def layer_perceptron(X, input_shape, layer_width, layer_name=None):
with tf.name_scope("weights"):
W = tf.Variable(tf.random_normal(input_shape + layer_width))
ts.variable_summaries(W)
with tf.name_scope("bias"):
bias = tf.Variable(tf.random_normal(layer_width))
ts.variable_summaries(bias)
return tf.sigmoid(tf.matmul(X, W) + bias)
def layer_perceptron(X, input_shape, layer_width, layer_name=None):
if layer_name is None:
layer_name = "layer"
with tf.name_scope(layer_name):
with tf.name_scope("weights"):
W = tf.Variable(tf.random_normal(input_shape + layer_width))
ts.variable_summaries(W)
with tf.name_scope("bias"):
bias = tf.Variable(tf.random_normal(layer_width))
ts.variable_summaries(bias)
with tf.name_scope("activate_function"):
activate = tf.sigmoid(tf.matmul(X, W) + bias)
return activate
def model_NN_softmax():
# placeHolder
ph_set = nn.placeholders_init()
# NN layer
layer1 = nn.layer_perceptron(ph_set["X"], [FLAGS.image_size],
[FLAGS.perceptron_output_shape_size],
"softmax_L1")
layer2 = nn.layer_perceptron(layer1, [FLAGS.perceptron_input_shape_size],
[FLAGS.perceptron_output_shape_size],
"softmax_L2")
layer3 = nn.layer_perceptron(layer2, [FLAGS.perceptron_input_shape_size],
[FLAGS.label_number], "softmax_L3")
# softmax layer
with tf.name_scope("softmax_func"):
W_softmax = tf.Variable(tf.zeros([10, 10]), name="W_softmax")
h = tf.nn.softmax(tf.matmul(layer3, W_softmax), name="h")
ts.variable_summaries(h)
# cross entropy function
with tf.name_scope("cross_entropy"):
cost = tf.reduce_mean(
-tf.reduce_sum(ph_set["Y"] * tf.log(h), reduction_indices=1))
ts.variable_summaries(cost)
# train op
train_op = tf.train.GradientDescentOptimizer(
FLAGS.learning_rate).minimize(cost)
# predicted label and batch batch_acc
predicted_label = tf.cast(tf.arg_max(h, 1, name="predicted_label"),
tf.float32)
with tf.name_scope("softmax_batch_acc"):
with tf.name_scope("accuracy"):
batch_acc = tf.reduce_mean(tf.cast(
tf.equal(predicted_label, ph_set["Y_label"]), tf.float32),
name="batch_acc")
tf.summary.scalar("accuracy", batch_acc)
with tf.name_scope("batch_hit_count"):
batch_hit_count = tf.reduce_sum(tf.cast(
tf.equal(predicted_label, ph_set["Y_label"]), tf.float32),
name="batch_hit_count")
tf.summary.scalar("hit_count", batch_hit_count)
# merge summary
summary = tf.summary.merge_all()
# init op
init_op = tf.global_variables_initializer()
# save tensor
tensor_set = {
"X": ph_set["X"],
"Y": ph_set["Y"],
"Y_label": ph_set["Y_label"],
"layer1": layer1,
"layer2": layer2,
"layer3": layer3,
"W_softmax ": W_softmax,
"h": h,
"cost": cost,
"train_op": train_op,
"predicted_label": predicted_label,
"batch_acc": batch_acc,
"batch_hit_count ": batch_hit_count,
"init_op": init_op,
"summary": summary,
}
return tensor_set
def model_NN():
# placeholder x, y, y_label
ph_set = nn.placeholders_init()
# NN layer
layer1 = nn.layer_perceptron(ph_set["X"], [FLAGS.image_size],
[FLAGS.perceptron_input_shape_size],
"layer_1")
layer2 = nn.layer_perceptron(layer1, [FLAGS.perceptron_input_shape_size],
[FLAGS.perceptron_output_shape_size],
"layer_2")
h = nn.layer_perceptron(layer2, [FLAGS.perceptron_input_shape_size],
[FLAGS.label_number], "layer_3")
# cost function
with tf.name_scope("cost_function"):
# TODO LOOK ME logistic regression does not work, for now use square error method
# cost function square error method
cost = tf.reduce_mean((h - ph_set["Y"])**2, name="cost")
# logistic regression
# cost = -tf.reduce_mean(Y * tf.log(h) + (1 - Y) * tf.log(1 - h), name="cost")
ts.variable_summaries(cost)
# train op
with tf.name_scope("train_op"):
train_op = tf.train.GradientDescentOptimizer(
FLAGS.learning_rate).minimize(cost)
# if train_op is None:
# pass
# tf.summary.histogram(train_op)
# predicted label and batch batch_acc
predicted_label = tf.cast(tf.arg_max(h, 1, name="predicted_label"),
tf.float32)
with tf.name_scope("NN_batch_acc"):
batch_acc = tf.reduce_mean(tf.cast(
tf.equal(predicted_label, ph_set["Y_label"]), tf.float32),
name="batch_acc")
tf.summary.scalar("accuracy", batch_acc)
batch_hit_count = tf.reduce_sum(tf.cast(
tf.equal(predicted_label, ph_set["Y_label"]), tf.float32),
name="batch_hit_count")
tf.summary.scalar("hit_count", batch_hit_count)
# merge summary
summary = tf.summary.merge_all()
# init op
init_op = tf.global_variables_initializer()
# save tensor
tensor_set = {
"X": ph_set["X"],
"Y": ph_set["Y"],
"Y_label": ph_set["Y_label"],
"layer1": layer1,
"layer2": layer2,
"h": h,
"cost": cost,
"train_op": train_op,
"predicted_label": predicted_label,
"batch_acc": batch_acc,
"batch_hit_count ": batch_hit_count,
"init_op": init_op,
"summary": summary,
}
return tensor_set