def conv_net_check3(x, keep_prob):
layer = create_conv2d(x, 10, strides=[2, 2], w_name='W1')
layer = tf.nn.relu(layer)
layer = tf.nn.max_pool(layer,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
layer = create_conv2d(layer, 20, strides=[2, 2], w_name='W2')
layer = tf.nn.relu(layer)
layer = tf.nn.max_pool(layer,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
layer = create_conv2d(layer, 40, strides=[2, 2], w_name='W3')
layer = tf.nn.relu(layer)
layer = tf.nn.max_pool(layer,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
layer = create_conv2d(layer, 80, strides=[2, 2], w_name='W4')
layer = tf.nn.relu(layer)
layer = tf.nn.max_pool(layer,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
tf.nn.dropout(layer, keep_prob=keep_prob)
layer = tf.contrib.layers.flatten(layer)
layer = tf.contrib.layers.fully_connected(layer, 400)
layer = tf.nn.dropout(layer, keep_prob)
layer = tf.contrib.layers.fully_connected(layer, 200)
layer = tf.nn.dropout(layer, keep_prob)
return tf.contrib.layers.fully_connected(layer, 10, activation_fn=None)
def make_logits(tensor, keep_prob):
global n
n += 1
# Inputs
# keep_prob = cnn.neural_net_keep_prob_input()
# Model
nn = cnn.create_conv2d(tensor, 128, strides=[32, 32], w_name='W1' + str(n))
nn = tf.nn.relu(nn)
nn = tf.nn.max_pool(nn,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
nn = cnn.create_conv2d(nn, 256, strides=[16, 16], w_name='W2' + str(n))
nn = tf.nn.relu(nn)
nn = tf.nn.max_pool(nn,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
# nn = cnn.create_conv2d(nn, 256, strides=[8, 8], w_name='W3'+str(n))
# nn = tf.nn.relu(nn)
# nn = tf.nn.max_pool(nn, ksize=[1,2,2,1], strides=[1,2,2,1], padding = 'SAME')
#
# nn = cnn.create_conv2d(nn, 512, strides=[8, 8], w_name='W4'+str(n))
# nn = tf.nn.relu(nn)
# nn = tf.nn.max_pool(nn, ksize=[1,2,2,1], strides=[1,2,2,1], padding = 'SAME')
# tf.nn.dropout(nn, keep_prob=keep_prob)
#
# nn = cnn.create_conv2d(nn, 512, strides=[6, 6], w_name='W5'+str(n))
# nn = tf.nn.relu(nn)
# nn = tf.nn.max_pool(nn, ksize=[1,2,2,1], strides=[1,2,2,1], padding = 'SAME')
#
# tf.nn.dropout(nn, keep_prob=keep_prob)
#
# nn = cnn.create_conv2d(nn, 1024, strides=[6, 6], w_name='W6'+str(n))
# nn = tf.nn.relu(nn)
# nn = tf.nn.max_pool(nn, ksize=[1,2,2,1], strides=[1,2,2,1], padding = 'SAME')
tf.nn.dropout(nn, keep_prob=keep_prob)
#
# nn = cnn.create_conv2d(nn, 1024, strides=[3, 3], w_name='W7'+str(n))
# nn = tf.nn.relu(nn)
# nn = tf.nn.max_pool(nn, ksize=[1,2,2,1], strides=[1,2,2,1], padding = 'SAME')
layer = tf.contrib.layers.flatten(nn)
# layer = tf.contrib.layers.fully_connected(layer, 2048, activation_fn=tf.nn.relu)
tf.nn.dropout(layer, keep_prob=keep_prob)
layer = tf.contrib.layers.fully_connected(layer,
1024,
activation_fn=tf.nn.relu)
# layer = tf.contrib.layers.fully_connected(layer, 100, activation_fn=tf.nn.relu)
layer = tf.contrib.layers.fully_connected(layer, 1, activation_fn=None)
return layer
def make_logits_for_live_prediction(tensor, keep_prob):
