myNet.py, Movidius NSC

import tensorflow as tf
import numpy as np

USING_SLICE = True


def leaky_relu(x, alpha=0.2):
    return tf.maximum(x, alpha * x)

def conv2d_leaky(input, kernel_shape, bias_shape, strides=1, relu=True, padding='SAME'):

    # Conv2D
    weights = tf.get_variable("weights", kernel_shape, initializer=tf.contrib.layers.xavier_initializer())
    biases = tf.get_variable("biases", bias_shape, initializer=tf.truncated_normal_initializer())
    output = tf.nn.conv2d(input, weights, strides=[1, strides, strides, 1], padding=padding)
    output = tf.nn.bias_add(output, biases)

    # ReLU (if required)
    if relu == False:
        print ('WARNING: reLU disabled')
    else:
        output = leaky_relu(output)
    return output

def decoder(volume):
    with tf.variable_scope("conv1") as scope:
      conv1 = conv2d_leaky(volume, [3, 3, volume.get_shape().as_list()[3], 128], [128], 1, True)
    with tf.variable_scope("conv2") as scope:
      conv2 = conv2d_leaky(conv1, [3, 3, conv1.get_shape().as_list()[3], 128], [128], 1, True)
    with tf.variable_scope("conv3") as scope:
      conv3 = conv2d_leaky(conv2, [3, 3, conv2.get_shape().as_list()[3], 96], [96], 1, True)
    with tf.variable_scope("conv4") as scope:
      conv4 = conv2d_leaky(conv3, [3, 3, conv3.get_shape().as_list()[3], 64], [64], 1, True)
    with tf.variable_scope("conv5") as scope:
      conv5 = conv2d_leaky(conv4, [3, 3, conv4.get_shape().as_list()[3], 32], [32], 1, True)
    with tf.variable_scope("conv6") as scope:
      conv6 = conv2d_leaky(conv5, [3, 3, conv5.get_shape().as_list()[3], 8], [8], 1, False)
    return conv6

def encoder(input_batch):
    with tf.variable_scope("conv1"):
      conv1 = conv2d_leaky(input_batch, [3, 3, 3, 16], [16], 2, True)
    with tf.variable_scope("conv2"):
      conv2 = conv2d_leaky(conv1, [3, 3, 16, 16], [16], 1, True)
    with tf.variable_scope("conv3"):
      conv3 = conv2d_leaky(conv2, [3, 3, 16, 32], [32], 2, True)
    with tf.variable_scope("conv4"):
      conv4 = conv2d_leaky(conv3, [3, 3, 32, 32], [32], 1, True)
    with tf.variable_scope("conv5"):
      conv5 = conv2d_leaky(conv4, [3, 3, 32, 64], [64], 2, True)
    with tf.variable_scope("conv6"):
      conv6 = conv2d_leaky(conv5, [3, 3, 64, 64], [64], 1, True)
    with tf.variable_scope("conv7"):
      conv7 = conv2d_leaky(conv6, [3, 3, 64, 96], [96], 2, True)
    with tf.variable_scope("conv8"):
      conv8 = conv2d_leaky(conv7, [3, 3, 96, 96], [96], 1, True)
    with tf.variable_scope("conv9"):
      conv9 = conv2d_leaky(conv8, [3, 3, 96, 128], [128], 2, True)
    with tf.variable_scope("conv10"):
      conv10 = conv2d_leaky(conv9, [3, 3, 128, 128], [128], 1, True)
    with tf.variable_scope("conv11"):
      conv11 = conv2d_leaky(conv10, [3, 3, 128, 192], [192], 2, True)
    with tf.variable_scope("conv12"):
      conv12 = conv2d_leaky(conv11, [3, 3, 192, 192], [192], 1, True)
    return conv12

def myNet(image, sliceOp=True):
  with tf.variable_scope("encoder-1") as scope:
    features1 = encoder(image)
  with tf.variable_scope("encoder-2") as scope:
    features2 = encoder(image)
  if sliceOp:
    features = tf.concat([features1, tf.slice(features2, [0, 0, 0, 0], [-1, -1, -1, 1])], -1)
  else:
    features = tf.concat([features1, features2], -1)
  output = decoder(features)
  return output

def run(name, height, width):
  with tf.Graph().as_default():
    image = tf.placeholder("float", [1, height, width, 3], name="input")
    output = myNet(image, USING_SLICE)
    with tf.variable_scope("result") as scope:
       output = output + output

    init = tf.group(tf.global_variables_initializer(),
                   tf.local_variables_initializer())

    saver = tf.train.Saver()

    with tf.Session() as sess:
        #loading weights TODO
        sess.run(init)
        saver.save(sess, "output/"+name)

run('myNet', 576, 960)