search

Yangr116 / BoxSnake

[ICCV 2023] BoxSnake official repository.
https://arxiv.org/pdf/2303.11630.pdf
Apache License 2.0
66 stars 6 forks source link

Installation issue #7

Closed nabeelkhalid92 closed 5 months ago

nabeelkhalid92 commented 7 months ago

Hi,

First of all thank you very much for providing such a nice platform.

I am running into some trouble with the installations, getting errors like:

  1. ERROR: Could not find a version that satisfies the requirement rasterizer==0.0.0 (from -r requirements.txt (line 60)) (from versions: none) ERROR: No matching distribution found for rasterizer==0.0.0 (from -r requirements.txt (line 60))
  2. ERROR: Could not find a version that satisfies the requirement networkx==2.8.6

  3. ERROR: Could not find a version that satisfies the requirement MultiScaleDeformableAttention==1.0 (from -r requirements.txt (line 36)) (from versions: none) ERROR: No matching distribution found for MultiScaleDeformableAttention==1.0 (from -r requirements.txt (line 36))

    1. ERROR: Could not find a version that satisfies the requirement bilateralfilter==0.1 (from versions: none) ERROR: No matching distribution found for bilateralfilter==0.1

    And some more errors like these. Can you please let me know why is that? Is it related to the python version? I already have tried Python 3.9 and 3.10.

I will wait for your response. Thank you and kind regards,

nabeelkhalid92 commented 7 months ago

Follow up! The installation worked, it had to do with the python version.

I have another question. I am training BoxSnake using my own custom dataset which contains microscopic data. The training runs smoothly at the start with AP for segmentation raising alongside the AP for detection but after a while, the AP for segmentation starts to decrease as the detection AP rises.

Are there any parameters in the config file related to POLYGON_HEAD or BOX_SUP I have to change other than the Num_classes etc?

I will be awaiting your response.

Thank you again :)

Yangr116 commented 7 months ago

Hi, can you tell me the size of your microscopic dataset? And can you visualize some cases?

nabeelkhalid92 commented 7 months ago

Thank you very much for your reply,

The dataset I am using is the LIVECell dataset: https://github.com/sartorius-research/LIVECell

Below you can find the AP decline case I am talking about:

After 1000 iterations: [04/02 19:12:55] d2.engine.defaults INFO: Evaluation results for testall in csv format: [04/02 19:12:55] d2.evaluation.testing INFO: copypaste: Task: bbox [04/02 19:12:55] d2.evaluation.testing INFO: copypaste: AP,AP50,AP75,APs,APm,APl [04/02 19:12:55] d2.evaluation.testing INFO: copypaste: 31.5135,65.8799,27.6709,32.3391,32.9443,32.1784 [04/02 19:12:55] d2.evaluation.testing INFO: copypaste: Task: segm [04/02 19:12:55] d2.evaluation.testing INFO: copypaste: AP,AP50,AP75,APs,APm,APl [04/02 19:12:55] d2.evaluation.testing INFO: copypaste: 27.3450,59.6818,22.6698,26.1375,27.3899,34.2320

After 1500 Iterations: [04/02 19:28:55] d2.engine.defaults INFO: Evaluation results for testall in csv format: [04/02 19:28:55] d2.evaluation.testing INFO: copypaste: Task: bbox [04/02 19:28:55] d2.evaluation.testing INFO: copypaste: AP,AP50,AP75,APs,APm,APl [04/02 19:28:55] d2.evaluation.testing INFO: copypaste: 35.1612,70.8197,32.3604,37.6642,32.4191,32.3361 [04/02 19:28:55] d2.evaluation.testing INFO: copypaste: Task: segm [04/02 19:28:55] d2.evaluation.testing INFO: copypaste: AP,AP50,AP75,APs,APm,APl [04/02 19:28:55] d2.evaluation.testing INFO: copypaste: 25.8201,64.4165,14.7160,24.7313,25.6564,33.5879

