def __init__(self,
input_dim,
output_dim,
optimizer,
activation=(utils.sigmoid, utils.sigmoid_prime)):
print("Instatiating Vanilla Recurrent")
NeuralLayer.__init__(self, input_dim, output_dim)
self.Wxz = np.random.randn(output_dim[0], input_dim[0]) * 0.01
self.Wyz = np.random.randn(output_dim[0], output_dim[0]) * 0.01
self.bz = np.zeros((output_dim[0], 1))
self.act_fxn = activation[0]
self.act_prime = activation[1]
self.optimizer = optimizer
# Stored values for single samples: This way, can preserve state for next single sample
# A convenience for testing (generating sentences from a single word)
self.y = np.zeros((output_dim[0], 1))
# Stored values for batch back propagation through time and batch updates
self.seq_x = None
self.seq_z = None
self.seq_y = None
def __init__(self, input_dim, pool_dim):
print("instantiating Max Pool")
self.pool_dim = pool_dim
output_dim = [
input_dim[0] / pool_dim[0], input_dim[1] / pool_dim[1],
input_dim[2] / pool_dim[2]
]
NeuralLayer.__init__(self, input_dim, output_dim)
# Stored values for back propagation and updates
self.batch_max_binary_filter = None
def __init__(self,
input_dim,
output_dim,
optimizer,
activation=(utils.sigmoid, utils.sigmoid_prime)):
print("instantiating Vanilla Feed Forward")
NeuralLayer.__init__(self, input_dim, output_dim)
self.input_dim, self.output_dim = input_dim, output_dim
self.Wxz = np.random.randn(output_dim[0], input_dim[0]) / np.sqrt(
input_dim[0])
self.bz = np.zeros(output_dim)
self.act_fxn = activation[0]
self.act_prime = activation[1]
self.optimizer = optimizer
# Stored values for batch back propagation and batch updates
self.batch_x = None
self.batch_z = None
def __init__(self, input_dim, num_patterns, filter_dim, optimizer, activation=(utils.softplus, utils.softplus_prime)):
print("instantiating Convolutional")
self.num_patterns = num_patterns
self.x_depth, self.x_height, self.x_width = input_dim[0], input_dim[1], input_dim[2]
self.w_depth, self.w_height, self.w_width = filter_dim[0], filter_dim[1], filter_dim[2]
self.z_depth, self.z_height, self.z_width \
= self.num_patterns, self.x_height-self.w_height+1, self.x_width-self.w_width+1
NeuralLayer.__init__(self, input_dim, [self.z_depth, self.z_height, self.z_width])
self.Wxz = np.random.randn(self.num_patterns, self.w_depth, self.w_height, self.w_width)
self.bz = np.zeros(self.num_patterns)
self.act_fxn = activation[0]
self.act_prime = activation[1]
self.optimizer = optimizer
# Stored values for back propagation and batch updates
self.batch_x = None
self.batch_z = None
def __init__(self, input_dim, output_dim):
NeuralLayer.__init__(self, input_dim, output_dim)
print("instantiating SoftMax")