def _build_net(self):
w2 = helpers.weight_variable([ENCODED_FEATURES, HIDDEN_UNITS_2])
b2 = helpers.bias_variable([HIDDEN_UNITS_2])
l2 = tf.nn.relu6(tf.matmul(self.x_placeholder, w2) + b2)
l2_drop = tf.nn.dropout(l2, self.keep_prob)
# Output layer
w5 = helpers.weight_variable([HIDDEN_UNITS_2, OUTPUT_CLASSES])
b5 = helpers.bias_variable([OUTPUT_CLASSES])
l5 = tf.nn.relu6(tf.matmul(l2_drop, w5) + b5)
return tf.nn.softmax(l5)
def _build(self):
# First hidden layer
w2 = helpers.weight_variable([INPUT_UNITS, HIDDEN_UNITS_1])
b2 = helpers.bias_variable([HIDDEN_UNITS_1])
l2 = tf.nn.relu(tf.matmul(self.x_placeholder, w2) + b2)
l2_drop = tf.nn.dropout(l2, self.keep_prob)
# Second hidden layer
w3 = helpers.weight_variable([HIDDEN_UNITS_1, HIDDEN_UNITS_2])
b3 = helpers.bias_variable([HIDDEN_UNITS_2])
l3 = tf.nn.relu(tf.matmul(l2_drop, w3) + b3)
l3_drop = tf.nn.dropout(l3, self.keep_prob)
# Output layer
w5 = helpers.weight_variable([HIDDEN_UNITS_2, OUTPUT_CLASSES])
b5 = helpers.bias_variable([OUTPUT_CLASSES])
return tf.nn.softmax(tf.matmul(l3_drop, w5) + b5)
def _build(self):
# First hidden layer
w2 = helpers.weight_variable([INPUT_UNITS, HIDDEN_UNITS_1])
b2 = helpers.bias_variable([HIDDEN_UNITS_1])
l2 = tf.nn.relu(tf.matmul(self.x_placeholder, w2) + b2)
l2_drop = tf.nn.dropout(l2, self.keep_prob)
# Second hidden layer
w3 = helpers.weight_variable([HIDDEN_UNITS_1, HIDDEN_UNITS_2])
b3 = helpers.bias_variable([HIDDEN_UNITS_2])
l3 = tf.nn.relu(tf.matmul(l2_drop, w3) + b3)
l3_drop = tf.nn.dropout(l3, self.keep_prob)
# Output layer
w5 = helpers.weight_variable([HIDDEN_UNITS_2, OUTPUT_CLASSES])
b5 = helpers.bias_variable([OUTPUT_CLASSES])
return tf.nn.softmax(tf.matmul(l3_drop, w5) + b5)
def loss(deep_features):
with tf.name_scope('softmax_loss'):
batch_labels = tf.placeholder(tf.float32, name='labels')
W_loss = helpers.weight_variable([2, 10], "W_loss")
bias_loss = tf.Variable(
tf.truncated_normal(shape=[10], stddev=1e-4, mean=1e-1),
'bias_loss')
