How to train my own datasets (format is like coco datasets)

[Xavier-Zeng]

Now I have converted my datasets format to coco format, andI want to train my own datasets using FCOS. I referenced GETTING_STARTED.md in mmdetection repo, and there is a tutorial in mmdetection repo to train my own datasets. But in FCOS repo, I find the file FCOS/maskrcnn_benchmark/data/datasets/coco.py is different like /mmdetection/mmdet/datasets/coco.py. Is there any suggestions?

[tianzhi0549]

@EDG-Zola You do not need to change this code.
In order to train FCOS on your own dataset, you need to,

1. Add you dataset to

   FCOS/fcos_core/config/paths_catalog.py

   Line 10 in efb76e4

   "coco_2017_train": {

   . Please use _coco_style as the suffix of your dataset names.
2. In https://github.com/tianzhi0549/FCOS/blob/master/configs/fcos/fcos_R_50_FPN_1x.yaml, change DATASETS to your own ones.
3. Modify MODEL.FCOS.NUM_CLASSES in

   FCOS/maskrcnn_benchmark/config/defaults.py

   Line 284 in ff8376b

   _C.MODEL.FCOS.NUM_CLASSES = 81 # the number of classes including background

   if your dataset has a different number of classes.

[sunpeng981712364]

Thanks for great works! just a refered question. If I have 29 classes, _C.MODEL.FCOS.NUM_CLASSES should be set to 30?

[sunpeng981712364]

@EDG-Zola You do not need to change this code.
In order to train FCOS on your own dataset, you need to,

1. Add you dataset to https://github.com/tianzhi0549/FCOS/blob/master/maskrcnn_benchmark/config/defaults.py. Please use _coco_style as the suffix of your dataset names.
2. In https://github.com/tianzhi0549/FCOS/blob/master/configs/fcos/fcos_R_50_FPN_1x.yaml, change DATASETS to your own ones.
3. Modify MODEL.FCOS.NUM_CLASSES in

   FCOS/maskrcnn_benchmark/config/defaults.py

   Line 284 in ff8376b

   _C.MODEL.FCOS.NUM_CLASSES = 81 # the number of classes including background

   
   if your dataset has a different number of classes.
   Thanks for great works! just a refered question. If I have 29 classes, _C.MODEL.FCOS.NUM_CLASSES should be set to 30?

[tianzhi0549]

@sunpeng981712364 If the 29 classes do not contain the background class, NUM_CLASSES should be set as 30.

[sunpeng981712364]

hi, I use fcos_demo.py to visualize the result and it seems right, But when I predict use tools/testnet.py with coco protocol, all the AP/AR is close to zero. Do I need to change tools/testnet.py
@tianzhi0549
Should the following code be add?
top_predictions = self.select_top_predictions(predictions)

[tianzhi0549]

@sunpeng981712364 I am not sure what is wrong with your code. It might be helpful to debug your code line by line.

[sunpeng981712364]

@tianzhi0549 谢谢您的及时回复(#^.^#)嘻嘻

[liuguanglyc]

Can one or two 1080Ti GPU be used to train?

[tianzhi0549]

@liuguanglyc I think you can, but maybe you need to use a smaller input size (e.g., 600px).

[heng2j]

Hi @tianzhi0549 , I am trying to train with my own dataset with fcos_R_101_FPN_2x.

However, I encountered the error that mentioned

RuntimeError: Error(s) in loading state_dict for GeneralizedRCNN:
size mismatch for rpn.head.cls_logits.weight: copying a param with shape torch.Size([80, 256, 3, 3]) from checkpoint, the shape in current model is torch.Size([1, 256, 3, 3]).
size mismatch for rpn.head.cls_logits.bias: copying a param with shape torch.Size([80]) from checkpoint, the shape in current model is torch.Size([11]).

I also removed all the previous checkpoints from ~/.torch/models/

Would you please advice on the steps to retrain a model with your own coco style dataset? Thank you so much!

Training Command

python -m torch.distributed.launch \
    --nproc_per_node=3 \
    --master_port=$((RANDOM + 10000)) \
    tools/train_net.py \
    --skip-test \
    --config-file configs/fcos/fcos_R_101_FPN_2x.yaml \
    DATALOADER.NUM_WORKERS 2 \
    OUTPUT_DIR training_dir/fcos_R_101_FPN_2x

fcos_R_101_FPN_2x.yaml

MODEL:
  META_ARCHITECTURE: "GeneralizedRCNN"
  WEIGHT: "../FCOS/FCOS_R_101_FPN_2x.pth"
  RPN_ONLY: True
  FCOS_ON: True
  BACKBONE:
    CONV_BODY: "R-101-FPN-RETINANET"
  RESNETS:
    BACKBONE_OUT_CHANNELS: 256
  RETINANET:
    USE_C5: False # FCOS uses P5 instead of C5
DATASETS:
  TRAIN: ("my_data_train_coco_style", "my_data_val_coco_style")
  TEST: ("my_data_test_coco_style",)
INPUT:
  MIN_SIZE_RANGE_TRAIN: (640, 800)
  MAX_SIZE_TRAIN: 1333
  MIN_SIZE_TEST: 800
  MAX_SIZE_TEST: 1333
DATALOADER:
  SIZE_DIVISIBILITY: 32
SOLVER:
  BASE_LR: 0.01
  WEIGHT_DECAY: 0.0001
  STEPS: (120000, 160000)
  MAX_ITER: 180000
  IMS_PER_BATCH: 4
  WARMUP_METHOD: "constant"

[heng2j]

Should I follow the retrain instruction from maskrcnn-benchmark, to trim the last layers. And also add the dataset statement in the _init.py file ?

[tianzhi0549]

@heng2j I don't think it is necessary if you have converted your datasets into the coco-style format.

[heng2j]

Hi @tianzhi0549 thank you for your quick response. And how about trim the last layers of I am retaining with the given FCOS_R_101_FPN_2x.pth?

[tianzhi0549]

@heng2j You might need to do that if you want to fine-tune from coco pre-trained models.