_200e_jpg.py

expected:

https://github.com/alibaba/EasyCV/blob/master/configs/selfsup/mocov2/mocov2_rn50_8xb32_200e_jpg.py

[bugs] coco format data is conflict with source concat

use mulit subset with glob classification id coco annotations will meet the classification id error at
# line77
self.cat2label = {cat_id: i for i, cat_id in enumerate(self.cat_ids)}

every subset datasource will generate cls id from 0~x

nvidia_dali install error

pip install http://pai-vision-data-hz.cn-hangzhou.oss-cdn.aliyun-inc.com/third_party/nvidia_dali_cuda100-0.25.0-1535750-py3-none-manylinux2014_x86_64.whl
Looking in indexes: https://mirrors.aliyun.com/pypi/simple/
Collecting nvidia-dali-cuda100==0.25.0
WARNING: Retrying (Retry(total=4, connect=None, read=None, redirect=None, status=None)) after connection broken by 'ConnectTimeoutError(<pip._vendor.urllib3.connection.HTTPConnection object at 0x7f5bee745978>, 'Connection to pai-vision-data-hz.cn-hangzhou.oss-cdn.aliyun-inc.com timed out. (connect timeout=15)')': /third_party/nvidia_dali_cuda100-0.25.0-1535750-py3-none-manylinux2014_x86_64.whl
WARNING: Retrying (Retry(total=3, connect=None, read=None

[Framework] Compatible with mmdet models

DAB-DETR mAP 42.52

Thanks for your great work! I wonder if the mAP 42.52 of dab-detr is trained via the exact config file "configs/detection/dab_detr/dab_detr_r50_8x2_50e_coco.py"?
If so, I see that you use dropout = 0.0 in DABDetrTransformer instead of the default dropout value 0.1. Is there a performance gain by setting dropout value to 0.0?

_pickle.PicklingError: Can't pickle <class 'easycv.datasets.shared.pipelines.third_transforms_wrapper.RandomResizedCrop'>: attribute lookup RandomResizedCrop on easycv.datasets.shared.pipelines.third_transforms_wrapper failed

I have an conda env with pip install -f requirements successfully and train moby success
then by replace configs/selfsup/dino/dino_deit_small_p16_8xb32_100e_jpg.py datalist and dataroot, training on a 8-V100 ,get failed with

nccl timeout error when read tfrecord format data from oss

The error log like this:
Current pipeline object is no longer valid.
[E ProcessGroupNCCL.cpp:294] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. To avoid this inconsistency, we are taking the entire process down.
terminate called after throwing an instance of 'std::runtime_error'
what(): [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(OpType=ALLREDUCE, Timeout(ms)=1800000) ran for 1808702 milliseconds before timing out.
Traceback (most recent call last):
File "/home/pai/lib/python3.6/runpy.py", line 193, in _run_module_as_main
"main", mod_spec)
File "/home/pai/lib/python3.6/runpy.py", line 85, in _run_code
exec(code, run_globals)
File "/home/pai/lib/python3.6/site-packages/easypai/torch/launch.py", line 428, in
main()
File "/home/pai/lib/python3.6/site-packages/easypai/torch/launch.py", line 414, in main
sigkill_handler(signal.SIGTERM, None) # not coming back
File "/home/pai/lib/python3.6/site-packages/easypai/torch/launch.py", line 389, in sigkill_handler
raise subprocess.CalledProcessError(returncode=last_return_code, cmd=cmd)
subprocess.CalledProcessError: Command '['/apsara/TempRoot/Odps/xxxxx/[email protected]#0/workspace/python_bin', '-u', 'tools/train.py', '--local_rank=1', 'configs/metric_learning/resnet50_jpg_nopk_tfrecord.py', '--work_dir', 'oss://xxxx/mtl_tf/', '--load_from', 'oss://xxxx/r50_imagenet_epoch_100.pth', '--launcher', 'pytorch', '--fp16']' died with <Signals.SIGABRT: 6>.

Then set os.environ["NCCL_DEBUG_SUBSYS"] = "ALL"
os.environ["NCCL_DEBUG"] = "INFO"
and rerun, the error log like this:

[Rank 0] Watchdog caught collective operation timeout: WorkNCCL(OpType=ALLREDUCE, Timeout(ms)=1800000) ran for 1808702 milliseconds before timing out.
Traceback (most recent call last):
File "tools/train.py", line 293, in
main()
File "tools/train.py", line 274, in main
data_loaders = [dataset.get_dataloader()]
File "/apsara/TempRoot/Odps/xxxx/[email protected]#0/workspace/easycv/datasets/shared/dali_tfrecord_imagenet.py", line 166, in get_dataloader
self.dali_pipe.build()
File "/home/pai/lib/python3.6/site-packages/nvidia/dali/pipeline.py", line 478, in build
self._pipe.Build(self._names_and_devices)
RuntimeError: Critical error when building pipeline:
Error when constructing operator: TFRecordReader encountered:
[/opt/dali/dali/operators/reader/loader/indexed_file_loader.h:105] Assert on "index_uris.size() == uris.size()" failed: Number of index files needs to match the number of data files

