|
<!DOCTYPE html> |
|
<html lang="en"> |
|
<head> |
|
<style> |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
body { |
|
box-sizing: border-box; |
|
min-width: 200px; |
|
max-width: 1110px; |
|
margin: 0 auto; |
|
padding: 45px; |
|
} |
|
|
|
@media (prefers-color-scheme: dark) { |
|
body { |
|
background-color: #0d1117; |
|
} |
|
} |
|
|
|
|
|
.github-fork-ribbon:before { |
|
background-color: #121612; |
|
} |
|
|
|
.banner-image { |
|
border-radius: 50px; |
|
} |
|
</style> |
|
|
|
<meta charset="utf-8"> |
|
<title> |
|
Animated AI |
|
</title> |
|
<meta content="width=device-width, initial-scale=1.0" name="viewport"> |
|
</head> |
|
<body> |
|
<article class="markdown-body"> |
|
<h1> |
|
<div> |
|
<img class="banner-image" src="rick.jpg" width="100%"> |
|
</div> |
|
</h1> |
|
<p> |
|
I'm creating animations and instructional videos about neural networks. |
|
Check out my |
|
<a> |
|
Patreon |
|
</a> |
|
and |
|
<a> |
|
YouTube channel |
|
</a> |
|
! |
|
</p> |
|
<h2> |
|
Convolution |
|
</h2> |
|
<h3> |
|
The Basic Algorithm |
|
</h3> |
|
Watch the companion YouTube video: |
|
<a> |
|
Fundamental Algorithm of Convolution in Neural Networks |
|
</a> |
|
. |
|
<img alt="The process of convolution in neural networks with a 3x3 kernel size" height="auto" src="rick.jpg" width="960"> |
|
<h3> |
|
Padding |
|
</h3> |
|
Companion video: |
|
<a> |
|
Convolution Padding - Neural Networks |
|
</a> |
|
<table> |
|
<tbody> |
|
<tr> |
|
<td> |
|
<img alt="The process of convolution in neural networks with a 3x3 kernel size" height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
<td> |
|
<img alt="The process of convolution in neural networks with a 3x3 kernel size and a padding of 'SAME', i.e., 1 pixel on the top, bottom, left, and right." height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
</tr> |
|
<tr> |
|
<td> |
|
No Padding AKA "Valid" |
|
</td> |
|
<td> |
|
[1,1,1,1] Padding AKA "Same" |
|
</td> |
|
</tr> |
|
</tbody> |
|
</table> |
|
<h3> |
|
Stride |
|
</h3> |
|
Companion video: |
|
<a> |
|
Stride - Convolution in Neural Networks |
|
</a> |
|
<table> |
|
<thead> |
|
<tr> |
|
<th> |
|
Stride 1 |
|
</th> |
|
<th> |
|
Stride 2 |
|
</th> |
|
</tr> |
|
</thead> |
|
<tbody> |
|
<tr> |
|
<td> |
|
<img alt="The process of convolution in neural networks with a 3x3 kernel size" height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
<td> |
|
<img alt="The process of convolution in neural networks with a 3x3 kernel size and a stride of 2x2" height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
</tr> |
|
<tr> |
|
<td> |
|
No Padding AKA "Valid"; Stride of 1 |
|
</td> |
|
<td> |
|
No Padding AKA "Valid"; Stride of 2 |
|
</td> |
|
</tr> |
|
<tr> |
|
<td> |
|
<img height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
<td> |
|
<img alt="The process of convolution in neural networks with a 3x3 kernel size, a stride of 2x2, and a padding of 'SAME', i.e., 1 pixel on the top, bottom, left, and right." height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
</tr> |
|
</tbody> |
|
</table> |
|
<h3> |
|
Groups, Depthwise, and Depthwise-Separable |
|
</h3> |
|
Watch the companion YouTube video: |
|
<a> |
|
Groups, Depthwise, and Depthwise-Separable Convolution (Neural Networks) |
|
</a> |
|
. |
|
<table> |
|
<tbody> |
|
<tr> |
|
<td> |
|
<img alt="The process of convolution in neural networks with a 3x3 kernel size" height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
<td> |
|
<img alt="The process of convolution in neural networks with a 3x3 kernel size and 2 groups" height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
</tr> |
|
<tr> |
|
<td> |
|
1 Group |
|
</td> |
|
<td> |
|
2 Groups |
|
</td> |
|
</tr> |
|
<tr> |
|
<td> |
|
<img alt="The process of convolution in neural networks with a 3x3 kernel size and 8 groups making it a depthwise convolution" height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
<td> |
|
<img alt="A depthwise convolution layer followed by a pointwise convolution layer making the combined process a depthwise-separable convolution" height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
</tr> |
|
<tr> |
|
<td> |
|
Depthwise (8 Groups) |
|
</td> |
|
<td> |
|
Depthwise-separable (8 Groups followed by pointwise) |
|
</td> |
|
</tr> |
|
</tbody> |
|
</table> |
|
<h2 id="pixel-shuffle"> |
|
Pixel Shuffle |
|
</h2> |
|
Watch the companion YouTube video: |
|
<a> |
|
Pixel Shuffle - Changing Resolution with Style |
|
</a> |
|
<h3> |
|
2x2 Block Size |
|
</h3> |
|
<table> |
|
<tbody> |
|
<tr> |
|
<td> |
|
<img alt="Pixel shuffle animation in neural networks with a block size of 2x2" height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
<td> |
|
<img alt="Pixel unshuffle animation in neural networks with a block size of 2x2" height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
</tr> |
|
<tr> |
|
<td> |
|
2x2 Pixel Shuffle |
|
</td> |
|
<td> |
|
2x2 Pixel Unshuffle |
|
</td> |
|
</tr> |
|
<tr> |
|
<td colspan="2"> |
|
<img alt="Pixel shuffle and unshuffle in neural networks with a block size of 2x2 looping animation" height="auto" src="rick.jpg" width="960"> |
|
</td> |
|
</tr> |
|
<tr> |
|
<td colspan="2"> |
|
2x2 Pixel Shuffle/Unshuffle Loop |
|
</td> |
|
</tr> |
|
</tbody> |
|
</table> |
|
<h3> |
|
3x3 Block Size |
|
</h3> |
|
<table> |
|
<tbody> |
|
<tr> |
|
<td> |
|
<img alt="Pixel shuffle animation in neural networks with a block size of 3x3" height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
<td> |
|
<img alt="Pixel unshuffle animation in neural networks with a block size of 3x3" height="auto" src="rick.jpg" width="480"> |
|
</td> |
|
</tr> |
|
<tr> |
|
<td> |
|
3x3 Pixel Shuffle |
|
</td> |
|
<td> |
|
3x3 Pixel Unshuffle |
|
</td> |
|
</tr> |
|
<tr> |
|
<td colspan="2"> |
|
<img alt="Pixel shuffle and unshuffle in neural networks with a block size of 3x3 looping animation" height="auto" src="rick.jpg" width="960"> |
|
</td> |
|
</tr> |
|
<tr> |
|
<td colspan="2"> |
|
3x3 Pixel Shuffle/Unshuffle Loop |
|
</td> |
|
</tr> |
|
</tbody> |
|
</table> |
|
<h2> |
|
</h2> |
|
<p> |
|
Licensed under the |
|
<a> |
|
MIT License |
|
</a> |
|
</p> |
|
</article> |
|
</body> |
|
</html> |
|
|