https://github.com/chenzx921020/DropBlock-mxnet

DropBlock implementation
https://github.com/chenzx921020/DropBlock-mxnet

Last synced: 2 months ago
JSON representation

DropBlock implementation

• Host: GitHub
• URL: https://github.com/chenzx921020/DropBlock-mxnet
• Owner: chenzx921020
• Created: 2018-11-16T06:53:31.000Z (about 6 years ago)
• Default Branch: master
• Last Pushed: 2018-11-27T07:22:08.000Z (about 6 years ago)
• Last Synced: 2024-08-01T22:39:59.170Z (5 months ago)
• Language: Python
• Size: 2.93 KB
• Stars: 6
• Watchers: 1
• Forks: 2
• Open Issues: 1
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• Awesome-MXNet - DropBlock

README

        
# 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