ssd300_vgg_voc.json

{
  "model": "ssd_vgg",
  "batchnorm": true,
  "input_info": {
    "sample_size":  [2, 3, 300, 300]
  },
  "basenet": "./vgg16_bn-6c64b313.pth", // Download this from https://download.pytorch.org/models/vgg16_bn-6c64b313.pth if you want to pretrain your own FP32 net
  "num_classes": 21,
  "dataset": "voc",
  "preprocessing": {
    "mean": [0.406, 0.456, 0.485],
    "std": [0.255, 0.224, 0.229],
    "normalize_coef": 255,
    "rgb": true
  },
  "batch_size": 256,
  "epochs": 290,
  "save_freq": 50,
  "optimizer": {
    "type": "Adam",
    "base_lr": 1e-4,
    "weight_decay": 5e-4,
    "schedule_type": "multistep",
    "steps": [2, 3, 4]
  },
  "ssd_params": {
    "clip": false,
    "variance": [0.1, 0.1, 0.2, 0.2],
    "max_sizes": [60, 111, 162, 213, 264, 315],
    "min_sizes": [30, 60, 111, 162, 213, 264],
    "steps": [8, 16, 32, 64, 100, 300],
    "aspect_ratios": [[2], [2, 3], [2, 3], [2, 3], [2], [2]],
    "flip": true
  }
}