squeezenet.py

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import tensorflow as tf
from tensorflow.contrib.framework import add_arg_scope
from tensorflow.contrib.layers.python.layers import utils
slim = tf.contrib.slim

def squeeze(inputs, num_outputs, fire_id):
    return slim.conv2d(inputs, num_outputs, 1, stride=1, scope="fire/squeeze/"+str(fire_id))

def expand(inputs, num_outputs, fire_id):
    with tf.variable_scope('expand'):
        e1x1 = slim.conv2d(inputs, num_outputs, 1, stride=1, scope='fire/ex1x1/'+str(fire_id))
        e3x3 = slim.conv2d(inputs, num_outputs, 3, scope='fire/ex3x3/'+str(fire_id))
    return tf.concat(3, [e1x1, e3x3])

def fire_module(x, s=16, e=64, fire_id=0):
    return expand(squeeze(x, s, fire_id), e, fire_id)

def inference(images, _a, phase_train=True, weight_decay=0.8, reuse=None):
    with tf.variable_scope('SqueezeNet', 'SqueezeNet', [images], reuse=reuse):
            x_image = tf.reshape(images, [-1, 160, 160, 3])
            net = slim.conv2d(x_image, 96, 7, scope="conv_1", stride=2) #[?,14,14,64]
            net = slim.max_pool2d(net, 3, stride=2, scope="maxpool1") #[?,6,6,64]
            net = fire_module(net, fire_id=2)
            net = fire_module(net, fire_id=3)
            net = fire_module(net, 32, 128, fire_id=4)
            net = slim.max_pool2d(net, 3, scope="maxpool3")
            net = fire_module(net, 32, 128, fire_id=5)
            net = fire_module(net, 48, 192, fire_id=6)
            net = fire_module(net, 48, 192, fire_id=7)
            net = fire_module(net, 64, 256, fire_id=8)
            net = slim.max_pool2d(net, 2, stride=2, scope="maxpool8")
            net = fire_module(net, 64, 256, fire_id=9)
            net = slim.dropout(net, 0.5)
            net = slim.conv2d(net, 1792, 1, scope="conv10", padding="VALID")
            net = slim.avg_pool2d(net, 9)
            net = slim.flatten(net)
    return net,{}