狠狠撸

狠狠撸Share a Scribd company logo

PyData London CNN Lightning Talk

?
?
E
Eddie Bell

This document discusses convolutional neural networks and image filtering. It defines key concepts like image size, filter size, padding, and stride that are used to calculate the size of convolved feature maps. Examples of VGG neural networks using 3x3 convolutions with padding and stride of 1 are provided. Resources on image size, model parameters, and parallelization techniques are also listed.

1 of 20
Downloaded 29 times
X - F + 2P S @ejlbell 1 +
Convolution 1 1 0 0 1 0 1 1 1 0 0 0 1 0 0 1 0 0 0 0 1 1 1 1 1 4 Image Convoluted Feature
Image Filters -1 -2 -1 0 0 0 +1 +2 +1 -1 0 +1 -2 0 +2 -1 0 +1
X X X - F + 2P S X = image size 1 +
X X F X = image size F = ?lter size X - F + 2P S 1 +
X XX = image size F = ?lter size P = padding F P X - F + 2P S 1 +
X XX = image size F = ?lter size P = padding S = stride P X - F + 2P S 1 +
X XX = image size F = ?lter size P = padding S = stride P X - F + 2P S 1 +
X XX = image size F = ?lter size P = padding S = stride P X - F + 2P S 1 +
X XX = image size F = ?lter size P = padding S = stride P X - F + 2P S 1 +
X XX = image size F = ?lter size P = padding S = stride P X - F + 2P S 1 +
X XX = image size F = ?lter size P = padding S = stride P X - F + 2P S 1 +
X XX = image size F = ?lter size P = padding S = stride P X - F + 2P S 1 +
Filter Size 1 × 1 3 × 3 7 × 7 9 × 95 × 5
Max Pooling 1 6 2 2 4 1 0 4 1 5 3 1 2 1 1 3 8 4 6 3 X Y
Example: VGG 19 layers 3x3 convolution pad 1 stride 1
19 layers 3x3 convolution pad 1 stride 1 X - F + 2P S 1 + = X Example: VGG
Resources ? image size ? parameters (dense vs conv) ? parallelisation (data vs model)
thanks @ejlbell
References

More Related Content

PyData London CNN Lightning Talk