After 2000 iterations: [04/02 19:42:45] d2.engine.defaults INFO: Evaluation results for testall in csv format: [04/02 19:42:45] d2.evaluation.testing INFO: copypaste: Task: bbox [04/02 19:42:45] d2.evaluation.testing INFO: copypaste: AP,AP50,AP75,APs,APm,APl [04/02 19:42:45] d2.evaluation.testing INFO: copypaste: 37.6725,73.6499,35.7813,39.2702,38.2316,39.5760 [04/02 19:42:45] d2.evaluation.testing INFO: copypaste: Task: segm [04/02 19:42:45] d2.evaluation.testing INFO: copypaste: AP,AP50,AP75,APs,APm,APl [04/02 19:42:45] d2.evaluation.testing INFO: copypaste: 21.3821,61.3990,7.6428,18.1947,23.4008,35.2326

After 2500 Iterations: [04/02 19:57:39] d2.engine.defaults INFO: Evaluation results for testall in csv format: [04/02 19:57:39] d2.evaluation.testing INFO: copypaste: Task: bbox [04/02 19:57:39] d2.evaluation.testing INFO: copypaste: AP,AP50,AP75,APs,APm,APl [04/02 19:57:39] d2.evaluation.testing INFO: copypaste: 37.3242,74.8505,34.4125,39.5199,36.0486,39.0514 [04/02 19:57:39] d2.evaluation.testing INFO: copypaste: Task: segm [04/02 19:57:39] d2.evaluation.testing INFO: copypaste: AP,AP50,AP75,APs,APm,APl [04/02 19:57:39] d2.evaluation.testing INFO: copypaste: 16.4391,56.2552,3.4211,12.2411,20.9670,33.4366

The Segmentation AP falls right down after this.

Below is the config file I am using, I made changes related to the LIVECell dataset:

BASE: "../Base-BoxSnake-RCNN-FPN.yaml" OUTPUT_DIR: "/raid/nabeelk/nabeelk/nabeelk/output_boxsnake/" MODEL: WEIGHTS: "/home/nabeelk/BoxSnake-master/configs/COCO-InstanceSegmentation/BoxSnake_RCNN/boxsnake_rcnn_R_50_FPN_coco_1x.pth" MASK_ON: True ROI_MASK_HEAD: NAME: "PolygonHead" POOLER_TYPE: "" POLYGON_HEAD: IN_FEATURES: ["p2", "p3", "p4", "p5"] PRED_WITHIN_BOX: False POLY_NUM_PTS: 64 CLS_AGNOSTIC_MASK: True PREPOOL: False UPSAMPLING: False FPN: NORM: "SyncBN" ANCHOR_GENERATOR: SIZES: [[8], [16], [32], [64], [128]] # One size for each in feature map ASPECT_RATIOS: [[0.5, 1.0, 2.0, 3.0, 4.0]] # Three aspect ratios (same for all in feature maps)

DIFFRAS: RESOLUTIONS: (64, 64, 64, 64, 64, 64, 64, 64) USE_RASTERIZED_GT: False INV_SMOOTHNESS_SCHED: (0.1,) RESNETS: DEPTH: 50 ROI_HEADS: NUM_CLASSES: 1 BATCH_SIZE_PER_IMAGE: 512 PROPOSAL_ONLY_GT: False BOX_SUP: ENABLE: True LOSS_POINTS_PROJ: True LOSS_POINTS_PROJ_WEIGHT: 1.0 LOSS_LOCAL_PAIRWISE: True LOSS_PAIRWISE_WARMUP_ITER: 10000 LOCAL_PAIRWISE_KERNEL_SIZE: 3 LOCAL_PAIRWISE_DILATION: 2 LOSS_LOCAL_PAIRWISE_WEIGHT: 0.5 LOSS_GLOBAL_PAIRWISE: True LOSS_GLOBAL_PAIRWISE_WEIGHT: 0.03 CROP_PREDICTS: True CROP_SIZE: 64 MASK_PADDING_SIZE: 4

RPN: IN_FEATURES: ["p2", "p3", "p4", "p5", "p6"] BATCH_SIZE_PER_IMAGE: 256 POST_NMS_TOPK_TEST: 3000 POST_NMS_TOPK_TRAIN: 3000 PRE_NMS_TOPK_TEST: 6000 PRE_NMS_TOPK_TRAIN: 12000 RETINANET: NUM_CLASSES: 1 TOPK_CANDIDATES_TEST: 3000 PIXEL_MEAN: [128, 128, 128] PIXEL_STD: [11.578, 11.578, 11.578]

