how to train centermask2 on own datasets about centermask2 HOT 13 OPEN

I successfully trained a model on my own dataset.

Firstly I create two own datasets in coco format:

• for test
  ./centermask2/datasets/numberplate/val/coco_numberplate.json
  ./centermask2/datasets/numberplate/val/*.jpeg
• for train
  ./centermask2/datasets/numberplate/train/coco_numberplate.json
  ./centermask2/datasets/numberplate/train/*.jpeg

Than I create own config files for model with 2 classes BACKGROUND and NUMBERPLATE in dir ./centermask2/configs/numberplates

• numberplate_V_39_eSE_FPN_ms_3x.yaml

_BASE_: "Base-NumberPlate-VoVNet.yaml"
MODEL:
  WEIGHTS: "https://www.dropbox.com/s/q98pypf96rhtd8y/vovnet39_ese_detectron2.pth?dl=1"
  VOVNET:
    CONV_BODY : "V-39-eSE"
SOLVER:
  STEPS: (0, 60000)
  MAX_ITER: 60000
OUTPUT_DIR: "output/numberplate/numberplate-V-39-ms-3x"

• Base-NumberPlate-VoVNet.yaml

_MODEL:
  META_ARCHITECTURE: "GeneralizedRCNN"
  BACKBONE:
    NAME: "build_fcos_vovnet_fpn_backbone"
    FREEZE_AT: 0
  VOVNET:
    OUT_FEATURES: ["stage3", "stage4", "stage5"]
  FPN:
    IN_FEATURES: ["stage3", "stage4", "stage5"]
  PROPOSAL_GENERATOR:
    NAME: "FCOS"  
  FCOS:
    POST_NMS_TOPK_TEST: 50
  # PIXEL_MEAN: [102.9801, 115.9465, 122.7717]
  MASK_ON: True
  MASKIOU_ON: True
  ROI_HEADS:
    NAME: "CenterROIHeads"
    IN_FEATURES: ["p3", "p4", "p5"]
  ROI_MASK_HEAD:
    NAME: "SpatialAttentionMaskHead"
    ASSIGN_CRITERION: "ratio"
    NUM_CONV: 4
    POOLER_RESOLUTION: 14
DATASETS:
  TRAIN: ("numberplate_train",)
  TEST: ("numberplate_val",)
SOLVER:
  CHECKPOINT_PERIOD: 6000
  IMS_PER_BATCH: 16
  BASE_LR: 0.01  # Note that RetinaNet uses a different default learning rate
  STEPS: (0, 60000)
  MAX_ITER: 60000
INPUT:
  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)

After that I added to the file ./centermask2/train_net.py in the section main the registration of my own dataset:

    # Regist own dataset.
    from detectron2.data.datasets import register_coco_instances
    # train data
    name        = "numberplate_train"
    json_file   = "/var/www/centermask2/datasets/numberplate/train/coco_numberplate.json"
    image_root  = "/var/www/centermask2/datasets/numberplate/train"
    # test data
    name_val        = "numberplate_val"
    json_file_val   = "./centermask2/datasets/numberplate/val/coco_numberplate.json"
    image_root_val  = "./centermask2/datasets/numberplate/val"
    # registr
    register_coco_instances(name, {}, json_file, image_root)
    register_coco_instances(name_val, {}, json_file_val, image_root_val)

In the end, I ran the command line:

DETECTRON2_DATASETS=./centermask2/datasets python3 train_net.py --config-file "configs/numberplates/numberplate_V_39_eSE_FPN_ms_3x.yaml" --num-gpus 1

from centermask2.

for prediction on own dataset I use simple script

import os
import time
import glob

import numpy as np
from matplotlib import pyplot as plt

from detectron2.engine import DefaultPredictor, default_argument_parser, default_setup, launch
from centermask.config import get_cfg

def setup(args):
    """
    Create configs and perform basic setups.
    """
    cfg = get_cfg()
    cfg.merge_from_file(args.config_file)
    cfg.merge_from_list(args.opts)
    cfg.freeze()
    default_setup(cfg, args)
    return cfg


def main(args):
    cfg = setup(args)
    
    outputs_cpu = []
    predictor = DefaultPredictor(cfg)
    for file_name in glob.glob("./images/*"):
        im = plt.imread(file_name)        
        
        start_time = time.time()
        outputs = predictor(im)
        output_cpu = outputs["instances"].to("cpu")
        print(f"[TIME] {time.time() - start_time}s")
        for i, mask in enumerate(np.array(output_cpu.get_fields()["pred_masks"])):
            plt.imsave(f"./masks/{os.path.basename(file_name)}_mask_{i}.png", mask*255)
        outputs_cpu.append(output_cpu)
    return output_cpu


if __name__ == "__main__":
    args = default_argument_parser().parse_args()
    print("Command Line Args:", args)
    
    # Regist own dataset.
    from detectron2.data.datasets import register_coco_instances
    # train data
    name        = "numberplate_train"
    json_file   = "/var/www/centermask2/datasets/numberplate/train/coco_numberplate.json"
    image_root  = "/var/www/centermask2/datasets/numberplate/train"
    # test data
    name_val        = "numberplate_val"
    json_file_val   = "/var/www/centermask2/datasets/numberplate/val/coco_numberplate.json"
    image_root_val  = "/var/www/centermask2/datasets/numberplate/val"
    # registr
    register_coco_instances(name, {}, json_file, image_root)
    register_coco_instances(name_val, {}, json_file_val, image_root_val)
    
    launch(
        main,
        args.num_gpus,
        num_machines=args.num_machines,
        machine_rank=args.machine_rank,
        dist_url=args.dist_url,
        args=(args,),
    )

run script

python3 predict_net.py --config-file "./configs/numberplates/numberplate_V_39_eSE_FPN_ms_3x.yaml"  MODEL.WEIGHTS .//output/numberplate/numberplate-V-39-ms-3x/model_0023999.pth

In this case, only 1 GPU is used.
If you want to run on more GPUs, use AsyncPredictor instead of DefaultPredictor

from centermask2.

need to change the ./centermask/config/defaults.py file ？
like :
cfg.MODEL.ROI_HEADS.NUM_CLASSES = num_classes
cfg.MODEL.SEM_SEG_HEAD.NUM_CLASSES = num_classes
cfg.MODEL.RETINANET.NUM_CLASSES = num_classes
cfg.MODEL.FCOS.NUM_CLASSES = num_classes

??

from centermask2.

It seems that if own datasets only has one class, the NUM_CLASSES should be 2 which contains background.
When training with my own datasets, if this value set as 1 , the loss will be very small and quickly decreases to 0. If this value set as 2, the loss seems normal.

from centermask2.

@Kenneth-X

Yes, except for cfg.MODEL.RETINANET.NUM_CLASSES = num_classes

from centermask2.

@Kenneth-X

Yes, except for cfg.MODEL.RETINANET.NUM_CLASSES = num_classes

thanks for that ,i can only find _C.MODEL.FCOS.NUM_CLASSES in defaults.py and i change it from 80 to 1
however, i can not find the others:
cfg.MODEL.ROI_HEADS.NUM_CLASSES
cfg.MODEL.SEM_SEG_HEAD.NUM_CLASSES
which file should i find ?

from centermask2.

@Kenneth-X

you just insert the config option in config file (.yaml) or command line.

For example,

python train_net.py --config-file <config.yaml> --num-gpus 8 MODEL.ROI_HEADS.NUM_CLASSES 1

from centermask2.

/numberplate/train/coco_numberplate.json"
    image_root  = "/var/www/centermask2/datasets/numberplate/train"

Is it required to include the .npy files of the masks in the train folder?

I am confused here; when I only include the .jpeg files of the input images and the .json file of the annotations, I get an error like this:

No such file or directory: '/datasets/new/train/image_000000.npy

and when I include the .npy files, the error is:

cannot identify image file <_io.BufferedReader name='/datasets/new/train/image_000000.npy'

from centermask2.

Please how can i predict with the model that i train?
plaise help me

from centermask2.

I successfully trained a model on my own dataset.

Firstly I create two own datasets in coco format:

• for test
  ./centermask2/datasets/numberplate/val/coco_numberplate.json
  ./centermask2/datasets/numberplate/val/*.jpeg
• for train
  ./centermask2/datasets/numberplate/train/coco_numberplate.json
  ./centermask2/datasets/numberplate/train/*.jpeg

Than I create own config files for model with 2 classes BACKGROUND and NUMBERPLATE in dir ./centermask2/configs/numberplates

• numberplate_V_39_eSE_FPN_ms_3x.yaml

_BASE_: "Base-NumberPlate-VoVNet.yaml"
MODEL:
  WEIGHTS: "https://www.dropbox.com/s/q98pypf96rhtd8y/vovnet39_ese_detectron2.pth?dl=1"
  VOVNET:
    CONV_BODY : "V-39-eSE"
SOLVER:
  STEPS: (0, 60000)
  MAX_ITER: 60000
OUTPUT_DIR: "output/numberplate/numberplate-V-39-ms-3x"

