diff --git a/main.py b/main.py
index df040c83613c6297bac6dc91ccc42a89eb66ea6a..6aa6b99442be5c9c664cc3b024d197ceaa14c040 100644
--- a/main.py
+++ b/main.py
@@ -17,8 +17,8 @@ logger = logging.get_logger(__name__)
# Load in CSV.
housing_data = pandas.read_csv('./Documents/other_housing.csv')
-# Initially only work with first 10, for testing purposes.
-housing_data = housing_data[0:3]
+# Initially only work with first few, for testing purposes.
+# housing_data = housing_data[0:3]
# Normalize data.
logger.info('')
@@ -32,5 +32,5 @@ logger.info('')
# logger.info('Normalized Targets: \n{0}'.format(targets))
# Start neural net.
-backprop = neural_net.BackPropNet(normalized_data.values)
+backprop = neural_net.BackPropNet(features)
backprop.train(features.values, targets.values)
diff --git a/neural_net.py b/neural_net.py
index 34ddec825f6a3015a9b2a1ee101f00269f65b76e..f975db463a0714eb2ccd83a6425fa168528f3265 100644
--- a/neural_net.py
+++ b/neural_net.py
@@ -130,10 +130,19 @@ class Normalizer():
for column in ignored_columns:
orig_data = orig_data.loc[:, orig_data.columns != column]
- # Target Data. Should this be normalized as well?
- # Add column to normalized data list.
- frame = pandas.DataFrame(orig_data[target_column])
- normalized_data = normalized_data.join(frame, how='outer')
+ # Process target data. Handles same as continous.
+ for column in target_column:
+ if column in orig_data.columns:
+ # Normalize.
+ self.squish_values(orig_data, column)
+
+ # Remove NaN references.
+ orig_data[column] = orig_data[column].fillna(value=0)
+
+ # Add column to normalized data list.
+ frame = pandas.DataFrame(orig_data[column])
+ normalized_data = normalized_data.join(frame, how='outer')
+ orig_data = orig_data.loc[:, orig_data.columns != column]
return normalized_data
@@ -181,6 +190,7 @@ class BackPropNet():
def __init__(self, data):
self.hidden_layer_size = 3
self.network = []
+ self.layer_inputs = None
self._create_architecture(data)
self.learning_rate = 0.001
@@ -192,11 +202,12 @@ class BackPropNet():
Weight values are randomized values near 0, using a normal distribution.
:param data: Data to reference for input layer count.
"""
+ logger.info('Columns: {0}'.format(data.columns))
# Create first hidden layer.
hidden_layer_1 = []
for index in range(self.hidden_layer_size):
hidden_layer_1.append([
- (numpy.random.randn() * 0.001) for index in range(len(data) + 1)
+ (numpy.random.randn() * 0.001) for index in range(len(data.columns) + 1)
])
# Create second hidden layer.
@@ -216,10 +227,10 @@ class BackPropNet():
self.network.append(hidden_layer_2)
self.network.append(output_layer)
- # logger.info('Network:')
+ logger.info('Network:')
index = 0
for layer in self.network:
- # logger.info('Layer {0}: {1}'.format(index, layer))
+ logger.info('Layer {0}: {1}'.format(index, layer))
index += 1
def _activation(self, weights, inputs):
@@ -239,23 +250,15 @@ class BackPropNet():
def _reverse_activation(self, weights, outputs):
"""
- Calculates reverse of initial values, prior to neuron firing.
+ Calculates derivative and adjusts weight accordingly.
:param weights: Weights of the given layer.
:param inputs: Previously calculated output.
- :return:
"""
- inputs = []
- # logger.info('Weights: {0}'.format(weights))
for output in outputs:
- # logger.info('Output to reverse: {0}'.format(output))
for index in range(len(weights) - 1):
- # logger.info('Output: {0}'.format(output[0]))
pre_sigmoid_value = self._reverse_sigmoid(output[0])
- output_rev = (weights[index] * pre_sigmoid_value)
- input = output[0] - output_rev
- inputs.append(input)
- # How do you calculate this and where does weight updating come in?
- return inputs
+ derivative = (weights[index] * pre_sigmoid_value)
+ weights[index] -= (derivative * self.learning_rate)
def _sigmoid(self, value):
"""
@@ -273,7 +276,7 @@ class BackPropNet():
"""
return ( self._sigmoid(value) * ( 1 - self._sigmoid(value) ) )
- def _forward_propagate(self, inputs):
+ def _forward_propagate(self, inputs, training=False):
"""
Walk forward through the neural network.
:param inputs: Initial inputs for network.
@@ -283,6 +286,8 @@ class BackPropNet():
# Iterate through each value in network, using previous outputs as new inputs.
for index in range(len(self.network)):
outputs = []
+ if training:
+ self.layer_inputs.append(inputs)
for neuron in self.network[index]:
outputs.append(self._activation(neuron, inputs))
inputs = outputs
@@ -292,17 +297,13 @@ class BackPropNet():
"""
Walk backward through the neural network, using derivatives.
:param inputs: Original output of network.
- :return: ???
"""
# Iterate backwards through network.
outputs = prediction
- inputs = None
for index in reversed(range(len(self.network))):
- inputs = []
for neuron in self.network[index]:
- inputs.append(self._reverse_activation(neuron, outputs))
- outputs = inputs
- return inputs
+ self._reverse_activation(neuron, outputs)
+ # TODO: Where does delta come in?
def _calculate_delta(self, prediction, targets):
"""
@@ -311,7 +312,7 @@ class BackPropNet():
:param targets: Desired prediction.
:return: Delta of error difference.
"""
- return ( (targets - prediction) ** 2)
+ return ( (targets - prediction) ** 2) # TODO: Is this ever used?
def train(self, features, targets):
"""
@@ -320,12 +321,15 @@ class BackPropNet():
"""
logger.info('Initial Inputs: {0}'.format(features))
prediction = []
+ self.layer_inputs = []
for index in range(len(features)):
- prediction.append(self._forward_propagate(features[index]))
+ prediction.append(self._forward_propagate(features[index], training=True))
+ # logger.info('Layer Inputs: {0}'.format(self.layer_inputs))
delta_error = self._calculate_delta(prediction, targets)
- logger.info('Delta Error: {0}'.format(delta_error))
- back_prop_result = self._backward_propagate(features, targets, prediction, delta_error)
- logger.info('Backprop Result: {0}'.format(back_prop_result))
+ # logger.info('Delta Error: {0}'.format(delta_error))
+ self._backward_propagate(features, targets, prediction, delta_error)
+ logger.info('Expected targets: {0}'.format(targets))
+ logger.info('Predicted targets: {0}'.format(prediction))
def predict(self, data):
"""