def __third_block(x):
from yolo.architecture import layer_functions as lf
layer_0 = __second_block(x)
layer_1 = lf.leaky_relu(lf.conv2d(layer_0, [1, 1, 192, 128], name='w_3_1'))
layer_2 = lf.leaky_relu(lf.conv2d(layer_1, [3, 3, 128, 256], name='w_3_2'))
layer_3 = lf.leaky_relu(lf.conv2d(layer_2, [1, 1, 256, 256], name='w_3_3'))
layer_4 = lf.leaky_relu(lf.conv2d(layer_3, [3, 3, 256, 512], name='w_3_4'))
return lf.max_pool(layer_4, 2, 2)
def __fifth_block(x):
layer_0 = __fourth_block(x)
layer_1 = lf.leaky_relu(
lf.conv2d(layer_0, [1, 1, 1024, 512], name='w_5_1'))
layer_2 = lf.leaky_relu(
lf.conv2d(layer_1, [3, 3, 512, 1024], name='w_5_2'))
layer_3 = lf.leaky_relu(
lf.conv2d(layer_2, [1, 1, 1024, 512], name='w_5_3'))
layer_4 = lf.leaky_relu(
lf.conv2d(layer_3, [3, 3, 512, 1024], name='w_5_4'))
layer_5 = tf.reduce_mean(layer_4, [1, 2])
w = lf.weight_variable([1024, 1000], name='pre_w_6')
return lf.leaky_relu(tf.matmul(tf.reshape(layer_5, [-1, 1024]), w))
def __fourth_block(x):
from yolo.architecture import layer_functions as lf
layer_0 = __third_block(x)
layer_1 = lf.leaky_relu(lf.conv2d(layer_0, [1, 1, 512, 256], name='w_4_1'))
layer_2 = lf.leaky_relu(lf.conv2d(layer_1, [3, 3, 256, 512], name='w_4_2'))
layer_3 = lf.leaky_relu(lf.conv2d(layer_2, [1, 1, 512, 256], name='w_4_3'))
layer_4 = lf.leaky_relu(lf.conv2d(layer_3, [3, 3, 256, 512], name='w_4_4'))
layer_5 = lf.leaky_relu(lf.conv2d(layer_4, [1, 1, 512, 256], name='w_4_5'))
layer_6 = lf.leaky_relu(lf.conv2d(layer_5, [3, 3, 256, 512], name='w_4_6'))
layer_7 = lf.leaky_relu(lf.conv2d(layer_6, [1, 1, 512, 256], name='w_4_7'))
layer_8 = lf.leaky_relu(lf.conv2d(layer_7, [3, 3, 256, 512], name='w_4_8'))
layer_9 = lf.leaky_relu(lf.conv2d(layer_8, [1, 1, 512, 512], name='w_4_9'))
layer_10 = lf.leaky_relu(
lf.conv2d(layer_9, [3, 3, 512, 1024], name='w_4_10'))
return lf.max_pool(layer_10, 2, 2)
def __seventh_block(x, keep_prob):
from yolo.architecture import layer_functions as lf
import tensorflow as tf
layer_0 = __sixth_block(x)
w = lf.weight_variable([7 * 7 * 1024, 4096], name='w_7')
conn = lf.leaky_relu(tf.matmul(tf.reshape(layer_0, [-1, 7 * 7 * 1024]), w))
return tf.nn.dropout(conn, rate=1 - keep_prob)
def __fifth_block(x):
from yolo.architecture import layer_functions as lf
layer_0 = __fourth_block(x)
layer_1 = lf.leaky_relu(lf.conv2d(layer_0, [1, 1, 1024, 512],
name='w_5_1'))
layer_2 = lf.leaky_relu(lf.conv2d(layer_1, [3, 3, 512, 1024],
name='w_5_2'))
layer_3 = lf.leaky_relu(lf.conv2d(layer_2, [1, 1, 1024, 512],
name='w_5_3'))
layer_4 = lf.leaky_relu(lf.conv2d(layer_3, [3, 3, 512, 1024],
name='w_5_4'))
layer_5 = lf.leaky_relu(
lf.conv2d(layer_4, [3, 3, 1024, 1024], name='w_5_5'))
return lf.leaky_relu(
lf.conv2d(layer_5, [3, 3, 1024, 1024], 2, name='w_5_6'))
def __second_block(x):
from yolo.architecture import layer_functions as lf
layer_0 = __first_block(x)
return lf.max_pool(
lf.leaky_relu(lf.conv2d(layer_0, [3, 3, 64, 192], 1, name='w_2')), 2,
2)
def __sixth_block(x):
from yolo.architecture import layer_functions as lf
layer_0 = __fifth_block(x)
layer_1 = lf.leaky_relu(
lf.conv2d(layer_0, [3, 3, 1024, 1024], name='w_6_1'))
return lf.leaky_relu(lf.conv2d(layer_1, [3, 3, 1024, 1024], name='w_6_2'))
def __first_block(x):
from yolo.architecture import layer_functions as lf
return lf.max_pool(
lf.leaky_relu(lf.conv2d(x, [7, 7, 3, 64], 2, name='w_1')), 2, 2)