• Base-NumberPlate-VoVNet.yaml

_MODEL:
  META_ARCHITECTURE: "GeneralizedRCNN"
  BACKBONE:
    NAME: "build_fcos_vovnet_fpn_backbone"
    FREEZE_AT: 0
  VOVNET:
    OUT_FEATURES: ["stage3", "stage4", "stage5"]
  FPN:
    IN_FEATURES: ["stage3", "stage4", "stage5"]
  PROPOSAL_GENERATOR:
    NAME: "FCOS"  
  FCOS:
    POST_NMS_TOPK_TEST: 50
  # PIXEL_MEAN: [102.9801, 115.9465, 122.7717]
  MASK_ON: True
  MASKIOU_ON: True
  ROI_HEADS:
    NAME: "CenterROIHeads"
    IN_FEATURES: ["p3", "p4", "p5"]
  ROI_MASK_HEAD:
    NAME: "SpatialAttentionMaskHead"
    ASSIGN_CRITERION: "ratio"
    NUM_CONV: 4
    POOLER_RESOLUTION: 14
DATASETS:
  TRAIN: ("numberplate_train",)
  TEST: ("numberplate_val",)
SOLVER:
  CHECKPOINT_PERIOD: 6000
  IMS_PER_BATCH: 16
  BASE_LR: 0.01  # Note that RetinaNet uses a different default learning rate
  STEPS: (0, 60000)
  MAX_ITER: 60000
INPUT:
  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)

After that I added to the file ./centermask2/train_net.py in the section main the registration of my own dataset:

    # Regist own dataset.
    from detectron2.data.datasets import register_coco_instances
    # train data
    name        = "numberplate_train"
    json_file   = "/var/www/centermask2/datasets/numberplate/train/coco_numberplate.json"
    image_root  = "/var/www/centermask2/datasets/numberplate/train"
    # test data
    name_val        = "numberplate_val"
    json_file_val   = "./centermask2/datasets/numberplate/val/coco_numberplate.json"
    image_root_val  = "./centermask2/datasets/numberplate/val"
    # registr
    register_coco_instances(name, {}, json_file, image_root)
    register_coco_instances(name_val, {}, json_file_val, image_root_val)

In the end, I ran the command line:

DETECTRON2_DATASETS=./centermask2/datasets python3 train_net.py --config-file "configs/numberplates/numberplate_V_39_eSE_FPN_ms_3x.yaml" --num-gpus 1

Hello @dimabendera,

Thanks for the information.

I am implementing centermask2 on my custom data as well. All your steps have been helpful. I just wanted to know in which format did you register your annotations? Everything seems to work if the segmentation mask is in polygon format but I am having a problem while working with the mask in RLE format? Could you provide some information regarding the same?

from centermask2.

It seems that if own datasets only has one class, the NUM_CLASSES should be 2 which contains background.
When training with my own datasets, if this value set as 1 , the loss will be very small and quickly decreases to 0. If this value set as 2, the loss seems normal.

you are right. If the dataset contains 2 classes. so we have to set these params to 3: 2 classes + background

ALSO!! don't forget to do the same definition when you run predict.py

from centermask2.

@Kenneth-X

you just insert the config option in config file (.yaml) or command line.

For example,

python train_net.py --config-file <config.yaml> --num-gpus 8 MODEL.ROI_HEADS.NUM_CLASSES 1

But how could I config in the config-file (.yaml), is there any format for that

from centermask2.

Hello everyone

I am trying to implement CenterMask2 https://github.com/youngwanLEE/centermask2 which makes use of Detectron2.
My annotations are in COCO RLE format, my data registration is successful. I visualized the data as well using the detectron2 visualizer, also I have set cfg.INPUT.MASk_FORMAT= bitmask, but somehow this error is being raised while I try to train the network.

Can somebody help me out here with what could be done?

File "train_net.py", line 181, in <module>
    launch(
  File "/workspace/detectron2/detectron2/engine/launch.py", line 84, in launch
    main_func(*args)
  File "train_net.py", line 151, in main
    return trainer.train()
  File "/workspace/detectron2/detectron2/engine/defaults.py", line 484, in train
    super().train(self.start_iter, self.max_iter)
  File "/workspace/detectron2/detectron2/engine/train_loop.py", line 155, in train
    self.run_step()
  File "/workspace/detectron2/detectron2/engine/defaults.py", line 494, in run_step
    self._trainer.run_step()
  File "/workspace/detectron2/detectron2/engine/train_loop.py", line 287, in run_step
    loss_dict = self.model(data)
  File "/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/workspace/detectron2/detectron2/modeling/meta_arch/rcnn.py", line 167, in forward
    _, detector_losses = self.roi_heads(images, features, proposals, gt_instances)
  File "/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/workspace/centermask2/centermask/modeling/centermask/center_heads.py", line 401, in forward
    losses, mask_features, selected_mask, labels, maskiou_targets = self._forward_mask(features, proposals)
  File "/workspace/centermask2/centermask/modeling/centermask/center_heads.py", line 476, in _forward_mask
    loss, selected_mask, labels, maskiou_targets = mask_rcnn_loss(mask_logits, proposals, self.maskiou_on)
  File "/workspace/centermask2/centermask/modeling/centermask/mask_head.py", line 80, in mask_rcnn_loss
    cropped_mask = crop(instances_per_image.gt_masks.polygons, instances_per_image.proposal_boxes.tensor)
AttributeError: 'BitMasks' object has no attribute 'polygons'

from centermask2.