[feature] expand cls model zoo

TypeError: evaluate() got an unexpected keyword argument 'gpu_collect'

Traceback (most recent call last):
File "tools/train.py", line 274, in
main()
File "tools/train.py", line 270, in main
use_fp16=args.fp16)
File "/home/sanjie.lp/projects/EasyCV/easycv/apis/train.py", line 278, in train_model
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
File "/home/sanjie.lp/softwares/anaconda2/envs/easycv/lib/python3.6/site-packages/mmcv/runner/epoch_based_runner.py", line 108, in run
self.call_hook('before_run')
File "/home/sanjie.lp/softwares/anaconda2/envs/easycv/lib/python3.6/site-packages/mmcv/runner/base_runner.py", line 309, in call_hook
getattr(hook, fn_name)(self)
File "/home/sanjie.lp/projects/EasyCV/easycv/hooks/eval_hook.py", line 55, in before_run
self.after_train_epoch(runner)
File "/home/sanjie.lp/projects/EasyCV/easycv/hooks/eval_hook.py", line 72, in after_train_epoch
self.evaluate(runner, results)
File "/home/sanjie.lp/projects/EasyCV/easycv/hooks/eval_hook.py", line 93, in evaluate
results, logger=runner.logger, **self.eval_kwargs)
TypeError: evaluate() got an unexpected keyword argument 'gpu_collect'

[doc] metric learning tutorial is missing

for metric learning task, only config files is provided, but detailed tutorials should be more helpful

nvidia-dali is too old, should be upgraded

currently install nvidia-dali by

! pip install http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/third_party/nvidia_dali_cuda100-0.25.0-1535750-py3-none-manylinux2014_x86_64.whl

which can not be automatically add to requirements of whl package

[feat] ConvMAE deserves to be implemented

paper: https://arxiv.org/abs/2205.03892

TypeError: create_model() got multiple values for argument 'pretrained'

I encountered this problem while running the first example.

python tools/train.py configs/classification/cifar10/swintiny_b64_5e_jpg.py --work_dir work_dirs/classification/cifar10/swintiny --fp16

Traceback (most recent call last):
File "tools/train.py", line 277, in
main()
File "tools/train.py", line 213, in main
model = build_model(cfg.model)
File "/home/zhenzhen/ev/tmp/EasyCV/package/easycv/models/builder.py", line 39, in build_model
return build(cfg, MODELS)
File "/home/zhenzhen/ev/tmp/EasyCV/package/easycv/models/builder.py", line 15, in build
return build_from_cfg(cfg, registry, default_args)
File "/home/zhenzhen/ev/tmp/EasyCV/package/easycv/utils/registry.py", line 80, in build_from_cfg
return obj_cls(**args)
File "/home/zhenzhen/ev/tmp/EasyCV/package/easycv/models/classification/classification.py", line 77, in init
self.backbone = builder.build_backbone(backbone)
File "/home/zhenzhen/ev/tmp/EasyCV/package/easycv/models/builder.py", line 19, in build_backbone
return build(cfg, BACKBONES)
File "/home/zhenzhen/ev/tmp/EasyCV/package/easycv/models/builder.py", line 15, in build
return build_from_cfg(cfg, registry, default_args)
File "/home/zhenzhen/ev/tmp/EasyCV/package/easycv/utils/registry.py", line 80, in build_from_cfg
return obj_cls(**args)
File "/home/zhenzhen/ev/tmp/EasyCV/package/easycv/models/backbones/pytorch_image_models_wrapper.py", line 100, in init
exportable, no_jit, **kwargs)
TypeError: create_model() got multiple values for argument 'pretrained'

Hook is already registered in hook while import mmdet.core

swint_cifar10 config bug

feature extract error

Could easycv suuport batch inference?

Thanks for your brilliant work!

I've looked at the inference demo of object detection but found batch inference is not supported yet.

Could you help support batch inference of detection models (ex. yolox)??

Appreciate!

Bug: run command --seed 42 --deterministic do not work

--seed 42 --deterministic
Results of two runs：

1. Epoch [1][50/7330] 59.7561 Epoch [1][100/7330] 48.4067 Epoch [1][150/7330] 44.8409
2. Epoch [1][50/7330] 58.8591 Epoch [1][100/7330] 48.8086 Epoch [1][150/7330] 45.2519

ModuleNotFoundError: No module named 'sailfish'

When I try to run the example of metric learning, I encountered the following error reports：
ModuleNotFoundError: No module named 'sailfish'

When I run the following command, error appears.
CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --nproc_per_node=1 --master_port=29500 tools/train.py configs/metric_learning/cub_resnet50_jpg.py --work_dir work_dirs/metric_learning/cub/r50 --launcher pytorch --fp16

The URL of this example: https://github.com/alibaba/EasyCV/blob/master/docs/source/tutorials/EasyCV%E5%BA%A6%E9%87%8F%E5%AD%A6%E4%B9%A0resnet50.ipynb

[doc] oss io tutorial does not introduce config ak via environment variable

https://github.com/alibaba/EasyCV/blob/master/docs/source/tutorials/file.md

add env config introduction including

1. how to config ak_id ak_secret host
2. how to config oss config file path

[Framework] Visualization enhancement

We should support visualization of images during training and evaluation, these images could be visualization of evaluation results such as bounding box, attention maps, and so on

