-
Notifications
You must be signed in to change notification settings - Fork 30
/
Copy pathindex.html
106 lines (103 loc) · 4.88 KB
/
index.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
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
<!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="super-convergence">Super Convergence</h1>
<h2 id="assignment">Assignment</h2>
<ol>
<li>Write a code that draws this curve (without the arrows). In submission, you’ll upload your drawn curve and code for that
<ol>
<li><img src="https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/11_SuperConvergence/assets/11s11.png?raw=true" alt="enter image description here"></li>
</ol>
</li>
<li>Write a code which
<ol>
<li>uses this new ResNet Architecture for Cifar10:
<ol>
<li>PrepLayer - Conv 3x3 s1, p1) >> BN >> RELU [64k]</li>
<li>Layer1 -
<ol>
<li>X = Conv 3x3 (s1, p1) >> MaxPool2D >> BN >> RELU [128k]</li>
<li>R1 = ResBlock( (Conv-BN-ReLU-Conv-BN-ReLU))(X) [128k]</li>
<li>Add(X, R1)</li>
</ol>
</li>
<li>Layer 2 -
<ol>
<li>Conv 3x3 [256k]</li>
<li>MaxPooling2D</li>
<li>BN</li>
<li>ReLU</li>
</ol>
</li>
<li>Layer 3 -
<ol>
<li>X = Conv 3x3 (s1, p1) >> MaxPool2D >> BN >> RELU [512k]</li>
<li>R2 = ResBlock( (Conv-BN-ReLU-Conv-BN-ReLU))(X) [512k]</li>
<li>Add(X, R2)</li>
</ol>
</li>
<li>MaxPooling with Kernel Size 4</li>
<li>FC Layer</li>
<li>SoftMax</li>
</ol>
</li>
<li>Uses One Cycle Policy such that:
<ol>
<li>Total Epochs = 24</li>
<li>Max at Epoch = 5</li>
<li>LRMIN = FIND</li>
<li>LRMAX = FIND</li>
<li>NO Annihilation</li>
</ol>
</li>
<li>Uses this transform -RandomCrop 32, 32 (after padding of 4) >> FlipLR >> Followed by CutOut(8, 8)</li>
<li>Batch size = 512</li>
<li>Target Accuracy: 90%.</li>
<li>The lesser the modular your code is (i.e. more the code you have written in your Colab file), less marks you’d get.</li>
</ol>
</li>
<li>Questions asked are:
<ol>
<li>Upload the code you used to draw your ZIGZAG or CYCLIC TRIANGLE plot.</li>
<li>Upload your triangle Plot which was drawn with your code.</li>
<li>Upload the link to your GitHub copy of Colab Code.</li>
<li>Upload the github link for the model as described in A11.</li>
<li>What is your test accuracy?</li>
</ol>
</li>
</ol>
<h2 id="solution">Solution</h2>
<p>Github link: <a href="https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/11_SuperConvergence/SuperConvergence.ipynb">https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/11_SuperConvergence/SuperConvergence.ipynb</a></p>
<p>Colab link: <a href="https://colab.research.google.com/github/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/11_SuperConvergence/SuperConvergence.ipynb">https://colab.research.google.com/github/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/11_SuperConvergence/SuperConvergence.ipynb</a></p>
<p>PySodium: <a href="https://github.com/satyajitghana/PySodium">https://github.com/satyajitghana/PySodium</a></p>
<p>Triangle Pattern: <a href="https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/11_SuperConvergence/CycleLR.ipynb">https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/11_SuperConvergence/CycleLR.ipynb</a></p>
<h3 id="code-for-pattern">code for pattern</h3>
<p><img src="https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/11_SuperConvergence/assets/pattern_code.PNG?raw=true" alt="enter image description here"></p>
<p><img src="https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/11_SuperConvergence/assets/one_cycle_fig.png?raw=true" alt="enter image description here"></p>
<h2 id="model-stats">Model Stats</h2>
<pre><code>Test Accuracy: 89.97
Train Accuracy: 92.66
Params: 6,573,120
</code></pre>
<h3 id="lr-finder">LR Finder</h3>
<pre><code>[ 2020-04-11 17:52:58,135 - sodium.sodium.runner ] INFO: sorted lrs : [0.609391, 0.61039, 0.6083919999999999, 0.611389, 0.607393, 0.606394, 0.63037, 0.613387, 0.626374, 0.612388]
[ 2020-04-11 17:52:58,137 - sodium.sodium.runner ] INFO: found the best lr : 0.609391
</code></pre>
<p><img src="https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/11_SuperConvergence/assets/lr_finder.png?raw=true" alt="enter image description here"></p>
<pre><code>[ 2020-04-11 17:53:02,878 - sodium.sodium.runner ] INFO: using max_lr : 0.609391
[ 2020-04-11 17:53:02,880 - sodium.sodium.runner ] INFO: using min_lr : 0.02437564
[ 2020-04-11 17:53:02,880 - sodium.sodium.runner ] INFO: using initial_lr : 0.02437564
</code></pre>
<h3 id="learning-rate">Learning Rate</h3>
<p><img src="https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/11_SuperConvergence/assets/lr_metric.png?raw=true" alt="enter image description here"></p>
<h3 id="model-accuracy-loss-curves">Model Accuracy-Loss Curves</h3>
<p><img src="https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/11_SuperConvergence/assets/model_stats.png?raw=true" alt="enter image description here"></p>
</div>
</body>
</html>