# Note: we don't use the bias here because it does not affect things. removing the
# bias also makes the analysis simpler.
logits = tf.matmul(deep_features, W_loss) + bias_loss
cross_entropy = -tf.reduce_mean(tf.mul(batch_labels,
tf.nn.log_softmax(logits)),
reduction_indices=[1])
xentropy_mean = tf.reduce_mean(cross_entropy, name="xentropy_mean")
tf.scalar_summary(xentropy_mean.op.name, xentropy_mean)
return batch_labels, logits, xentropy_mean
def _build_decoder(self, encoder):
weights = helpers.weight_variable([ENCODED_FEATURES, INPUT_UNITS])
biases = helpers.bias_variable([INPUT_UNITS])
l3 = tf.nn.relu(tf.matmul(encoder, weights) + biases)
return tf.nn.softmax(l3)
def _build_encoder(self):
weights = helpers.weight_variable([INPUT_UNITS, ENCODED_FEATURES])
biases = helpers.bias_variable([ENCODED_FEATURES])
return tf.nn.relu(tf.matmul(self.x_placeholder, weights) + biases)
def inference():
x = tf.placeholder(tf.float32, shape=[None, 784], name='input')
image = tf.reshape(x, [-1, 28, 28, 1])
with tf.name_scope('conv_layer_1'):
W_conv1 = helpers.weight_variable([5, 5, 1, 32], 'W_conv1')
b_conv1 = helpers.bias_variable([32], 'bias_conv1')
alphas_conv1 = helpers.bias_variable([32], 'alpha_conv1')
layer_conv_1 = helpers.prelu(
helpers.conv2d(image, W_conv1) + b_conv1, alphas_conv1)
W_conv1_b = helpers.weight_variable([5, 5, 32, 32], 'W_conv1_b')
b_conv1_b = helpers.bias_variable([32], 'bias_conv1_b')
alphas_conv1_b = helpers.bias_variable([32], 'alpha_conv1_b')
layer_conv_1_b = helpers.prelu(
helpers.conv2d(layer_conv_1, W_conv1_b) + b_conv1_b,
alphas_conv1_b)
stage_1_pool = helpers.max_pool_2x2(layer_conv_1_b)
with tf.name_scope('conv_layer_2'):
W_conv2 = helpers.weight_variable([5, 5, 32, 64], "W_conv2")
b_conv2 = helpers.bias_variable([64], 'bias_conv2')
alphas_conv2 = helpers.bias_variable([64], 'alpha_conv2')
layer_conv_2 = helpers.prelu(
helpers.conv2d(stage_1_pool, W_conv2) + b_conv2, alphas_conv2)
W_conv2_b = helpers.weight_variable([5, 5, 64, 64], "W_conv2_b")
b_conv2_b = helpers.bias_variable([64], 'bias_conv2_b')
alphas_conv2_b = helpers.bias_variable([64], 'alpha_conv2_b')
layer_conv_2_b = helpers.prelu(
helpers.conv2d(layer_conv_2, W_conv2_b) + b_conv2_b,
alphas_conv2_b)
stage_2_pool = helpers.max_pool_2x2(layer_conv_2_b)
# stage_2_pool_flat = tf.reshape(stage_2_pool, [-1, 7 * 7 * 64])
with tf.name_scope('conv_layer_3'):
W_conv3 = helpers.weight_variable([5, 5, 64, 128], "W_conv3")
b_conv3 = helpers.bias_variable([128], 'bias_conv3')
alphas_conv3 = helpers.bias_variable([128], 'alpha_conv3')
layer_conv_3 = helpers.prelu(
helpers.conv2d(stage_2_pool, W_conv3) + b_conv3, alphas_conv3)
# stage_3_pool = helpers.max_pool_2x2(layer_conv_3)
# stage_3_pool_flat = tf.reshape(stage_3_pool, [-1, 4 * 4 * 256])
W_conv3_b = helpers.weight_variable([5, 5, 128, 128], "W_conv3_b")
b_conv3_b = helpers.bias_variable([128], 'bias_conv3_b')
alphas_conv3_b = helpers.bias_variable([128], 'alpha_conv3_b')
layer_conv_3_b = helpers.prelu(
helpers.conv2d(layer_conv_3, W_conv3_b) + b_conv3_b,
alphas_conv3_b)
stage_3_pool = helpers.max_pool_2x2(layer_conv_3_b)
stage_3_pool_flat = tf.reshape(stage_3_pool, [-1, 4 * 4 * 128])
with tf.name_scope('fc_layer_1'):
W_fc1 = helpers.weight_variable([4 * 4 * 128, 2], "W_fc1")
# W_fc1 = helpers.weight_variable([7 * 7 * 64, 2], "W_fc1")
b_fc1 = helpers.bias_variable([2], 'bias_fc1')
alphas_fc1 = helpers.bias_variable([2], 'alpha_conv3')
output = helpers.prelu(
tf.matmul(stage_3_pool_flat, W_fc1) + b_fc1, alphas_fc1)
# with tf.name_scope('fc_output'):
# W_output = helpers.weight_variable([500, 10], "W_putput")
# b_output = helpers.bias_variable([10], 'bias_output')
# output = tf.nn.relu(tf.matmul(h_fc1, W_output) + b_output)
# with tf.name_scope('output'):
# W_output = helpers.weight_variable([2, 10], "W_output")
# b_output = helpers.bias_variable([10])
# output = tf.nn.relu(tf.matmul(h_fc2, W_output) + b_output)
return x, output
