The code is for a plain neural network built on Tensorflow to classify handsigns. Architecture of the model is as follows :
CONV2D -> RELU -> MAXPOOL -> CONV2D -> RELU -> MAXPOOL -> FLATTEN -> FULLYCONNECTED
Cost function is minimised using Softmax Cross Entropy with logits (https://www.tensorflow.org/api_docs/python/tf/nn/softmax_cross_entropy_with_logits) Accuracy on training set : 94.3 %. Can be improved by increasing number of epochs. (Hardware constraints)
The same hand signs dataset is trained on a deep 50 layer neural network.
Identity block: CONV2D -> BATCHNORM -> RELU -> CONV2D -> BATCHNORM + RELU where another CONV2D -> BATCHNORM block skips the first two and gets added at the +
Full Network: CONV2D -> BATCHNORM -> RELU -> MAXPOOL -> CONVBLOCK -> IDBLOCK2 -> CONVBLOCK -> IDBLOCK3 -> CONVBLOCK -> IDBLOCK5 -> CONVBLOCK -> IDBLOCK2 -> AVGPOOL -> TOPLAYER
Some neural network classifiers for the MNIST and Fashion MNIST datasets. Experimented with regular neural network in Tensorflow as well as a Convolutional Neural Network to explore accuracy improvement.
Generating new Dinosaur names from a list of existing names using a character level language model. Gradient clipping used inside the optimization function to prevent exploding gradients. Sampling carries out generation of new words.