dropout.py

keep_prob = 0.5 

def train_step(X):
    hidden_layer_1 = np.maximum(0, np.dot(W1, X) + b1)
    dropout_mask_1 = np.random.binomial(1, keep_prob, hidden_layer_1.shape)
    hidden_layer_1 *= dropout_mask_1
    hidden_layer_2 = np.maximum(0, np.dot(W2, hidden_layer_1) + b2)
    dropout_mask_2 = np.random.binomial(1, keep_prob, hidden_layer_2.shape)
    hidden_layer_2 *= dropout_mask_2
    out = np.dot(W3, hidden_layer_2) + b3
  
    # backward pass: compute gradients... (not shown)
    # perform parameter update... (not shown)
  
def predict(X):
    # ensembled forward pass
    hidden_layer_1 = np.maximum(0, np.dot(W1, X) + b1) * keep_prob # NOTE: scale the activations
    hidden_layer_2 = np.maximum(0, np.dot(W2, hidden_layer_1) + b2) * keep_prob # NOTE: scale the activations
    out = np.dot(W3, hidden_layer_2) + b3