def shallow2(inpt, inpt_size, is_training):
'''
18-ticnn
'''
end_points = {}
keep_prob = tf.cond(is_training, lambda: tf.constant(0.5),
lambda: tf.constant(1.0))
with tf.name_scope('reshape'):
x_image = tf.reshape(inpt, [-1, inpt_size, 1, 1])
## first conv-----------------------------
num_feature = 16
out = tu.add_conv1d_layer(x_image,
num_feature,
9,
2,
is_training=is_training,
layer_name='conv1')
end_points['conv1'] = out
out = tu.batch_normalization(out, is_training=is_training)
out = tu.max_pool(out, ksize=4, stride=2, layer_name='pool1')
end_points['pool1'] = out
tu.print_activations(out)
out = tu.add_conv1d_layer(out,
num_feature,
9,
2,
is_training=is_training,
layer_name='conv2')
end_points['conv2'] = out
out = tu.max_pool(out, ksize=4, stride=2, layer_name='pool2')
end_points['pool2'] = out
tu.print_activations(out)
## fully connected layers-----------------------------
with tf.name_scope('fc1'):
out = tu.add_fc_layer(out,
100,
relu=True,
BN=True,
is_training=is_training) # previously 256 nodes
out = tf.nn.dropout(out, keep_prob)
with tf.name_scope('fc2-1'):
out1 = tu.add_fc_layer(out, 3)
with tf.name_scope('fc2-2'):
print(end_points.keys())
out2 = tu.add_fc_layer(out, 1)
end_points['class_end'] = out1
end_points['speed_end'] = out2
return end_points
def speednet1(inpt, inpt_size,
is_training):
keep_prob = tf.cond(is_training, lambda: tf.constant(0.5), lambda: tf.constant(1.0))
end_point={}
with tf.name_scope('reshape'):
x_image = tf.reshape(inpt, [-1, inpt_size,1, 1])
with tf.name_scope('block1'):
out = tu.add_conv1d_layer(x_image,32,9,3,is_training=is_training)
out = tu.max_pool(out, ksize=4)
end_point['block1'] = out
with tf.name_scope('block8'):
out2 = tu.add_fc_layer(out, 128, relu=True, BN=True,
is_training=is_training)
out2 = tf.nn.dropout(out2, keep_prob)
with tf.name_scope('block9'):
out2 = tu.add_fc_layer(out2, 1)
with tf.name_scope('block2'):
out = tu.add_conv1d_layer(out,64,9,is_training=is_training)
out = tu.max_pool(out, ksize=4)
end_point['block2'] = out
with tf.name_scope('block3'):
out = tu.add_conv1d_layer(out,64,9,is_training=is_training)
out = tu.max_pool(out, ksize=4)
with tf.name_scope('block4'):
out = tu.add_conv1d_layer(out,64,9,is_training=is_training)
out = tu.max_pool(out, ksize=4)
# with tf.name_scope('block5'):
# out = tu.add_conv1d_layer(out,64,9,is_training=is_training)
# out = tu.max_pool(out, ksize=4)
with tf.name_scope('block6'):
out1 = tu.add_fc_layer(out, 256, relu=True, BN=True,
is_training=is_training)
out1 = tf.nn.dropout(out1, keep_prob)
with tf.name_scope('block7'):
out1 = tu.add_fc_layer(out1, 3)
end_point['class_end'] = out1
end_point['speed_end'] = out2
return end_point
def res_block(inpt, out_channel, name_scope=None, down_sample=False):
with tf.name_scope(name_scope):
if down_sample:
inpt = tu.add_conv1d_layer(inpt,
out_channel,
3,
2,
BN=False,
layer_name='conv0')
out = tu.add_conv1d_layer(inpt,
out_channel // 2,
3,
1,
BN=False,
layer_name='conv1')
out = tu.add_conv1d_layer(out,
out_channel,
3,
1,
BN=True,
layer_name='conv2')
return inpt + out
def simpnet2(inpt, inpt_size, is_training):
keep_prob = tf.cond(is_training, lambda: tf.constant(0.5),
lambda: tf.constant(1.0))
with tf.name_scope('reshape'):
x_image = tf.reshape(inpt, [-1, inpt_size, 1, 1])
with tf.name_scope('block1'):
out = tu.add_conv1d_layer(x_image, 32, 64, is_training=is_training)
out = tu.max_pool(out)
with tf.name_scope('block2'):
out = tu.add_conv1d_layer(out, 64, 9, is_training=is_training)
out = tu.max_pool(out)
with tf.name_scope('block3'):
