search

msracver / Deformable-ConvNets

Deformable Convolutional Networks
MIT License
4.04k stars 957 forks source link

The accuracy of deeplab #103

Open Mendel1 opened 7 years ago

Mendel1 commented 7 years ago

I retrain the deeplab part and the result is 61% for baseline while 66% for dcn. This is a lot below the official result. I only changed the gpu to 2 and remove the multiprocess in loader. Anyone know why it came to be that? the config file for dcn

---
MXNET_VERSION: "mxnet"
output_path: "./output/voc12"
symbol: resnet_v1_101_deeplab_dcn
gpus: '0,1'
SCALES:
- 360
- 600
default:
  frequent: 10
  kvstore: device
dataset:
  NUM_CLASSES: 21
  dataset: PascalVOC
  dataset_path: "./data/VOCdevkit2012/"
  image_set: 2012_train_seg
  root_path: "./data/"
  test_image_set: 2012_val_seg
network:
  FIXED_PARAMS:
  - conv1
  - bn_conv1
  - res2
  - bn2
  - gamma
  - beta
  FIXED_PARAMS_SHARED:
  - conv1
  - bn_conv1
  - res2
  - bn2
  - res3
  - bn3
  - res4
  - bn4
  - gamma
  - beta
  IMAGE_STRIDE: 0
  PIXEL_MEANS:
  - 103.06
  - 115.90
  - 123.15
  pretrained: "./model/pretrained_model/resnet_v1_101"
  pretrained_epoch: 0
TRAIN:
  warmup: false
  warmup_lr: 0.00005
  warmup_step: 1000
  begin_epoch: 0
  end_epoch: 12
  lr: 0.0005
  lr_step: '8'
  model_prefix: "deeplab_resnet_v1_101_voc12_segmentation_dcn"
  FLIP: true
  BATCH_IMAGES: 1
  ENABLE_CROP: False
  CROP_HEIGHT: 768
  CROP_WIDTH: 1024
  RESUME: false
  SHUFFLE: true
TEST:
  BATCH_IMAGES: 1
  test_epoch: 12
larsoncs commented 6 years ago

@Mendel1 have you retrain other part, for exampe ,rfcn and faster rcnn? what is the performace of other parts? thank you.