Example of 3D convolutional network with TensorFlow

conv3dnet.py

import tensorflow as tf
import numpy as np

FC_SIZE = 1024
DTYPE = tf.float32


def _weight_variable(name, shape):
    return tf.get_variable(name, shape, DTYPE, tf.truncated_normal_initializer(stddev=0.1))


def _bias_variable(name, shape):
    return tf.get_variable(name, shape, DTYPE, tf.constant_initializer(0.1, dtype=DTYPE))


def inference(boxes, dataconfig):
    prev_layer = boxes

    in_filters = dataconfig.num_props
    with tf.variable_scope('conv1') as scope:
        out_filters = 16
        kernel = _weight_variable('weights', [5, 5, 5, in_filters, out_filters])
        conv = tf.nn.conv3d(prev_layer, kernel, [1, 1, 1, 1, 1], padding='SAME')
        biases = _bias_variable('biases', [out_filters])
        bias = tf.nn.bias_add(conv, biases)
        conv1 = tf.nn.relu(bias, name=scope.name)

        prev_layer = conv1
        in_filters = out_filters

    pool1 = tf.nn.max_pool3d(prev_layer, ksize=[1, 3, 3, 3, 1], strides=[1, 2, 2, 2, 1], padding='SAME')
    norm1 = pool1  # tf.nn.lrn(pool1, 4, bias=1.0, alpha=0.001 / 9.0, beta = 0.75, name='norm1')

    prev_layer = norm1

    with tf.variable_scope('conv2') as scope:
        out_filters = 32
        kernel = _weight_variable('weights', [5, 5, 5, in_filters, out_filters])
        conv = tf.nn.conv3d(prev_layer, kernel, [1, 1, 1, 1, 1], padding='SAME')
        biases = _bias_variable('biases', [out_filters])
        bias = tf.nn.bias_add(conv, biases)
        conv2 = tf.nn.relu(bias, name=scope.name)

        prev_layer = conv2
        in_filters = out_filters

    # normalize prev_layer here
    prev_layer = tf.nn.max_pool3d(prev_layer, ksize=[1, 3, 3, 3, 1], strides=[1, 2, 2, 2, 1], padding='SAME')

    with tf.variable_scope('conv3_1') as scope:
        out_filters = 64
        kernel = _weight_variable('weights', [5, 5, 5, in_filters, out_filters])
        conv = tf.nn.conv3d(prev_layer, kernel, [1, 1, 1, 1, 1], padding='SAME')
        biases = _bias_variable('biases', [out_filters])
        bias = tf.nn.bias_add(conv, biases)
        prev_layer = tf.nn.relu(bias, name=scope.name)
        in_filters = out_filters

    with tf.variable_scope('conv3_2') as scope:
        out_filters = 64
        kernel = _weight_variable('weights', [5, 5, 5, in_filters, out_filters])
        conv = tf.nn.conv3d(prev_layer, kernel, [1, 1, 1, 1, 1], padding='SAME')
        biases = _bias_variable('biases', [out_filters])
        bias = tf.nn.bias_add(conv, biases)
        prev_layer = tf.nn.relu(bias, name=scope.name)
        in_filters = out_filters

    with tf.variable_scope('conv3_3') as scope:
        out_filters = 32
        kernel = _weight_variable('weights', [5, 5, 5, in_filters, out_filters])
        conv = tf.nn.conv3d(prev_layer, kernel, [1, 1, 1, 1, 1], padding='SAME')
        biases = _bias_variable('biases', [out_filters])
        bias = tf.nn.bias_add(conv, biases)
        prev_layer = tf.nn.relu(bias, name=scope.name)
        in_filters = out_filters

    # normalize prev_layer here
    prev_layer = tf.nn.max_pool3d(prev_layer, ksize=[1, 3, 3, 3, 1], strides=[1, 2, 2, 2, 1], padding='SAME')


    with tf.variable_scope('local3') as scope:
        dim = np.prod(prev_layer.get_shape().as_list()[1:])
        prev_layer_flat = tf.reshape(prev_layer, [-1, dim])
        weights = _weight_variable('weights', [dim, FC_SIZE])
        biases = _bias_variable('biases', [FC_SIZE])
        local3 = tf.nn.relu(tf.matmul(prev_layer_flat, weights) + biases, name=scope.name)

    prev_layer = local3

    with tf.variable_scope('local4') as scope:
        dim = np.prod(prev_layer.get_shape().as_list()[1:])
        prev_layer_flat = tf.reshape(prev_layer, [-1, dim])
        weights = _weight_variable('weights', [dim, FC_SIZE])
        biases = _bias_variable('biases', [FC_SIZE])
        local4 = tf.nn.relu(tf.matmul(prev_layer_flat, weights) + biases, name=scope.name)

    prev_layer = local4

    with tf.variable_scope('softmax_linear') as scope:
        dim = np.prod(prev_layer.get_shape().as_list()[1:])
        weights = _weight_variable('weights', [dim, dataconfig.num_classes])
        biases = _bias_variable('biases', [dataconfig.num_classes])
        softmax_linear = tf.add(tf.matmul(prev_layer, weights), biases, name=scope.name)

    return softmax_linear


def loss(logits, labels):
    cross_entropy = tf.nn.softmax_cross_entropy_with_logits(
        logits, labels, name='cross_entropy_per_example')

    return tf.reduce_mean(cross_entropy, name='xentropy_mean')

Hi dansileshi. I was wondering if you had the rest of the code that you used to make this run. I'm trying to adapt this into a demo 3D CNN that will classify weather there is a sphere or a cube in a set of synthetic 3D images I made. Specifically, I'm wondering what trainer you used and how to connect the inference and loss to the trainer and run it on a 4D matrix containing the 3D images and an array of labels.
Thanks!

Hi dansileshi, I am also interested in what @rceballos98 is asking. I am also interested in running a similar kind of classification. @rceballos98, what are you using the demo for, sounds cool! I'm just trying to see if I can do better than just 2D image classification by taking advantage of point cloud data.