out = tu.add_conv1d_layer(out, 64, 9, is_training=is_training)
out = tu.max_pool(out)
with tf.name_scope('block4'):
out = tu.add_conv1d_layer(out, 64, 9, is_training=is_training)
out = tu.max_pool(out)
with tf.name_scope('block5'):
out = tu.add_conv1d_layer(out, 64, 9, is_training=is_training)
out = tu.max_pool(out)
with tf.name_scope('block6'):
out = tu.add_fc_layer(out,
128,
relu=True,
BN=True,
is_training=is_training)
out = tf.nn.dropout(out, keep_prob)
with tf.name_scope('block7'):
y_conv = tu.add_fc_layer(out, 3)
return y_conv
def cAlex(inpt, inpt_size, is_training):
keep_prob = tf.cond(is_training, lambda: tf.constant(0.5),
lambda: tf.constant(1.0))
with tf.name_scope('reshape'):
x_image = tf.reshape(inpt, [-1, inpt_size, 1, 1])
## first conv-----------------------------
with tf.name_scope('conv1'):
conv1_out = tu.add_conv1d_layer(x_image,
16,
kernel_size=64,
stride=16,
is_training=is_training)
# output size = INPUT_SIZE/stride_size=256 --->[-1,256,1,num_feature1]
# this formula is for 'padding="SAME"' situation
with tf.name_scope('pool1'):
h_pool1 = tu.max_pool(conv1_out, ksize=3, stride=2, padding='VALID')
tu.print_activations(h_pool1)
# output size: (255-3+1)/2=127(ceiling) --->[-1,127,1,num_feature1]
## second conv-----------------------------
with tf.name_scope('conv2'):
conv2_out = tu.add_conv1d_layer(h_pool1,
32,
kernel_size=25,
is_training=is_training)
with tf.name_scope('pool2'):
h_pool2 = tu.max_pool(conv2_out, ksize=3, stride=2, padding='VALID')
tu.print_activations(h_pool2)
# output size: (127-3+1)/2=63 ---> [-1,63,1,num_feature2]
## third conv-----------------------------
with tf.name_scope('conv3'):
conv3_out = tu.add_conv1d_layer(h_pool2,
32,
9,
print_activation=False,
is_training=is_training)
with tf.name_scope('conv4'):
conv4_out = tu.add_conv1d_layer(conv3_out,
32,
9,
print_activation=False,
is_training=is_training)
with tf.name_scope('conv5'):
conv5_out = tu.add_conv1d_layer(conv4_out,
25,
9,
print_activation=False,
is_training=is_training)
with tf.name_scope('pool3'):
h_pool3 = tu.max_pool(conv5_out, ksize=3, stride=2, padding='VALID')
tu.print_activations(h_pool3)
# output size: (63-3+1)/2=31 ---> [-1,31,1,num_feature5]
## fully connected layers-----------------------------
with tf.name_scope('fc1'):
fc1_out = tu.add_fc_layer(h_pool3,
128,
relu=True,
BN=True,
is_training=is_training)
with tf.name_scope('dropout1'):
h_fc1_drop = tf.nn.dropout(fc1_out, keep_prob)
with tf.name_scope('fc2'):
fc1_out = tu.add_fc_layer(h_fc1_drop,
128,
relu=True,
BN=True,
is_training=is_training)
with tf.name_scope('dropout2'):
h_fc2_drop = tf.nn.dropout(fc1_out, keep_prob)
with tf.name_scope('fc3'):
y_conv = tu.add_fc_layer(h_fc2_drop, 3)
return y_conv
def cvgg19_with_pindex(inpt, inpt_size, is_training):
end_points = {}
keep_prob = tf.cond(is_training, lambda: tf.constant(0.5),
lambda: tf.constant(1.0))
with tf.name_scope('reshape'):
x_image = tf.reshape(inpt, [-1, inpt_size, 1, 1])
## first conv-----------------------------
with tf.name_scope('block1'):
num_feature = 32
out = tu.add_conv1d_layer(x_image,
num_feature,
9,
2,
BN=False,
layer_name='conv1')
end_points['conv1'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
2,
BN=False,
layer_name='conv2')
end_points['conv2'] = out
out = tu.batch_normalization(out, is_training=is_training)
out, pool_args = tu.max_pool_with_argmax(out, 4, layer_name='pool1')
end_points['pool1'] = out
end_points['pool1_arg'] = pool_args
tu.print_activations(out)
## second conv-----------------------------
with tf.name_scope('block2'):
