Densenet_SVHN

A tensorflow implementation of dense net on Street view house number(SVHN) dataset.

Model

A brief description of the Model is provided below.

The files I have modified significantly are:

1. preprocess.py

2. train.py

3. models.py

High level view

Input layer -> Block 1 -> Transition 1 -> Block 2 -> Transition 2 -> Block 3 -> Batch Normalization -> Relu -> Global average pooling -> Fully connected layer

• Block consists of 4 Dense layers.

• Dense layer is made of the following sequence:

1. Batch Normalization

2. Relu

3. Convolutional 2d layer

4. Concatination of the previous layers output to the previous element(Convolutional 2d layer)

• Transition Layer is made of the following sequence:

1. Batch Normalization

2. Relu

3. Convolutional 2d layer

4. Average Pooling(stride=2)

Hyper parameters and other essential attributes

• Input dimension = [100, 32, 32, 1] (Trained using batches of 100 images)

• Ouput dimension = [10]

• epoch = 800

Preprocessing

• Balanced subsampling on training dataset.

• Converting SVHN images from RGB to grayscale.

• The training and validation data are stored in HDF5 binary data format.