Cat-Dog-Classifier:

Cat-dog classifications using a deep neural network.

Network Architecture:

-	Layer	I/O Shape
1	Conv2d_1	(150, 150, 3) / (148, 148, 32)
2	ReLU_1	(148, 148, 32) / (148, 148, 32)
3	MaxPooling2d_1	(148, 148, 32) / (74, 74, 32)
4	Conv2d_2	(74, 74, 32) / (73, 73, 64)
5	ReLU_2	(73, 73, 64) / (73, 73, 64)
6	MaxPooling2d_2	(73, 73, 64) / (36, 36, 64)
7	Conv2d_3	(36, 36, 64) / (34, 34, 128)
8	ReLU_3	(34, 34, 128) / (34, 34, 128)
9	MaxPooling2d_3	(34, 34, 128) / (17, 17, 128)
10	Conv2d_4	(17, 17, 128) / (16, 16, 256)
11	ReLU_4	(16, 16, 256) / (16, 16, 256)
12	MaxPooling2d_3	(16, 16, 256) / (8, 8, 256)
13	Flatten	(8, 8, 256) / 16384
14	Dropout_1	16384 / 16384
15	Dense_1	16384 / 500
16	ReLU_5	500 / 500
17	Dropout_2	500 / 500
18	Dense_2	500 / 1
19	Sigmoid	1 / 1

Preprocessing:

Images are rescaled to be 150 * 150 pixels images.
The color range is re-scaled from [0, 255] to be in range of [0, 1].

Data augmentation:

Images go through random flipping, height/width shifting, zooming transformations.

Training:

The model trained for 15 epochs with path size of 128. Nadam is used as the optimizer with 0.001 learning rate and binary cross entropy as the loss function.

Results:

The model successfully classified 88% of the test data set.