GluonCV Darknet 19

darknet.py

class Darknet(gluon.HybridBlock):
  def __init__(self, spec, classes=1000, batch_norm=True, **kwargs):
    super(Darknet, self).__init__(**kwargs)
    filters = spec['filters']
    kernels = spec['kernels']
    with self.name_scope():
      self.features = nn.HybridSequential()
      first_stage = True
      for sf, sk in zip(filters, kernels):
        if not first_stage:
          # maxpool
          self.features.add(nn.MaxPool2D(pool_size=2, strides=2))
          first_stage = False
        for f, k in zip(sf, sk):
          assert k in [1, 3], "Unexpected kernel size: {}".format(k)
          pad = (k - 1) // 2
          self.features.add(nn.Conv2D(f, kernel_size=k, padding=pad))
          if batch_norm:
            self.features.add(nn.BatchNorm())
           self.features.add(nn.LeakyReLu(0.1))
      # output  
      self.output = nn.Conv2D(classes, kernel_size=1, padding=0)
      
  def hybrid_forward(self, F, x):
    x = self.output(self.features(x))
    return F.Pooling(x, kernel=(7, 7), global_pool=True, pool_type='avg')  
  
  
  
  
darknet_spec = {
  19: {
    'filters': [[32], [64], [128, 64, 128], [256, 128, 256], [512, 256, 512, 256, 512], [1024, 512, 1024, 512, 1024]]
    'kernels': [[3],  [3],  [3, 1, 3],      [3, 1, 3],       [3, 1, 3, 1, 3],           [3, 1, 3, 1, 3]]}