Training param tuning

Training param in prototxt

There has two types data augmentation method for different application

https://github.com/eric612/MobileNet-YOLO/issues/29

For example , I will choose adaptive aspect ratio in fisheye videos which were pixelwise geometry distortion

BGR to RGB

Convert channel in training and test phase , add parameter in prototxt

layer {
  name: "data"
  type: "AnnotatedData"
  ...
  transform_param {
    scale: ...
    cvt_bgr2rgb: true 
    ...
  }
  ...
}

keep aspect ratio

This type may decrease accuracy when input size < 416 (compare with "Adaptive aspect ratio")

• Set preprocessing resize mode to "FIT_LARGE_SIZE_AND_PAD"
• Inference use "FIT_LARGE_SIZE_AND_PAD" resize

Adaptive aspect ratio

This type may break k-mean anchors rule and effect accuracy about ±1% in my test

• Set preprocessing resize mode to "WARP"
• Expand param set to {VOC:4.0 , COCO:1.5 , ...}
• Inference use "WARP" resize

Warm up training

If solver type set to "SGD" , you may need set learning rate policy like this

Pre-trained weights and batch size

total_batch_size = iter_size * batch_size

If pre-trained weights use

• Classification model (like imagenet)

  total_batch_size set to 64 at least

• Detection model (like ms-coco)

  total_batch_size set to 16 at least ("PASCAL-VOC")

  total_batch_size set to 32 at least , recommend to 64 ("MS-COCO")

Loss do not convergence or not stable

You can try below training tricks

• Decrease/Increase expand_param , batch sampler or jitter scale (data augmentation)
• Set larger batch size
• Set lr_mult:0.1 or 0.2 and also decay_mult at first 0 ~ x layers (x is 10 for MobileNet) , base on your backbone network
• Changes solver type
• Set rms_decay at range 0.9~0.98 in solver prototxt (rmsprop only)