Getting Started with CNN (GDG Gandhinagar)

Get Start with Convolution Neural Network

Transcript

3. Topics:- 1. Computer Vision 2. CNN & Types 3. Edge Detection 4. Construction & Working of CNN 5. CNNs architectures
4. Computer Vision:-
5. What is CNN? • • • • •
6. Image Classification Neural Style Transfer Object Detection
7. Edge Detection
8. 3 0 1 2 7 4 1 5 8 9 3 1 2 7 2 5 1 3 0 1 3 1 7 8 4 2 1 6 2 8 2 4 5 2 3 9 E.G.:- * 1 0 -1 1 0 -1 1 0 -1 Convolution Step 1 :-- (3*1 + 1*1 + 2*1 + 0*0 + 5*0 + 7*0 + 1*-1 + 8*-1 + 2*-1) “1 Step Aagad Vadho” Step 2 :-- (0*1 + 5*1 + 7*1 + 1*0 + 8*0 + 2*0 + 2*-1 + 9*-1 + 5*-1) . “Evi Rite Niche” Step 5 :-- (1*1 + 2*1 + 0*1 + 5*0 + 7*0 + 1*0 + 8*-1 + 2*-1 + 3*-1) -5 -4 0 8 -10 -2 2 3 0 -2 -4 -7 -3 -2 -3 -16 6 x 6 Vertical Filter of 3x3 =
9. Samaj me aya? 1 0 -1 2 0 -2 2 0 -1 3 0 -3 10 0 -10 3 0 -3 Sobel Filter Schar Filter
11. * * = = 6x6x3 3x3x3 4x4 Horizontal Vertical 4x4x2 (n-f+1) x (n-f+1) x nc’ Where:- nc’ = no. of channels Here( 4x4x2 )
12. Padding (p):- 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 * = Pela mate ni formula (Valid) :- (n-f+1 , n-f+1) After adding padding ni formula (Same) :- (n+2p-f+1 , n+2p-f+1) nXn fXf pXp n+2p-f+1 = n p = (f-1)/2
13. Stride (s) :- a b c d e f g h i j k l m n o p q r s t u v w x y z 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 3 4 4 1 0 2 -1 0 3 1 2 3 4 5 6 7 8 9 * = | ((n + 2p -f)/s) +1 | x | ((n + 2p -f)/s) +1 | 7x7 3x3 3x3 S = 2
14. Pooling :- 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 6 8 14 16 4x4 2x2 Used for Image compression Hyperparameters :- 1. Filter Size 2. Strides 3. Max Pool or Average Pool nh x nw x nc | ((nh-f)/s)+1 | x | ((nh-f)/s)+1 | x nc Note :- Parameters che j nai
15. Image Conv - 1 Pool - 1 Conv - 2 Pool - 1 asi = 32x32x3 f=5x5 s=1 28x28x6 14x14x6 10x10x16 5x5x16 NN Layer 1 NN Layer 2 NN Layer 3 120x1 84x1 10x1 E.g. :- Layer 1 Layer 2 f =2x2 s=2 maxpool f=5x5 s=1 f =2x2 s=2 maxpool Part 1
16. Conv. formula :- | ((n + 2p -f)/s) +1 | Pool. formula :- | ((n-f)/s)+1 | Parameters for. :- [ {f x f} + 1] x nc asi = Activation Size ash = Activation Shape = (solve asi) Activation Size ↓ Parameters ↑ Part 2
19. What do you mean ? Everything we learned, was a bluff ?
20. List of Transfer Learning models for C.V. :- • • • • • • •
21. Data Augmentations :-
22. Object detection :- Image Classification Classification with localization Detection
23. ... Conv. Layer Softmax Object categories :- - Pedestrian - Car - MotorCycle - Background {0,0} {1,1} bx,by,bh,bw .
24. Object categories :- - Pedestrian - Car - MotorCycle - Background Pc Bx By Bh Bw Cp Cc Cm 0 ? ? ? ? ? ? ? 1 Bx By Bh Bw 0 1 0 y =
25. Yolo Algorithm :- ( You only look once ) Pc Bx By Bh Bw Cp Cc Cm y = 1 Bx By Bh Bw 0 1 0 0 ? ? ? ? ? ? ? Target Output = 3x3x8 4 & 6
27. Intersection Over Union ( IOU ) :- IOU = Size of -------------------- Size of “ Correct ” if IOU ≥ 0.5
28. Anchor Boxes :- Pc Bx By Bh Bw Cp Cc Cm Pc Bx By Bh Bw Cp Cc Cm y = } } } Anchor Box 1 } } } } } } Anchor Box 2 } } } Target Output = 3x3x2x8
30. Implementation on TensorFlow...
31. How to implement :-
33. Image Generator :-
34. How to Use ImageGen. :-
36. Image Augmentation :- Main purpose to use Image Augmentations is to avoid overfiting Is it Overfitting?
37. How to use Image augmentations in T.F. :-
39. Transfer Learning :- Trained model Data Lock The layers Call our Data Fit it ! Step to use T.L. -->
40. Dropouts in Transfer Learning :- Using a transfer learning also results overfitting sometimes. To overcome that we have this “Ram Baan Ellaj”, called Dropouts.
41. How to use dropouts :-
44. I am Done...
45. Questions ?
46. Thanks for your valuable time... LinkedIn :- https://www.linkedin.com/in/manthan-ladva-506a04116/ Twitter :- https://twitter.com/manthan_ladva Github :- https://github.com/manthan-ladva
47. … Manthan Mitaben Ladva Signing Off...