nn = cnn.create_conv2d(tensor, 64, strides=[4, 4], w_name='W1')
nn = tf.nn.relu(nn)
nn = tf.nn.max_pool(nn,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
nn = cnn.create_conv2d(nn, 128, strides=[4, 4], w_name='W2')
nn = tf.nn.relu(nn)
nn = tf.nn.max_pool(nn,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
nn = cnn.create_conv2d(nn, 256, strides=[2, 2], w_name='W4')
nn = tf.nn.relu(nn)
nn = tf.nn.max_pool(nn,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
nn = cnn.create_conv2d(nn, 512, strides=[2, 2], w_name='W5')
nn = tf.nn.relu(nn)
nn = tf.nn.max_pool(nn,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
# nn = cnn.create_conv2d(nn, 512, strides=[2, 2], w_name='W6')
# nn = tf.nn.relu(nn)
# nn = tf.nn.max_pool(nn, ksize=[1,2,2,1], strides=[1,2,2,1], padding = 'SAME')
#
# nn = cnn.create_conv2d(nn, 1024, strides=[2, 2], w_name='W7')
# nn = tf.nn.relu(nn)
# nn = tf.nn.max_pool(nn, ksize=[1,2,2,1], strides=[1,2,2,1], padding = 'SAME')
tf.nn.dropout(nn, keep_prob=keep_prob)
layer = tf.contrib.layers.fully_connected(
nn, 1024) # activation_fn=tf.nn.relu)
tf.nn.dropout(layer, keep_prob=keep_prob)
layer = tf.contrib.layers.fully_connected(
layer, 512) #, activation_fn=tf.nn.relu)
tf.nn.dropout(layer, keep_prob=keep_prob)
layer = tf.contrib.layers.fully_connected(layer, 1, activation_fn=None)
return layer
keep_prob = cnn.neural_net_keep_prob_input()
# Model
# nn1 = cnn.create_conv2d(x, 32, strides=[8, 8], w_name='W1')
w_size, c_strides = cnn.get_weights_shape(x, 32, [8, 8])
W1 = tf.get_variable('W1',
w_size,
initializer=tf.contrib.layers.xavier_initializer(seed=0))
Z1 = tf.nn.conv2d(x, W1, strides=c_strides, padding='SAME', name='W1_conv2d')
nn2 = tf.nn.relu(Z1)
nn3 = tf.nn.max_pool(nn2,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
nn = cnn.create_conv2d(nn3, 64, strides=[4, 4], w_name='W2')
nn = tf.nn.relu(nn)
nn = tf.nn.max_pool(nn,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
nn = cnn.create_conv2d(nn, 128, strides=[3, 3], w_name='W3')
nn = tf.nn.relu(nn)
nn = tf.nn.max_pool(nn,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
nn = cnn.create_conv2d(nn, 256, strides=[3, 3], w_name='W4')
nn = tf.nn.relu(nn)
nn = tf.nn.max_pool(nn,
ksize=[1, 2, 2, 1],
imw = 90
imh = 120
n_classes = 1
epochs = 15
batch_size = 64
keep_probability = 0.5
tf.reset_default_graph()
# Inputs
x = cnn.neural_net_image_input((imw, imh, 3))
y = cnn.neural_net_label_input(n_classes)
keep_prob = cnn.neural_net_keep_prob_input()
# Model
nn = cnn.create_conv2d(x, 64, strides=[3, 3], w_name='W1')
nn = tf.nn.relu(nn, name='W1_activated')
nn = tf.nn.max_pool(nn,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
nn = cnn.create_conv2d(nn, 128, strides=[3, 3], w_name='W2')
nn = tf.nn.relu(nn, name='W2_activated')
nn = tf.nn.max_pool(nn,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
nn = cnn.create_conv2d(nn, 256, strides=[2, 2], w_name='W3')
nn = tf.nn.relu(nn, name='W3_activated')
