def inference(images):
k = 32
b1 = 6
b2 = 12
b3 = 24
b4 = 16
# conv1
conv1 = helper.convolution(images, [7, 7, 3, k], helper.DECAY, 2, 'conv1')
# pool1
pool1 = tf.nn.max_pool(conv1, [1, 3, 3, 1], [1, 2, 2, 1],
'SAME',
name='pool1')
input1 = helper.dense_block(pool1, k, b1, 'conv1')
transition1 = helper.transition(input1, k, b1, 'transition1')
input2 = helper.dense_block(transition1, k, b2, 'conv2')
transition2 = helper.transition(input2, k, b2, 'transition2')
input3 = helper.dense_block(transition2, k, b3, 'conv3')
transition3 = helper.transition(input3, k, b3, 'transition3')
input4 = helper.dense_block(transition3, k, b4, 'conv4')
avg_pool = tf.nn.avg_pool(input4,
ksize=[1, 7, 7, 1],
strides=[1, 1, 1, 1],
padding='SAME')
dense = tf.contrib.layers.flatten(avg_pool)
with tf.variable_scope("fc") as scope:
kernel = helper.variable_with_weight_decay(
'weights',
shape=[51200, helper.NUM_CLASSES],
stddev=5e-2,
wd=helper.DECAY)
conv = tf.matmul(dense, kernel)
biases = helper.variable_on_cpu('biases', helper.NUM_CLASSES,
tf.constant_initializer(0.0))
output = tf.nn.bias_add(conv, biases)
return output
def inference(input):
conv1 = helper.convolution(input, [7, 7, 3, 64], helper.DECAY, 2, 'conv1')
pool1 = tf.nn.max_pool(conv1,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool1')
block1_1 = helper.block_convolution(pool1, 64, helper.DECAY, 1,
'block1_1') + pool1
block1_2 = helper.block_convolution(block1_1, 64, helper.DECAY, 1,
'block1_2') + block1_1
block1_3 = helper.block_convolution(
block1_2, 64, helper.DECAY, 2, 'block1_3') + helper.convolution(
block1_2, [1, 1, 64, 64], helper.DECAY, 2, 'proj1')
block2_1 = helper.block_convolution(
block1_3, 128, helper.DECAY, 1, 'block2_1') + helper.convolution(
block1_3, [1, 1, 64, 128], helper.DECAY, 1, 'proj2')
block2_2 = helper.block_convolution(block2_1, 128, helper.DECAY, 1,
'block2_2') + block2_1
block2_3 = helper.block_convolution(block2_2, 128, helper.DECAY, 1,
'block2_3') + block2_2
block2_4 = helper.block_convolution(
block2_3, 128, helper.DECAY, 2, 'block2_4') + helper.convolution(
block2_3, [1, 1, 128, 128], helper.DECAY, 2, 'proj3')
block3_1 = helper.block_convolution(
block2_4, 256, helper.DECAY, 1, 'block3_1') + helper.convolution(
block2_4, [1, 1, 128, 256], helper.DECAY, 1, 'proj4')
block3_2 = helper.block_convolution(block3_1, 256, helper.DECAY, 1,
'block3_2') + block3_1
block3_3 = helper.block_convolution(block3_2, 256, helper.DECAY, 1,
'block3_3') + block3_2
block3_4 = helper.block_convolution(block3_3, 256, helper.DECAY, 1,
'block3_4') + block3_3
block3_5 = helper.block_convolution(block3_4, 256, helper.DECAY, 1,
'block3_5') + block3_4
block3_6 = helper.block_convolution(
block3_5, 256, helper.DECAY, 2, 'block3_6') + helper.convolution(
block3_5, [1, 1, 256, 256], helper.DECAY, 2, 'proj5')
block4_1 = helper.block_convolution(
block3_6, 512, helper.DECAY, 1, 'block4_1') + helper.convolution(
block3_6, [1, 1, 256, 512], helper.DECAY, 1, 'proj6')
block4_2 = helper.block_convolution(block4_1, 512, helper.DECAY, 1,
'block4_2') + block4_1
block4_3 = helper.block_convolution(block4_2, 512, helper.DECAY, 1,
'block4_3') + block4_2
avg_pool = tf.nn.avg_pool(block4_3,
ksize=[1, 7, 7, 1],
strides=[1, 1, 1, 1],
padding='SAME')
dense = tf.contrib.layers.flatten(avg_pool)
with tf.variable_scope("fc") as scope:
kernel = helper.variable_with_weight_decay(
'weights',
shape=[25088, helper.NUM_CLASSES],
stddev=5e-2,
wd=helper.DECAY)
conv = tf.matmul(dense, kernel)
biases = helper.variable_on_cpu('biases', helper.NUM_CLASSES,
tf.constant_initializer(0.0))
output = tf.nn.bias_add(conv, biases)
return output
def gaussian_blur(image, kernel, verbose=False):
return convolution(image, kernel, average=True, verbose=verbose)
def inference(images, dropout, is_training):
L = 20
N = (L - 6) // 4
d4 = 1024
d4_red = d4 / 4
d3 = d4 / 2
d3_red = d3 / 4
d2 = d3 / 2
d2_red = d2 / 4
d1 = d2 / 2
d1_red = d1 / 4
# conv1
conv1 = helper.convolution(images, [3, 3, 3, d1_red], helper.DECAY, 1,
'conv1', dropout, is_training)
conv1 = tf.nn.avg_pool(conv1,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
conv2 = helper.convolution(conv1, [3, 3, d1_red, d1], helper.DECAY, 1,
'conv2', dropout, is_training)
conv2 = tf.nn.avg_pool(conv2,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
x1 = N * [d1]
red3 = N * [d1_red]
x3 = N * [d1]
red5 = N * [d1_red]
x5 = N * [d1]
pool_proj = N * [d1]
block1 = helper.dense_inception(conv2, 'block1', dropout, is_training,
red3, red5, x1, x3, x5, pool_proj)
transition1 = helper.transition(block1, 'transition1', d2, dropout,
is_training)
x1 = N * [d2]
red3 = N * [d2_red]
x3 = N * [d2]
red5 = N * [d2_red]
x5 = N * [d2]
pool_proj = N * [d2]
block2 = helper.dense_inception(transition1, 'block2', dropout,
is_training, red3, red5, x1, x3, x5,
pool_proj)
transition2 = helper.transition(block2, 'transition2', d3, dropout,
is_training)
x1 = N * [d3]
red3 = N * [d3_red]
x3 = N * [d3]
red5 = N * [d3_red]
x5 = N * [d3]
pool_proj = N * [d3]
block3 = helper.dense_inception(transition2, 'block3', dropout,
is_training, red3, red5, x1, x3, x5,
pool_proj)
transition3 = helper.transition(block3, 'transition3', d4, dropout,
is_training)
x1 = N * [d4]
red3 = N * [d4_red]
x3 = N * [d4]
red5 = N * [d4_red]
x5 = N * [d4]
pool_proj = N * [d4]
block4 = helper.dense_inception(transition3, 'block4', dropout,
is_training, red3, red5, x1, x3, x5,
pool_proj)
output = helper.softmax_classifier(block4, 'output', dropout, is_training)
return output