ajithvcoder/getsetgo_keras-beginner

Pretrained Convolutional Neural Network models(CNN) (Imagenet challange ILSVRC) are used to predict input image .Codes to extract specific layers ,fine tune and visualize those models are also provided.

Using pretrained Convolutional Neural Network models(CNN) in Keras

Credits

-https://mlblr.com/
-https://keras.io/
-https://colab.research.google.com/

codes used here are executed and tested only in colab.research.google.com please make sure you have enabled GPU access from notebook settings for faster execution process

CNN models

Models used here are pretrained on ImageNet dataset ImageNet Large Scale Visual Recognition Challenge (ILSVRC) with python and keras deep learning library.

1.Resnet50 model

• Classifying input image with probabilities notebook link
  
• Visualizing resnet50 model notebook link
  

2.VGG16 model

• Classifying input image with probabilities notebook link
• Visuvalizing vgg16 model notebook link

3.VGG19 model

• Classifying input image with probabilities notebook link
• Extracting last layer of vgg19 and using them to predict notebook link
• Visuvalizing model extracted model notebook link

4.InceptionV3 model

• Classifying input image with probabilities notebook link
• Extracting last layer of vgg19 and using them to predict notebook link
• Visuvalizing model extracted model notebook link

What is a pretrained model?
Models that has been already trained on a particular data set with number of classes

Use of pretrained model: Instead of training from scratch/starting with random weight initialiation pre-trained models can be used with other datasets. [keywords: fine tuning-transfer learning]

Model info:

Resnet50 model Architecture speciality:Use of residual blocks(skip connections) enabled training much deeper network by handling vanishing and exploding gradient problems https://arxiv.org/abs/1512.03385

VGG16 model Architecture speciality: Use of 3x3 convolution filters with depth of 16 weight layers https://arxiv.org/abs/1409.1556

VGG19 model Architecture speciality: Use of 3x3 convolution filters with depth of 19 weight layers https://arxiv.org/abs/1409.1556

InceptionV3 model Architecture speciality:1x1,3x3,5x5,max pooling convolutions performed with less computation cost-sparsely deep connected network-Hebbian principle(neurons that fire together wire together) https://arxiv.org/pdf/1409.4842.pdf