-
Notifications
You must be signed in to change notification settings - Fork 30
/
README.html
44 lines (41 loc) · 2.38 KB
/
README.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>README</title>
<link rel="stylesheet" href="https://stackedit.io/style.css" />
</head>
<body class="stackedit">
<div class="stackedit__html"><h1 id="assignment">Assignment:</h1>
<ol>
<li>Go through this repository: <a href="https://github.com/kuangliu/pytorch-cifar">https://github.com/kuangliu/pytorch-cifar (Links to an external site.)</a></li>
<li>Extract the ResNet18 model from this repository and add it to your API/repo.</li>
<li>Use your data loader, model loading, train, and test code to train ResNet18 on Cifar10</li>
<li>Your Target is 85% accuracy. No limit on the number of epochs. Use default ResNet18 code (so params are fixed).</li>
<li>Once done finish S8-Assignment-Solution.</li>
</ol>
<h1 id="solution">Solution</h1>
<p>Github-Notebook : <a href="https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/08_AdvancedArchitecture/ResNet18.ipynb">https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/08_AdvancedArchitecture/ResNet18.ipynb</a></p>
<p>Colab-Notebook : <a href="https://colab.research.google.com/github/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/08_AdvancedArchitecture/ResNet18.ipynb">https://colab.research.google.com/github/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/08_AdvancedArchitecture/ResNet18.ipynb</a></p>
<p>PySodium Library : <a href="https://github.com/satyajitghana/PySodium">https://github.com/satyajitghana/PySodium</a></p>
<pre><code>Model : ResNet18
Data Augmentation : RandomCrop, RandomHorizontalFlip
Epochs : 15
Scheduler : OneCycleLR
Optimizer : SGD
Train Accuracy (Best) : 95.63
Test Accuracy (Best) : 90.84
</code></pre>
<h2 id="metrics">Metrics</h2>
<p><img src="https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/08_AdvancedArchitecture/metrics.png?raw=true" alt="metrics.png"></p>
<h2 id="conclusions-">Conclusions :</h2>
<ul>
<li>It was quite fun to create a library, there are a lot of design choices you have to make and stick with it throughout. You get to learn how to write generic functions and how to export them.</li>
<li>ResNet18 model overfits, so we added data augmentation to fix that</li>
<li>OneCycleLR helps achieve high accuracy within very less epochs</li>
<li>The model still overfits, maybe add L2 Regularization ?</li>
</ul>
</div>
</body>
</html>