Reduction of Feature map size in Convolutional Neural Network

Question

In CNN, the way we reduce the feature map size at layers is we use pooling. Pooling makes feature map size into half.

For the following network, if I want a new layer with feature map size somewhere around the middle between 30 and 15, what could be the way I can make changes of the network?

{'l_struct': [-1, -1, -1], 'grid_size': -1, 'l_name': 'data', 'rf_size': -1, 'stride': -1, 'n_channels': 3, 'Featuremap_size': 480}
{'l_struct': [3, 1, 1], 'grid_size': 1, 'l_name': 'conv_1', 'rf_size': 3, 'stride': 1, 'n_channels': 8, 'Featuremap_size': 480}
{'l_struct': [1, 1, 0], 'grid_size': 1, 'l_name': 'relu_1', 'rf_size': 3, 'stride': 1, 'n_channels': 8, 'Featuremap_size': 480}
{'l_struct': [2, 2, 0], 'grid_size': 2, 'l_name': 'pool_1', 'rf_size': 4, 'stride': 2, 'n_channels': 8, 'Featuremap_size': 240}
{'l_struct': [3, 1, 1], 'grid_size': 2, 'l_name': 'conv_2', 'rf_size': 8, 'stride': 2, 'n_channels': 8, 'Featuremap_size': 240}
{'l_struct': [1, 1, 0], 'grid_size': 2, 'l_name': 'relu_2', 'rf_size': 8, 'stride': 2, 'n_channels': 8, 'Featuremap_size': 240}
{'l_struct': [2, 2, 0], 'grid_size': 4, 'l_name': 'pool_2', 'rf_size': 10, 'stride': 4, 'n_channels': 8, 'Featuremap_size': 120}
{'l_struct': [3, 1, 1], 'grid_size': 4, 'l_name': 'conv_3', 'rf_size': 18, 'stride': 4, 'n_channels': 16, 'Featuremap_size': 120}
{'l_struct': [1, 1, 0], 'grid_size': 4, 'l_name': 'relu_3', 'rf_size': 18, 'stride': 4, 'n_channels': 16, 'Featuremap_size': 120}
{'l_struct': [2, 2, 0], 'grid_size': 8, 'l_name': 'pool_3', 'rf_size': 22, 'stride': 8, 'n_channels': 16, 'Featuremap_size': 60}
{'l_struct': [3, 1, 1], 'grid_size': 8, 'l_name': 'conv_4', 'rf_size': 38, 'stride': 8, 'n_channels': 16, 'Featuremap_size': 60}
{'l_struct': [1, 1, 0], 'grid_size': 8, 'l_name': 'relu_4', 'rf_size': 38, 'stride': 8, 'n_channels': 16, 'Featuremap_size': 60}
{'l_struct': [2, 2, 0], 'grid_size': 16, 'l_name': 'pool_4', 'rf_size': 46, 'stride': 16, 'n_channels': 16, 'Featuremap_size': 30}
{'l_struct': [3, 1, 1], 'grid_size': 16, 'l_name': 'conv_5', 'rf_size': 78, 'stride': 16, 'n_channels': 32, 'Featuremap_size': 30}
{'l_struct': [1, 1, 0], 'grid_size': 16, 'l_name': 'relu_5', 'rf_size': 78, 'stride': 16, 'n_channels': 32, 'Featuremap_size': 30}
{'l_struct': [2, 2, 0], 'grid_size': 32, 'l_name': 'pool_5', 'rf_size': 94, 'stride': 32, 'n_channels': 32, 'Featuremap_size': 15}
{'l_struct': [3, 1, 1], 'grid_size': 32, 'l_name': 'conv_6', 'rf_size': 158, 'stride': 32, 'n_channels': 32, 'Featuremap_size': 15}
{'l_struct': [1, 1, 0], 'grid_size': 32, 'l_name': 'relu_6', 'rf_size': 158, 'stride': 32, 'n_channels': 32, 'Featuremap_size': 15}

shimao · Accepted Answer · 2018-08-15T06:18:15.030

1

To obtain a feature map of some arbitrary size, resizing with bilinear interpolation is a good bet since it is differentiable.

edit: actually any reasonable interpolation mode (such as nearest neighbors) would work as long as the gradients are implemented.

edited Aug 15 '18 at 06:18

answered Aug 15 '18 at 04:48

shimao

22,706
2
42
81

Is that just normal bilinear interpolation applied to images? – batuman Aug 15 '18 at 06:15
@batuman yep, except you'd apply it to feature maps. – shimao Aug 15 '18 at 06:16

Reduction of Feature map size in Convolutional Neural Network

1 Answers1