num_feature = num_feature * 2
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv3')
end_points['conv3'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
is_training=is_training,
layer_name='conv4',
print_activation=True)
end_points['conv4'] = out
out, pool_args = tu.max_pool_with_argmax(out, 4, layer_name='pool2')
end_points['pool2_arg'] = pool_args
end_points['pool2'] = out
tu.print_activations(out)
## third conv-----------------------------
with tf.name_scope('block3'):
num_feature = num_feature * 2
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv6')
end_points['conv6'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv7')
end_points['conv7'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
is_training=is_training,
layer_name='conv8',
print_activation=True)
end_points['conv8'] = out
out, pool_args = tu.max_pool_with_argmax(out, 4, layer_name='pool3')
end_points['pool3_arg'] = pool_args
end_points['pool3'] = out
tu.print_activations(out)
## forth conv-----------------------------
with tf.name_scope('block4'):
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv10')
end_points['conv10'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv11')
end_points['conv11'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
is_training=is_training,
layer_name='conv12',
print_activation=True)
end_points['conv12'] = out
out, pool_args = tu.max_pool_with_argmax(out, 4, layer_name='pool4')
end_points['pool4_arg'] = pool_args
# out = tu.global_average_pool(out)
end_points['pool4'] = out
tu.print_activations(out)
## fully connected layers-----------------------------
with tf.name_scope('fc1'):
out = tu.add_fc_layer(out,
256,
relu=True,
BN=True,
is_training=is_training)
out = tf.nn.dropout(out, keep_prob)
with tf.name_scope('fc4'):
out = tu.add_fc_layer(out, 3)
with tf.name_scope('block8'):
print(end_points.keys())
out2 = tu.add_fc_layer(end_points['pool4'],
256,
relu=True,
BN=True,
is_training=is_training)
out2 = tf.nn.dropout(out2, keep_prob)
with tf.name_scope('block9'):
out2 = tu.add_fc_layer(out2, 1)
end_points['class_end'] = out
end_points['speed_end'] = out2
return end_points
def cvgg19_3(inpt, inpt_size, is_training, num_class=3):
end_points = {}
keep_prob = tf.cond(is_training, lambda: tf.constant(0.5),
lambda: tf.constant(1.0))
with tf.name_scope('reshape'):
x_image = tf.reshape(inpt, [-1, inpt_size, 1, 1])
## first conv-----------------------------
num_feature = 16
out = tu.add_conv1d_layer(x_image,
num_feature,
9,
2,
BN=False,
layer_name='conv1')
end_points['conv1'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
2,
BN=False,
layer_name='conv2')
end_points['conv2'] = out
out = tu.batch_normalization(out, is_training=is_training)
out = tu.max_pool(out, ksize=4, layer_name='pool1')
end_points['pool1'] = out
tu.print_activations(out)
## second conv-----------------------------
num_feature = num_feature * 2
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv3')
end_points['conv3'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
is_training=is_training,
layer_name='conv4',
print_activation=True)
end_points['conv4'] = out
out = tu.max_pool(out, ksize=4, layer_name='pool2')
end_points['pool2'] = out
tu.print_activations(out)
## third conv-----------------------------
num_feature = num_feature * 2
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv6')
end_points['conv5'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv7')
end_points['conv6'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
