rajendra7406/dogsandcats

Classifying dogs and cats using CNN. And many more.

Dogs and Cat classifier.

Along with cloning this repository. Clone this repository https://github.com/fchollet/deep-learning-models.

Dependencies:

• keras
• tensorflow/theano
• numpy
• matplotlib
• bcolz
• pillow
• h5

Files:

1. image classifier in 7 step . ipynb :

   • Predicts image offline.
   • Image must be one of 1000 categories in imagenet dataset.

2. CNN - 1. ipynb:

   • Notebook contains:

     1. Simple convolutional architecture:
        • 1 convolution layer
        • 1 maxpooling layer
        • 1 flatten layer
        • 1 dense layer

   • Ran on small dataset.

   • Saved the architecture in h5 format.

3. CNN - 2 . ipynb:

   • Loaded the architecture.
   • Predicting the image.

4. Vgg architecture from scratch . ipynb:

   • Created vgg-16 cnn architecture from scratch.
   • Architecture contains:
     1. convolution layers with zero padding layer. ( several convolution layers with varying no of filters. Maxpooling layer in between)
     2. Flatten layer
     3. 2 Dense layers.
     4. 3rd dense layer with 1 ouput for cats and dogs ( in original there are 1000 outputs )
   • Tried to saved the model.

5. Advanced vgg architecture . ipynb:

   • Splitting the convolutional computation part & remaining part of the model.
   • used bcolz for faster run time.
   • Reduced underfitting:
     • Removal/reduced dropout.
   • Reduced overfitting:
     • Data Augmentation.
     • Batch Normalization.

6. Fine tuning and Back Propagation:

   • Fine tuning is to modify existing model and create a secondary model.
   • Doing Back Propagation in few lines