安装pai-easycv时出现错误

linux系统下安装pai-easycv报错如下：
note: This error originates from a subprocess, and is likely not a problem with pip.
ERROR: Failed building wheel for opencv-python-headless
Failed to build opencv-python-headless
ERROR: Could not build wheels for opencv-python-headless, which is required to install pyproject.toml-based projects
感谢解答

ci test is needed

we should add ci test pipeline based on github actions

vitdet outofmomery

请问vitdet 默认用的是什么gpu， 2080会提示out of memory

训练vitdet出错

我按照 https://zhuanlan.zhihu.com/p/528733299 里的教程一步步配置，不同的是我是单机双卡，所以我的命令是CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port=29500 tools/train.py configs/detection/vitdet/vitdet_100e.py --work_dir easycv/vitdet --launcher pytorch --fp16，训练报错
Traceback (most recent call last):
File "/home/xlh2/EasyCV2/tools/train.py", line 277, in
main()
File "/home/xlh2/EasyCV2/tools/train.py", line 266, in main
train_model(
File "/home/xlh2/EasyCV2/easycv/apis/train.py", line 269, in train_model
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
File "/home/xlh2/.conda/envs/easycv/lib/python3.9/site-packages/mmcv/runner/epoch_based_runner.py", line 127, in run
epoch_runner(data_loaders[i], **kwargs)
File "/home/xlh2/EasyCV2/easycv/runner/ev_runner.py", line 105, in train
self.run_iter(data_batch, train_mode=True)
File "/home/xlh2/EasyCV2/easycv/runner/ev_runner.py", line 72, in run_iter
outputs = self.model.train_step(data_batch, self.optimizer,
File "/home/xlh2/.conda/envs/easycv/lib/python3.9/site-packages/mmcv/parallel/distributed.py", line 48, in train_step
self._sync_params()
File "/home/xlh2/.conda/envs/easycv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1185, in getattr
raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'MMDistributedDataParallel' object has no attribute '_sync_params'

Yolox预测过程出错

我在使用Yolox教程里的预测过程，不过我没有自己训练，而是打算先用你们提供的模型参数，export之后再测试一下，预测过程报错：There are syntax errors in config file /home/xlh2/EasyCV2/.easycv_cache/epoch_300m_e_config.py: unmatched '}' (, line 1)。经过调试，我发现错误出现在 easycv/predictors/detector.py 第90行：config_str = config_str[config_str.find('base'):]。在export过程中没有往pth文件里写入_base_,从而导致返回0索引值，进而返回无意义的”｛“。

[feature] fcos implementation

FCOS (https://arxiv.org/abs/1904.01355) is a practically used one-stage detection algorithm during industry, besides, it is a fundamental model on which many exicting work have been done, we need this model.

attribute lookup RandomResizedCrop on easycv.datasets.shared.pipelines.third_transforms_wrapper failed

_pickle.PicklingError while train by launcher pytorch

difference between export model and original model ?

why should we have to export model before predicting , rather than use it directly ?
为什么我们需要export 模型，然后再预测呢，为什么不直接预测呢 ？

yolox的例子中https://github.com/alibaba/EasyCV/blob/master/docs/source/tutorials/EasyCV%E5%9B%BE%E5%83%8F%E6%A3%80%E6%B5%8BYoloX.ipynb
在导出模型之前，对nms阈值进行了修改， 如果用vitdet，请问需要怎么调整呢？需要吗，谢谢

No "backend" parameters in tool/quantize.py

Hi，I just read the "YOLOX Compression Tutorial", in there the quantize functional is called by

python tools/quantize.py \
		${CONFIG_PATH} \
		${MODEL_PATH} \
		--work_dir ${WORK_DIR} \
		--device ${DEVICE} \
		--backend ${BACKEND}

But I find no 'backend' parameters in tool/quantize.py, weather it is missed?
And moreover, I have found that pruning functional use quantize.py as well. Is that wrong writen?

it seems tools/test.py duplicated with tools/eval.py

So many grammar mistakes in README

Listed here and noted by brackets "()".

Functionality & Extensibility

In addition to SSL, EasyCV also support(s) image classification, object detection, metric learning, and more area(s) will be supported in the future. Although convering different area(s), EasyCV decompose(s) the framework into different componets such as dataset, model (and) running hook, making it easy to add new compoenets and combining it with existing modules.

EasyCV provide(s) simple and comprehensive interface for inference. Additionaly, all models are supported on PAI-EAS, which can be easily deployed as online service and support automatic scaling and service monitoring.

Efficiency

EasyCV support(s) multi-gpu and multi worker training. EasyCV use(s) DALI to accelerate data io and preprocessing process, and use(s) TorchAccelerator and fp16 to accelerate training process. For inference optimization, EasyCV export(s) model using jit script, which can be optimized by PAI-Blade

`configs/metric_learning/sop_timm_swinb_local.py` raise TypeError: __init__() got an unexpected keyword argument 'm_per_class'