def create_layers(self):
self.layer_conv1, self.weights_conv1 = new_conv_layer(
prev_layer=self.X_img,
in_channels=self.num_channels,
out_channels=self.num_filters1,
filter_size=self.conv_filter_size1,
name='conv-1')
self.layer_conv2, self.weights_conv2 = new_conv_layer(
prev_layer=self.layer_conv1,
in_channels=self.num_filters1,
out_channels=self.num_filters2,
filter_size=self.conv_filter_size2,
name='conv-2')
self.flattened, self.num_features = new_flatten_layer(
prev_layer=self.layer_conv2)
self.fc_layer1 = new_fc_layer(prev_layer=self.flattened,
num_inputs=self.num_features,
num_outputs=self.fc_size,
name='fc-1',
relu=True)
self.fc_layer2 = new_fc_layer(prev_layer=self.fc_layer1,
num_inputs=self.fc_size,
num_outputs=self.num_classes,
name='fc-2')
self.prediction = tf.nn.softmax(self.fc_layer2)
self.y_pred_cls = tf.argmax(self.prediction, axis=1)
def configure_network1(dataset, input_dim):
global x, y_true, num_classes
num_features = int(input_dim)
fc_size1 = int(1024)
print("num_classes {} num_features {}".format(num_classes, num_features))
layer_fc1 = new_fc_layer(input=dataset,
num_inputs=num_features,
num_outputs=fc_size1,
use_relu=False)
layer_fc2 = new_fc_layer(input=layer_fc1,
num_inputs=fc_size1,
num_outputs=num_classes,
use_relu=False)
y_pred = tf.nn.softmax(layer_fc2)
print(y_pred)
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=layer_fc2,
labels=y_true)
cost = tf.reduce_mean(cross_entropy)
return y_pred, cost
def create_layers(self):
self.layer_conv1, self.weights_conv1 = new_conv_layer(
prev_layer=self.X_img,
in_channels=self.num_channels,
out_channels=self.num_filters1,
filter_size=self.conv_filter_size1,
name=self.name + 'conv-1')
self.layer_conv2, self.weights_conv2 = new_conv_layer(
prev_layer=self.layer_conv1,
in_channels=self.num_filters1,
out_channels=self.num_filters2,
filter_size=self.conv_filter_size2,
name=self.name + 'conv-2')
self.flattened, self.num_features = new_flatten_layer(
prev_layer=self.layer_conv2)
self.fc_layer1 = new_fc_layer(prev_layer=self.flattened,
num_inputs=self.num_features,
num_outputs=self.fc_size,
name=self.name + 'fc-1',
relu=True)