def nn_layer(input_tensor, input_dim, output_dim, layer_name, act=tf.nn.relu):
with tf.name_scope(layer_name):
with tf.name_scope(weights):
weights = tf.Variable(tf.truncate_normal(input_dim, output_dim), stddev=0.1)
variable_summaries(weights, layer_name + '/weights')
with tf.name_scope(biases):
biases = tf.Variable(tf.constant(0.0, shape=[output_dim]))
variable_summaries(biases, layer_name + '/biases')
with tf.name_scope('Wx_plus_b'):
preactivate = tf.matmul(input_tensor, weights) + biases
#记录输出节点在激活函数前的分布
tf.summary.histogram(layer_name + '/preactivations', preactivate)
activations = act(preactivate, name='activation')#激活函数
#记录输出节点激活后的分布
tf.summary.histogram(layer_name + '/activations', activations)
return activations
def neuron_layer(X, n_neurons, name, activation=None):
with tf.name_scope(
name): # create name scope (-> will look nicer in TensorBoard)
n_inputs = int(X.get_shape()[1]) # retrieve nbr of inputs
stddev = 2 / np.sqrt(n_inputs) # help faster convergence
init = tf.truncate_normal(
(n_inputs, n_neurons), stddev=stddev
) # contain weights between each input and each neuron, hence n_inputs*n_neurons
# -> use of truncated normal ensures no large weights which could slow down the training
W = tf.Variables(
init,
name="weights") # random init. is important to break symmetry !!!
b = tf.Variable(tf.zeros([n_neurons]), name="biases")
z = tf.matmul(X, W) + b
if activation == "relu":
return tf.nn.relu(z)
else:
return z
def weight_variable(shape):
return tf.Variable(tf.truncate_normal(shape, stddev=0.1))
def weight_variable(shape): initial = tf.truncate_normal(shape, stddev=0.1) return tf.Variable(initial)
def cnn_model(train_x, train_y, test_x, test_y, device='/cpu:0', epoch=1):
with tf.device(device):
x = tf.placeholder(tf.float32, [None, 48, 48, 3])
y = tf.placeholder(tf.float32, [None, 12])
keep_prob = tf.placeholder(tf.float32)
conv1_w = tf.Variable(tf.truncate_normal([8, 8, 3, 8]), tf.float32)
conv1_b = tf.Variable(tf.zeros([8]))
conv2_w = tf.Variable(tf.truncate_normal([4, 4, 8, 16]), tf.float32)
conv2_b = tf.Variable(tf.zeros([16]))
conv3_w = tf.Variable(tf.truncate_normal([4, 4, 16, 32]), tf.float32)
conv3_b = tf.Variable(tf.zeros[32])
fc1_w = tf.Variable(tf.truncate_normal([48 * 48 * 32, 200]),
tf.float32)
fc1_b = tf.Variable(tf.ones[200])
fc2_w = tf.Variable(tf.truncate_normal([200, 200]), tf.float32)
fc2_b = tf.Variable(tf.ones[200])
fc3_w = tf.Variable(tf.truncate_normal([200, 12]), tf.float32)
fc3_b = tf.Variable(tf.ones[12])
#MODEL
conv1 = tf.nn.relu(
tf.nn.conv2d(x, conv1_w, strides=[1, 1, 1, 1], padding='SAME'))
max_1 = tf.nn.max_pool(conv1,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
conv2 = tf.nn.relu(
tf.nn.conv2d(max_1, conv2_w, strides=[1, 1, 1, 1], padding='SAME'))
max_2 = tf.nn.max_pool(conv2,
kstrides=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
conv3 = tf.nn.relu(
tf.nn.conv2d(max_2, conv3_w, stride[1, 1, 1, 1], padding='SAME'))
max_3 = tf.nn.max_pool(conv3,
kstrides=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
reshaped = tf.reshape(max_3, [-1, 48 * 48 * 32])
fc1 = tf.nn.dropout(tf.nn.relu(tf.matmul(reshaped, fc1_w) + fc1_b),
keep_prob)
fc2 = tf.nn.dropout(tf.nn.relu(tf.matmul(fc1, fc2_w) + fc2_b),
keep_prob)
fc3 = tf.matmul(fc2, tf3_w) + fc3_b
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=fc3))
train = tf.train.AdamOptimizer().minimize(cross_entropy)
cross_prediction = tf.equal(tf.argmax(fc3, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(cross_prediction, tf.float32))
sess = tf.InteractiveSession()
tf.initiate_global_variables().run()
for i in range(epoch):
if i % 100 == 0:
print(sess.run(accuracy, feed_dict={x: x, y: y, keep_prob: 0.6}))