SOLVER: OPTIMIZER: "ADAM" BASE_LR: 1e-4 WEIGHT_DECAY: 0.1 WEIGHT_DECAY_NORM: 0.0 STEPS: (17000, 18000) MAX_ITER: 20000 IMS_PER_BATCH: 16 CHECKPOINT_PERIOD: 500 CLIP_GRADIENTS: ENABLED: False DATASETS: TRAIN: ("trainall",) TEST: ("testall",) INPUT: MIN_SIZE_TRAIN: (440, 480, 520, 560, 580, 620)
VIS_PERIOD: 0 TEST: DETECTIONS_PER_IMAGE: 3000 EVAL_PERIOD: 500 DATALOADER: NUM_WORKERS: 12 [04/02 18:41:15] detectron2 INFO: Running with full config: CFG_FILE_STR: BoxSnake_RCNN/allconfig.yaml COMMENT: NONE CUDNN_BENCHMARK: false DATALOADER: ASPECT_RATIO_GROUPING: true FILTER_EMPTY_ANNOTATIONS: true NUM_WORKERS: 12 REPEAT_THRESHOLD: 0.0 SAMPLER_TRAIN: TrainingSampler DATASETS: PRECOMPUTED_PROPOSAL_TOPK_TEST: 1000 PRECOMPUTED_PROPOSAL_TOPK_TRAIN: 2000 PROPOSAL_FILES_TEST: [] PROPOSAL_FILES_TRAIN: [] TEST:

Please let me know what else can be done. Thank you again :)

Yangr116 commented 7 months ago

You can refer to this config to revise your config in terms of ANCHOR_GENERATOR:

  ANCHOR_GENERATOR:
    SIZES: [[4], [9], [17], [31], [64], [127]]  # One size for each in feature map
    ASPECT_RATIOS: [[0.25, 0.5, 1.0, 2.0, 4.0]]  # Three aspect ratios (same for all in feature maps)

This will fit with the object size of your dataset.

In addition, the default parameters are set for a common environment. For the cell segmentation, the RGB color may influence the model's performance. Especially in the weakly supervised setting, the model is prone to overfitting. So you can adjust PAIRWISE.SIGMA of this line to find the best sigma for the pairwise loss for your dataset.

nabeelkhalid92 commented 7 months ago

I have tried changing the local.pairwise sigma, I have used 0.5, 1.0, 1.5, 2.4, 2.5, 2.6, 3.0, 3.5 and it doesn't make much of a difference unfortunately.

Are there any other parameters which I can play around with? Thank you

Yangr116 commented 7 months ago

Can you visualize some samples?

python demo/demo.py  \
   --config-file configs/COCO-InstanceSegmentation/BoxSnake_RCNN/boxsnake_rcnn_R_50_FPN_1x.yaml \
   --input demo/demo.jpg \
   --output ${/your/visualized/dir} \
   --confidence-threshold 0.5 \
   --opts MODEL.WEIGHTS ${your/checkpoints/boxsnake_rcnn_R_50_FPN_coco_1x.pth}
nabeelkhalid92 commented 7 months ago

skbr3_1 skbr3_3 Here are some results from the model before it starts to overfit.

nabeelkhalid92 commented 7 months ago

And this a result after the model does overfit huh7_1

Yangr116 commented 7 months ago

You can see that the colors of the objects are similar to the background, so you need to reduce sigma of local pairwise since our method mainly relies on the RGB color feature. Can you try sigma=0.1 or sigma=0.01?

In addition, I suggest that you can use some image augmentation methods to improve the contrast between objects and backgrounds.

Julppe commented 5 months ago

Hello, I'm running into the same issues during the installation. Could you specify which Python version worked for you?

Thank you!

Yangr116 commented 5 months ago

We rely on python=3.8. You can prepare the env by:

conda create --name boxsnake python=3.8 -y
conda activate boxsnake

conda install pytorch==1.9.0 torchvision==0.10.0 cudatoolkit=11.1 -c pytorch -c nvidia
python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'

git clone https://github.com/Yangr116/BoxSnake.git
cd BoxSnake
pip install -r requirements.txt
bash scripts/auto_build.sh