ftZHOU / deepLearnPraticals

Tensorflow, tensorboard, keras....

Deep learning praticals

1. First step in Keras: classifying handwritten digits (MNIST)

The usage of Tensorboard/ ModelCheckpoint for training and saving different models

1. One input layer + one Dense layer + one output layer(softmax)
   Corssentropy loss + Adam optimizer + no normalization

2. Same model with normalization (lamda layer)
   

3. Two hidden layers (Dense+relu) + One output layer (softmax)

4. add regularizer (in the base of the model below)

   1. L2 penality x = Dense(256,kernel_regularizer=regularizers.l2(1e-5))(xl)
      Test loss: 0.158592546591 Test accuracy: 0.9762
   2. Dropout x = Dense(hidden1)(x) x = Activation('relu')(x) x = Dropout(0.5)(x) Test loss: 0.0925885819901 Test accuracy: 0.9746
      

Reference link

A Friendly Introduction to Cross-Entropy Loss