horseorhuman

• I have built three CNN models to classify horse or human from the freely available dataset on tensorflow called "horseorhuman"
• They are built in python using Tensorflow libraries
• This dataset contains 1027 training images and 256 test images.
• Images are of dimensions 300 x 300 x 3
• Image label is "0" for horse "1" for human.
• Three models are described as

model1

• 4 convolution layers, activation is "relu" each conv layer followed by a max pool layer.
• Then 2 fully connected layers.

Training accuracy	Training loss	Test accuracy	Test loss
100 %	4.05e-05	90.23 %	3.8

model2

• This model is inspired from "ResNet50" model

Training accuracy	Training loss	Test accuracy	Test loss
98.8 %	0.03	92.97 %	0.45

• Both are run on Epochs = 100

model1 vs model2

model3

• This is a transfer learning model
• "MobileNetV2" is used.
• Training is set to false
• weights from "imagenet" are used.
• Two additional layers are added in the end.
• With the model frozen, only the newly added additional layers are trained for Epochs = 20.

Training accuracy	Training loss	Test accuracy	Test loss
99.6 %	0.01	100 %	0.003

model3 performance

Remarks

• Model3 standsout
• Models 1 & 2 are overfitting on Training data.

ToDo

• Avoiding overfitting for Models 1 & 2