README.md 1.7 KB
Newer Older
Syed.Samad Shakeel's avatar
Syed.Samad Shakeel committed
1
2
3
4
5
6
7
8
9
This is a test task where I was asked to add a custom openCV filter in convNet and see the difference in results. Also, I was asked to show a gradcam view of the model.

I decided to add laplacian filters.

The gradcam images have been added in the repository.

Given below was my convNet model summary after adding the filter which is at "conv2d_15":-
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
Syed.Samad Shakeel's avatar
Syed.Samad Shakeel committed
10
11

_________________________________________________________________
Syed.Samad Shakeel's avatar
Syed.Samad Shakeel committed
12
13
14
15
16
17
18
19
20
21
22
23
24
25
conv2d_13 (Conv2D)           (None, 26, 26, 32)        320       
_________________________________________________________________
conv2d_14 (Conv2D)           (None, 24, 24, 64)        18496     
_________________________________________________________________
max_pooling2d_7 (MaxPooling2 (None, 12, 12, 64)        0         
_________________________________________________________________
conv2d_15 (Conv2D)           (None, 5, 5, 1)           577       
(Custom Filter)       
_________________________________________________________________
flatten_7 (Flatten)          (None, 25)                0         
_________________________________________________________________
dense_13 (Dense)             (None, 128)               3328      
_________________________________________________________________
dense_14 (Dense)             (None, 10)                1290      
Syed.Samad Shakeel's avatar
Syed.Samad Shakeel committed
26
_________________________________________________________________
Syed.Samad Shakeel's avatar
Syed.Samad Shakeel committed
27

Syed.Samad Shakeel's avatar
Syed.Samad Shakeel committed
28
29
30
31
32
33
Total params: 24,011
Trainable params: 23,434
Non-trainable params: 577

The validation accuracy of the model without the added filters was 98%.
But the validation accuracy of the model after adding the filters was 79.5% which is not good.