Awesome

DropBlock

Implementation of DropBlock: A regularization method for convolutional networks in mxnet.

Usage

initail parameters

In your symbol,between convolutional operators, you need to previously compute the feature map size and confirm the mask size, this part will be improved in future commit.

self.block_mask = nd.ones((256, 48, 7, 7)) # mask size:(batch_size,channel,mask_size,mask_size)

operator implementation

In my experiment, feature map size is 7, the schedule for drop block probability has been finished, you can set step and prob range in operator.

drop_layer = mx.sym.Custom(conv5,drop_prob=0.0,block_size=3,drop_prob_max=0.3,step=15000,block_factor_prob = 0.04 ,op_type = 'DropBlock')

For different task, maybe you need to try different parameters for many times.

TODO

DropBlock for 3D convolution