is_training=is_training,
layer_name='conv8',
print_activation=True)
end_points['conv7'] = out
out = tu.max_pool(out, ksize=4, layer_name='pool3')
end_points['pool3'] = out
tu.print_activations(out)
## forth conv-----------------------------
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv10')
end_points['conv8'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv11')
end_points['conv9'] = out
out1 = tu.add_conv1d_layer(out,
num_class,
1,
is_training=is_training,
layer_name='conv10',
print_activation=True)
out2 = tu.add_conv1d_layer(out,
1,
1,
is_training=is_training,
layer_name='conv11',
print_activation=True)
end_points['conv10'] = out1
# out = tu.max_pool(out, ksize=4, layer_name='pool4')
out1 = tu.global_average_pool(out1)
out2 = tu.global_average_pool(out2)
end_points['GAP'] = out1
tu.print_activations(out1)
out1 = tf.reshape(out1, [-1, num_class])
out2 = tf.reshape(out2, [-1, 1])
print("out1" + str(out1.get_shape().as_list()))
print("out2" + str(out2.get_shape().as_list()))
end_points['class_end'] = out1
end_points['speed_end'] = out2
return end_points
def cvgg19_2(inpt, inpt_size, is_training):
end_points = {}
keep_prob = tf.cond(is_training, lambda: tf.constant(0.5),
lambda: tf.constant(1.0))
with tf.name_scope('reshape'):
x_image = tf.reshape(inpt, [-1, inpt_size, 1, 1])
## first conv-----------------------------
num_feature = 16
out = tu.add_conv1d_layer(x_image,
num_feature,
9,
2,
BN=False,
layer_name='conv1')
end_points['conv1'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
2,
BN=False,
layer_name='conv2')
end_points['conv2'] = out
out = tu.batch_normalization(out, is_training=is_training)
out = tu.max_pool(out, ksize=4, layer_name='pool1')
end_points['pool1'] = out
tu.print_activations(out)
## second conv-----------------------------
num_feature = num_feature * 2
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv3')
end_points['conv3'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
is_training=is_training,
layer_name='conv4',
print_activation=True)
end_points['conv4'] = out
out = tu.max_pool(out, ksize=4, layer_name='pool2')
end_points['pool2'] = out
tu.print_activations(out)
## third conv-----------------------------
num_feature = num_feature * 2
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv6')
end_points['conv5'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv7')
end_points['conv6'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
is_training=is_training,
layer_name='conv8',
print_activation=True)
end_points['conv7'] = out
out = tu.max_pool(out, ksize=4, layer_name='pool3')
end_points['pool3'] = out
tu.print_activations(out)
## forth conv-----------------------------
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv10')
end_points['conv8'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
BN=False,
layer_name='conv11')
end_points['conv9'] = out
out = tu.add_conv1d_layer(out,
num_feature,
9,
is_training=is_training,
layer_name='conv12',
print_activation=True)
end_points['conv10'] = out
# out = tu.max_pool(out, ksize=4, layer_name='pool4')
out = tu.global_average_pool(out)
end_points['GAP'] = out
tu.print_activations(out)
## fully connected layers-----------------------------
with tf.name_scope('fc1'):
out = tu.add_fc_layer(out,
256,
relu=True,
BN=True,
is_training=is_training)
out = tf.nn.dropout(out, keep_prob)
with tf.name_scope('fc2-1'):
out1 = tu.add_fc_layer(out, 3)
with tf.name_scope('fc2-2'):
out2 = tu.add_fc_layer(out, 1)
end_points['class_end'] = out1
end_points['speed_end'] = out2
return end_points