def make_full_connected_layer(self, key, input, weight_shape, bias_shape):
with tf.name_scope(key):
# Weights
if 'type' in self.config[key]['weight']:
weight_type = self.config[key]['weight']['type']
else:
weight_type = 'random_normal'
if 'name' in self.config[key]['weight']:
weight_name = self.config[key]['weight']['name']
else:
weight_name = 'Weight_' + key
self.model_params['weights'][key] = \
Weights.define(weight_shape, weight_type=weight_type, weight_name=weight_name)
# Bias
if 'type' in self.config[key]['bias']:
bias_type = self.config[key]['bias']['type']
else:
bias_type = 'random_normal'
if 'name' in self.config[key]['bias']:
bias_name = self.config[key]['bias']['name']
else:
bias_name = 'Bias_' + key
self.model_params['bias'][key] = \
Bias.define(bias_shape, bias_type=bias_type, bias_name=bias_name)
if 'name' in self.config[key]:
layer_name = self.config[key]['name']
else:
layer_name = key
max_pool_layer = mlp_layer(input,
self.model_params['weights'][key],
self.model_params['bias'][key],
layer_name=layer_name)
self.model_params['layers'][key] = max_pool_layer
def make_predictions(self):
with tf.device(self.device):
with tf.name_scope('Predictions'):
value = mlp_layer(self.predict_params['input'], self.model_params['weight'],
self.model_params['bias'],
activation_type=self.config['activation_fn'])
self.params['logits'] = tf.nn.softmax(value)
self.predict_params['predict_class'] = \
tf.argmax(self.params['logits'], dimension=1, name='predict_class')
self.predict_params['predict_one_hot'] = \
tf.one_hot(self.predict_params['predict_class'],
depth=self.num_classes, on_value=1.0,
off_value=0.0, axis=-1)
def make_output_layer(self):
prev_layer_key = 'layer_' + str(self.num_layers)
layer_key = 'output_layer'
prev_layer = self.model_params['layers'][prev_layer_key]
if 'name' in self.config[layer_key]:
layer_name = self.config[layer_key]['name']
else:
layer_name = 'Output_Layer'
if 'activation_fn' in self.config[layer_key]:
activation_fn = self.config[layer_key]['activation_fn']
else:
activation_fn = 'sigmoid'
if prev_layer_key in self.config:
in_units = self.config[prev_layer_key]['num_nodes']
else:
in_units = self.num_features
weight_shape = (in_units, self.num_classes)
bias_shape = (self.num_classes, )
if 'type' in self.config[layer_key]['weight']:
weight_type = self.config[layer_key]['weight']['type']
else:
weight_type = 'random_normal'
if 'name' in self.config[layer_key]['weight']:
weight_name = self.config[layer_key]['weight']['name']
else:
weight_name = 'Weight_output'
if 'type' in self.config[layer_key]['bias']:
bias_type = self.config[layer_key]['bias']['type']
else:
bias_type = 'random_normal'
if 'name' in self.config[layer_key]['bias']:
bias_name = self.config[layer_key]['bias']['name']
else:
bias_name = 'Bias_output'
weight = Weights.define(weight_shape,
weight_type=weight_type,
weight_name=weight_name)
bias = Bias.define(bias_shape,
bias_type=bias_type,
bias_name=bias_name)
self.model_params['weights']['output_layer'] = weight
self.model_params['bias']['output_layer'] = bias
self.model_params['layers']['output_layer'] = \
mlp_layer(prev_layer, weight, bias, activation_type=activation_fn, layer_name=layer_name)
def create_graph(self, num_classes, num_features):
start = time.time()
self.num_features = num_features
self.num_classes = num_classes
self.graph = tf.Graph()
# Step 1: Creating placeholders for inputs
self.make_placeholders_for_inputs(self.graph)
# Step 2: Creating initial parameters for the variables
self.make_parameters(self.graph)
# Step 3: Make predictions for the data
final_layer = mlp_layer(self.predict_params['input'], self.model_params['weight'],
self.model_params['bias'],
activation_type=self.config['activation_fn'])
self.make_predictions(self.graph, final_layer, self.num_classes)
# Step 4: Perform optimization operation
self.make_optimization(self.graph, final_layer)
# Step 5: Calculate accuracies
self.make_accuracy(self.graph)
self.start_session(self.graph)
end = time.time()
print('Graph completed')
def make_layers(self, number_of_layers=1):
prev_layer = self.predict_params['input']
for layer_no in range(1, number_of_layers + 1):
weight_key = 'weight_' + str(layer_no)
bias_key = 'bias_' + str(layer_no)
layer_key = 'layer_' + str(layer_no)
if 'name' in self.config[layer_key]:
layer_name = self.config[layer_key]['name']
else:
layer_name = 'Layer_' + str(layer_no)
if 'activation_fn' in self.config[layer_key]:
activation_fn = self.config[layer_key]['activation_fn']
else:
activation_fn = 'sigmoid'
weight = self.model_params['weights'][weight_key]
bias = self.model_params['bias'][bias_key]
layer = mlp_layer(prev_layer,
weight,
bias,
layer_name=layer_name,
activation_type=activation_fn)
self.model_params['layers'][layer_key] = layer
prev_layer = self.model_params['layers'][layer_key]