What is class_quality ?

[kdh-awraw1019]

I dont know what is class_quality

code : mmdet\ppal\sampler\difficulty_calibrated_uncertainty_sampler.py line : 41 class_qualities = ckpt['state_dict']['bbox_head.class_quality'].numpy() in def _get_classwise_weight(self, results_json):

I did load checkpoint file retinanet_r50_caffe_fpn_mstrain_1x_coco-586977a0.pth from https://mmdetection.readthedocs.io/en/stable/model_zoo.html

but, the checkpoint file has not key of dict 'bbox_head.class_quality'.

so I cannot run this command python tools/run_al_coco.py --config ./al_configs/coco/ppal_retinanet_coco.py --model retinanet

And, when I run this command in anaconda prompt ternimal this error occured..

obviously, I did installation with your guide

python tools/run_al_coco.py --config al_configs/coco/ppal_retinanet_coco.py --model retinanet Traceback (most recent call last): File "tools/run_al_coco.py", line 5, in from mmdet.ppal.sampler import * ModuleNotFoundError: No module named 'mmdet.ppal'

My environments

RTX3060 anaconda3 pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 mmcv-full==1.5.0 mmdet==2.27.0

[ChenhongyiYang]

Hi,

Thank you for being interested in our work.

For the first question, the class_quality equals to (1 - class_difficulty), which is introduced in Section 3.3 of our paper. It is computed and updated using EMA during model training, so a pre-trained model provided by mmdetection will not include it.

For the second question, you need to run python setup.py install under the PPAL project instead of a newly cloned mmdetection.

Best, Chenhongyi

[kdh-awraw1019]

Thank you for your comment

So, is the correct and only way to install PPAL is to just run the 'python setup.py install' command?

[ChenhongyiYang]

Hi,

You can also run pip install -v -e .

[kdh-awraw1019]

Hi,

I still have question.

How can I find the code for the process of calculating the difficulty of all predicted boxes during learning and the process of updating to EMA to calculate image uncertainty in DCUS (Section3.3)?

[ChenhongyiYang]

Hi,

You can find the implementations here

[kdh-awraw1019]

OK, And...

I want to fine the code for the process of calculating the similarity of Images in CCMS

I found implementation about diversity sampling (k-greedy-centers and kmeans) https://github.com/ChenhongyiYang/PPAL/blob/main/mmdet/ppal/sampler/diversity_sampler.py

but, I could not find calculating similarity...

[ChenhongyiYang]

Hi, you can find it here

[kdh-awraw1019]

Hi, New questions.

1. In run_al_coco.py, 'torch.distributed.launch' is in train / eval / unlabeled command.

   • Can I run 'torch.distributed.launch' in windows 10?

2. KeyError: 'bbox_head.class_quality' (Runned in pycharm)

Traceback (most recent call last): File "D:\Workspace\PPAL\tools\run_al_coco.py", line 168, in run(i+1, i!=int(cfg.get('round_num'))-1) File "D:\Workspace\PPAL\tools\run_al_coco.py", line 129, in run uncertainty_sampler.al_round(round_uncertainty_inference_json, round_labeled_json, round_uncertainty_new_labeled_json, round_uncertainty_new_unlabeled_json) File "D:\Workspace\PPAL\mmdet\ppal\sampler\difficulty_calibrated_uncertainty_sampler.py", line 136, in al_round sampled_img_ids, rest_img_ids = self.al_acquisition(result_path, last_label_path) File "D:\Workspace\PPAL\mmdet\ppal\sampler\difficulty_calibrated_uncertainty_sampler.py", line 55, in al_acquisition class_weights = self._get_classwise_weight(result_json) File "D:\Workspace\PPAL\mmdet\ppal\sampler\difficulty_calibrated_uncertainty_sampler.py", line 41, in _get_classwise_weight class_qualities = ckpt['state_dict']['bbox_head.class_quality'].numpy() KeyError: 'bbox_head.class_quality'

config file (ppal_retinanet_coco.py)

Paths

config_dir = 'D:/Workspace/PPAL/configs/coco_active_learning/' work_dir = 'D:/Workspace/PPAL/work_dirs/'

Environment setting

python_path = 'C:/Users/aaaaa/Anaconda3/envs/ppal/' port = 29500 gpus = 1

Data setting

oracle_path = 'D:/Workspace/PPAL/data/coco/annotations/instances_train2017.json' init_label_json = 'D:/Workspace/PPAL/data/active_learning/coco/coco_2365_labeled_1.json' init_unlabeled_json = 'D:/Workspace/PPAL/data/active_learning/coco/coco_2365_unlabeled_1.json' init_model = None

Config setting

train_config = config_dir + 'al_train/retinanet_26e.py' uncertainty_infer_config = config_dir + 'al_inference/retinanet_uncertainty.py' diversity_infer_config = config_dir + 'al_inference/retinanet_diversity.py'

Active learning setting

round_num = 5 budget = 2365 budget_expand_ratio = 4 uncertainty_pool_size = budget budget_expand_ratio + gpus - (budget budget_expand_ratio) % gpus

Sampler setting

uncertainty_sampler_config = dict( type='DCUSSampler', n_sample_images=uncertainty_pool_size, oracle_annotation_path=oracle_path, score_thr=0.05, class_weight_ub=0.2, class_weight_alpha=0.3, dataset_type='coco') diversity_sampler_config = dict( type='DiversitySampler', n_sample_images=budget, oracle_annotation_path=oracle_path, dataset_type='coco')

output_dir = work_dir + 'retinanet_coco_ppal_5rounds_2percent_to_10percent'

and I wanna know about 'port'

//////////////////////////////////////////////////////

if I run ./tools/run_al_coco.py —config ./al_configs/coco/ppal_retinanet_coco.py —model retinanet command in Terminal,

its print this: (ppal) D:\Workspace\PPAL>tools\run_al_coco.py --config ./al_configs/coco/ppal_retinanet_coco.py --model retinanet

(ppal) D:\Workspace\PPAL> [main 2023-01-30T06:19:35.953Z] update#setState idle [main 2023-01-30T06:19:36.681Z] Starting extension host with pid 19264 (fork() took 12 ms). [main 2023-01-30T06:20:05.963Z] update#setState checking for updates [main 2023-01-30T06:20:07.381Z] update#setState idle [9676:0130/152135.996:ERROR:gpu_init.cc(481)] Passthrough is not supported, GL is disabled, ANGLE is

[sangeethnrs]

@kdh-awraw1019 i do have the same error and how to solve it? the checkpoint file has not key of dict 'bbox_head.class_quality'.