def _build_network_lenet(self, x, y, is_training):
"""
Lenet network construction
:param x: tensor
Raw data
:param y: tensor
Labels given
:param is_training: boolean
Whether it is in the training step.
:return:
"""
x_resh = tf.reshape(x, [-1, self._input_width, self._input_height, self._input_channels])
outfc = []
for i in range(self._label_nums): # We train the six labels at the same time
label_name = "label%d_" % i
conv1 = conv_layer(x_resh, 5, 1, 32, is_training, name=label_name + "conv1")
pool1 = avg_pool_layer(conv1, 2, 2, name=label_name + "pool1")
conv2 = conv_layer(pool1, 5, 1, 64, is_training, name=label_name + "conv2")
pool2 = avg_pool_layer(conv2, 2, 2, name=label_name + "pool2")
fc_in = tf.reshape(pool2, [-1, 5 * 5 * 64])
fc3 = fc_layer(fc_in, 1024, is_training, name=label_name + "fc3", relu_flag=True)
dropout3 = tf.nn.dropout(fc3, self._keep_prob)
fc4 = fc_layer(dropout3, self._classes, is_training, name=label_name + "fc4", relu_flag=True)
outfc.append(fc4)
out = tf.reshape(tf.concat(outfc, axis=1), [-1, self._label_nums, self._classes])
loss = weighted_loss(out, y, self._classes, self._loss_array)
accu, precision, recall, f1 = self._get_network_measure(y, out)
return loss, out, accu, precision, recall, f1
def _build_network_alexnet(self, x, y, is_training):
"""
Alexnet network construction.
:param x: tensor
Raw data
:param y: tensor
Labels given
:param is_training: boolean
Whether it is in the training step.
:return:
"""
x_resh = tf.reshape(x, [-1, self._input_width, self._input_height, self._input_channels])
outfc = []
for i in range(self._label_nums): # We train the six labels at the same time
label_name = "label%d_" % i
conv1 = conv_layer(x_resh, 11, 1, 64, is_training, name=label_name + "conv1")
pool1 = avg_pool_layer(conv1, 2, 2, name=label_name + "pool1")
norm1 = norm_layer(pool1, 4)
conv2 = conv_layer(norm1, 5, 1, 192, is_training, name=label_name + "conv2", groups=2)
pool2 = avg_pool_layer(conv2, 2, 2, name=label_name + "pool2")
norm2 = norm_layer(pool2, 4)
conv3 = conv_layer(norm2, 3, 1, 384, is_training, name=label_name + "conv3")
conv4 = conv_layer(conv3, 3, 1, 384, is_training, name=label_name + "conv4")
conv5 = conv_layer(conv4, 3, 1, 256, is_training, name=label_name + "conv5")
pool5 = avg_pool_layer(conv5, 2, 2, name=label_name + "pool5")
fc_in = tf.reshape(pool5, [-1, 3 * 3 * 256])
fc6 = fc_layer(fc_in, 4096, is_training, name=label_name + "fc6", relu_flag=True)
dropout6 = tf.nn.dropout(fc6, self._keep_prob)
fc7 = fc_layer(dropout6, 4096, is_training, name=label_name + "fc7", relu_flag=True)
dropout7 = tf.nn.dropout(fc7, self._keep_prob)
fc8 = fc_layer(dropout7, self._classes, is_training, name=label_name + "fc8", relu_flag=True)
outfc.append(fc8)
out = tf.reshape(tf.concat(outfc, axis=1), [-1, self._label_nums, self._classes])
loss = weighted_loss(out, y, self._classes, self._loss_array)
accu, precision, recall, f1 = self._get_network_measure(y, out)
return loss, out, accu, precision, recall, f1