Detectron

Detectron is the object detection collection from Facebook Research. Original repo is here. Since facebook research version could encounter some errors during parsing yaml file, I have fixed this issue by fixing pyyamml version in requirement.txt. Related Issue In my repository I use my forked version of Detectron and works well now.

How to use

Build image by yourself

• Clone this repository with submodule: git clone --recursive git@github.com:TeeboneTing/Detectron_k8s.git

• Build image: make build

• If you would like to push to your dockerhub repository, please change your username in Makefile line 3 and make push_dockerhub

Acutally you can run docker image without build by yourself

• Run image by docker run -ti teeboneding/detectron bash

• Execute example inference command inside container from GETTING_STARTED:

python tools/infer_simple.py 
    --cfg configs/12_2017_baselines/e2e_mask_rcnn_R-101-FPN_2x.yaml 
    --output-dir /tmp/detectron-visualizations 
    --image-ext jpg 
    --wts https://dl.fbaipublicfiles.com/detectron/35861858/12_2017_baselines/e2e_mask_rcnn_R-101-FPN_2x.yaml.02_32_51.SgT4y1cO/output/train/coco_2014_train:coco_2014_valminusminival/generalized_rcnn/model_final.pkl 
    demo

Train RetinaNet

Prepare Dataset

• According to dataset readme, create your own dataset with COCO format and put into path detectron/datasets/data.

• To add your dataset into Detectron catalog, please refer to dataset_catalog.py. Add your dataset information in _DATASETS dict.

Start Training

1. Make a soft link for dataset to

   correct place

2. Setup RetinaNet config files and model output path

3. Inside the container with commands below:

ln -s /tmp/detectron/coco /detectron/detectron/datasets/data/coco ; 
python tools/train_net.py --cfg new_configs/getting_started/retinanet_X-101-32x8d-FPN_1x.yaml OUTPUT_DIR /tmp/detectron/detectron/model

Current Training Parameters

ref: retinanet_X-101-32x8d-FPN_1x.yaml

MODEL:
  TYPE: retinanet
  CONV_BODY: FPN.add_fpn_ResNet101_conv5_body
  NUM_CLASSES: 2 # Class number should be your dataset class number + 1 (for background class)
NUM_GPUS: 1 # Change your GPU number here
SOLVER:
  WEIGHT_DECAY: 0.0001
  LR_POLICY: steps_with_decay
  BASE_LR: 0.001 # Original LR is 0.01 which is too high for a pretrained model
  GAMMA: 0.1
  MAX_ITER: 90000
  STEPS: [0, 60000, 80000]
FPN:
  FPN_ON: True
  MULTILEVEL_RPN: True
  RPN_MAX_LEVEL: 7
  RPN_MIN_LEVEL: 3
  COARSEST_STRIDE: 128

  EXTRA_CONV_LEVELS: True
RESNETS:
  STRIDE_1X1: False  # default True for MSRA; False for C2 or Torch models
  TRANS_FUNC: bottleneck_transformation
  NUM_GROUPS: 32
  WIDTH_PER_GROUP: 8
RETINANET:
  RETINANET_ON: True
  NUM_CONVS: 4
  ASPECT_RATIOS: (1.0, 2.0, 0.5)
  SCALES_PER_OCTAVE: 3
  ANCHOR_SCALE: 4
  LOSS_GAMMA: 2.0
  LOSS_ALPHA: 0.25
TRAIN:
  WEIGHTS: /tmp/detectron/detectron/X-101-32x8d.pkl # ImageNet pretrained model. Download link: https://dl.fbaipublicfiles.com/detectron/ImageNetPretrained/20171220/X-101-32x8d.pkl
  DATASETS: ('ubitus_anime2019',) # Replace your training/validation dataset name here
  #SCALES: (500,)
  #MAX_SIZE: 833
  #IMS_PER_BATCH: 4
  #BATCH_SIZE_PER_IM: 128
  RPN_PRE_NMS_TOP_N: 2000  # Per FPN level
  RPN_STRADDLE_THRESH: -1  # default 0
#TEST:
#  DATASETS: ('coco_2014_minival',)
#  SCALE: 800
#  MAX_SIZE: 1333
#  NMS: 0.5
#  RPN_PRE_NMS_TOP_N: 10000  # Per FPN level
#  RPN_POST_NMS_TOP_N: 2000
OUTPUT_DIR: .

DATA_LOADER:
  MINIBATCH_QUEUE_SIZE: 64

Some example output images: