Tensorflow中单应性的实现 [英] Implementation of homography in Tensorflow

问题描述

我正在尝试在 TensorFlow 中实现一个学习预测单应性的网络(

I am trying to implement a network in TensorFlow that learns to predict homography (Wiki). My network will output a 4-dimensional vector, which will be used to translate 2 images horizontally and vertically. Then, I calculate the error with a ground truth using a central crop (tf.image.central_crop)of these warped images. I tried implementing it, and for the translation part, I used tf.contrib.image.translate. But, the gradients are not flowing to the variables of the network. How can I fix this problem? This is the error that I am getting:

ValueError: No gradients provided for any variable, check your graph for ops that do not support gradients, between variables ["<tf.Variable 'get_tr/w1:0' shape=(3, 3, 6, 64) dtype=float32_ref>", "<tf.Variable 'get_tr/b1:0' shape=(20, 298, 298, 64) dtype=float32_ref>", "<tf.Variable 'get_tr/w2:0' shape=(3, 3, 64, 64) dtype=float32_ref>", "<tf.Variable 'get_tr/b2:0' shape=(20, 296, 296, 64) dtype=float32_ref>", "<tf.Variable 'get_tr/w3:0' shape=(3, 3, 64, 128) dtype=float32_ref>", "<tf.Variable 'get_tr/b3:0' shape=(20, 147, 147, 128) dtype=float32_ref>", "<tf.Variable 'get_tr/w4:0' shape=(3, 3, 128, 128) dtype=float32_ref>", "<tf.Variable 'get_tr/b4:0' shape=(20, 73, 73, 128) dtype=float32_ref>", "<tf.Variable 'get_tr/w5:0' shape=(5, 5, 128, 128) dtype=float32_ref>", "<tf.Variable 'get_tr/fc1/kernel:0' shape=(609408, 512) dtype=float32_ref>", "<tf.Variable 'get_tr/fc1/bias:0' shape=(512,) dtype=float32_ref>", "<tf.Variable 'get_tr/fc2/kernel:0' shape=(512, 1024) dtype=float32_ref>", "<tf.Variable 'get_tr/fc2/bias:0' shape=(1024,) dtype=float32_ref>", "<tf.Variable 'get_tr/fc_o/kernel:0' shape=(1024, 4) dtype=float32_ref>", "<tf.Variable 'get_tr/fc_o/bias:0' shape=(4,) dtype=float32_ref>"] and loss Tensor("mean_squared_error/value:0", shape=(), dtype=float32).

Here is the code for getting this transformation vector.

def get_transform_vectors(self):
    # Start of transformation prediction network

    image = tf.concat((self.img_train_1, self.img_train_2), 3)

    with tf.variable_scope('get_tr'):
        w1 = tf.Variable(tf.truncated_normal(shape=[3, 3, 6, 64], stddev=0.1),
                    name='w1')
        conv1 = tf.nn.conv2d(image,
                                    filter=w1,
                                    strides=(1,1,1,1),
                                    padding="VALID",
                                    name='conv1')

        b1 = tf.Variable(tf.truncated_normal(shape=tf.shape(conv1), stddev=0.1),
                    name='b1')
        conv1_ = tf.nn.relu(conv1+b1, name='conv1_')

        w2 = tf.Variable(tf.truncated_normal(shape=[3, 3, 64, 64], stddev=0.1),
                    name='w2')
        conv2 = tf.nn.conv2d(conv1_,
                                    filter=w2,
                                    strides=(1,1,1,1),
                                    padding="VALID",
                                    name='conv2')
        b2 = tf.Variable(tf.truncated_normal(shape=tf.shape(conv2), stddev=0.1),
                    name='b2')
        conv2_ = tf.nn.relu(conv2+b2, name='conv2_')

        w3 = tf.Variable(tf.truncated_normal(shape=[3, 3, 64, 128], stddev=0.1),
                    name='w3')
        conv3 = tf.nn.conv2d(conv2_,
                                    filter=w3,
                                    strides=(1,2,2,1),
                                    padding="VALID",
                                    name='conv3')
        b3 = tf.Variable(tf.truncated_normal(shape=tf.shape(conv3), stddev=0.1),
                    name='b3')
        conv3_ = tf.nn.relu(conv3+b3, name='conv3_')

        w4 = tf.Variable(tf.truncated_normal(shape=[3, 3, 128, 128], stddev=0.1),
                    name='w4')
        conv4 = tf.nn.conv2d(conv3_,
                                    filter=w4,
                                    strides=(1,2,2,1),
                                    padding="VALID",
                                    name='conv4')

        b4 = tf.Variable(tf.truncated_normal(shape=tf.shape(conv4), stddev=0.1),
                    name='b4')
        conv4_ = tf.nn.relu(conv4+b4, name='conv4_')

        w5 = tf.Variable(tf.truncated_normal(shape=[5, 5, 128, 128], stddev=0.1),
                    name='w5')
        conv5 = tf.nn.conv2d(conv4_,
                                    filter=w5,
                                    strides=(1,1,1,1),
                                    padding="VALID",
                                    name='conv5')

        conv5_ = tf.contrib.layers.flatten(conv5)

        fc1 = tf.layers.dense(conv5_, 512, activation=tf.nn.relu, name='fc1')
        fc2 = tf.layers.dense(fc1, 1024, activation=tf.nn.relu, name='fc2')
        fc_o = tf.layers.dense(fc2, 4, name='fc_o')

    return fc_o

This is the code for translating the image and calculating the loss.

self.img_o = tf.contrib.image.translate(self.img_train_1,
                                            tf.cast(tf.reshape(self.tr_vector[:, 0:2], [self.batch_size,2]),
                                                    dtype=tf.float32)) +\
                 tf.contrib.image.translate(self.img_train_2,
                                            tf.cast(tf.reshape(self.tr_vector[:, 2:4], [self.batch_size,2]),
                                                    dtype=tf.float32))

self.loss = tf.losses.mean_squared_error(self.img_o, self.img_label)
self.optim = tf.train.AdamOptimizer().minimize(self.loss)

解决方案

Since you are only interested in optimizing translations, I came up with this function to perform a translation and have gradients for the vectors, using @tf.custom_gradient:

import tensorflow as tf

@tf.custom_gradient
def my_img_translate(imgs, translates):
    # Interpolation model has to be fixed due to limitations of tf.custom_gradient
    interpolation = 'NEAREST'
    imgs_translated = tf.contrib.image.translate(imgs, translates, interpolation=interpolation)
    def grad(img_translated_grads):
        translates_x = translates[:, 0]
        translates_y = translates[:, 1]
        translates_zero = tf.zeros_like(translates_x)
        # X gradients
        imgs_x_grad = (imgs[:, :, :-2] - imgs[:, :, 2:]) / 2
        imgs_x_grad = tf.concat([(imgs[:, :, :1] - imgs[:, :, 1:2]),
                                 imgs_x_grad,
                                 (imgs[:, :, -2:-1] - imgs[:, :, -1:])], axis=2)
        imgs_x_grad_translated = tf.contrib.image.translate(
            imgs_x_grad, tf.stack([translates_x, translates_zero], axis=1),
            interpolation=interpolation)
        translates_x_grad = tf.reduce_sum(img_translated_grads * imgs_x_grad_translated, axis=(1, 2, 3))
        # Y gradients
        imgs_y_grad = (imgs[:, :-2] - imgs[:, 2:]) / 2
        imgs_y_grad = tf.concat([(imgs[:, :1] - imgs[:, 1:2]),
                                 imgs_y_grad,
                                 (imgs[:, -2:-1] - imgs[:, -1:])], axis=1)
        imgs_y_grad_translated = tf.contrib.image.translate(
            imgs_y_grad, tf.stack([translates_zero, translates_y], axis=1),
            interpolation=interpolation)
        translates_y_grad = tf.reduce_sum(img_translated_grads * imgs_y_grad_translated, axis=(1, 2, 3))
        # Complete gradient
        translates_grad = tf.stack([translates_x_grad, translates_y_grad], axis=1)
        return None, translates_grad
    return imgs_translated, grad

Note that in this case I am not returning any gradient for the images, since these will not be optimized (but if you wanted that, you could in principle use the builtin gradient for the translation operation).

I tested this for a simple use case of translating an image so its center had the highest value:

import tensorflow as tf
import numpy as np

@tf.custom_gradient
def my_img_translate(imgs, translates):
    # Interpolation model has to be fixed due to limitations of tf.custom_gradient
    interpolation = 'NEAREST'
    imgs_translated = tf.contrib.image.translate(imgs, translates, interpolation=interpolation)
    def grad(img_translated_grads):
        translates_x = translates[:, 0]
        translates_y = translates[:, 1]
        translates_zero = tf.zeros_like(translates_x)
        # X gradients
        imgs_x_grad = (imgs[:, :, :-2] - imgs[:, :, 2:]) / 2
        imgs_x_grad = tf.concat([(imgs[:, :, :1] - imgs[:, :, 1:2]),
                                 imgs_x_grad,
                                 (imgs[:, :, -2:-1] - imgs[:, :, -1:])], axis=2)
        imgs_x_grad_translated = tf.contrib.image.translate(
            imgs_x_grad, tf.stack([translates_x, translates_zero], axis=1),
            interpolation=interpolation)
        translates_x_grad = tf.reduce_sum(img_translated_grads * imgs_x_grad_translated, axis=(1, 2, 3))
        # Y gradients
        imgs_y_grad = (imgs[:, :-2] - imgs[:, 2:]) / 2
        imgs_y_grad = tf.concat([(imgs[:, :1] - imgs[:, 1:2]),
                                 imgs_y_grad,
                                 (imgs[:, -2:-1] - imgs[:, -1:])], axis=1)
        imgs_y_grad_translated = tf.contrib.image.translate(
            imgs_y_grad, tf.stack([translates_zero, translates_y], axis=1),
            interpolation=interpolation)
        translates_y_grad = tf.reduce_sum(img_translated_grads * imgs_y_grad_translated, axis=(1, 2, 3))
        # Complete gradient
        translates_grad = tf.stack([translates_x_grad, translates_y_grad], axis=1)
        return None, translates_grad
    return imgs_translated, grad

# Test operations
imgs = tf.placeholder(tf.float32, [None, None, None, None])
translates = tf.Variable([0, 0], dtype=tf.float32)
translates_tiled = tf.tile(translates[tf.newaxis], (tf.shape(imgs)[0], 1))
imgs_translated = my_img_translate(imgs, translates_tiled)
imgs_midpoint = imgs_translated[:, tf.shape(imgs_translated)[1] // 2, tf.shape(imgs_translated)[2] // 2]
loss = -tf.reduce_sum(tf.square(imgs_midpoint))
train_op = tf.train.GradientDescentOptimizer(10).minimize(loss)
init_op = tf.global_variables_initializer()

with tf.Session() as sess:
    # Make test image
    xs, ys = np.meshgrid(np.linspace(-2, 1, 100), np.linspace(-1, 2, 75))
    test_img = (1 / (1 + np.square(xs))) * (1 / (1 + np.square(ys)))
    test_img /= np.max(test_img)
    test_img_batch = test_img[np.newaxis, :, :, np.newaxis]
    # Train
    sess.run(init_op)
    for _ in range(100):
        sess.run(train_op, feed_dict={imgs: test_img_batch})
    # Show result
    test_img_result = sess.run(imgs_translated, feed_dict={imgs: test_img_batch})[0, :, :, 0]
    plt.figure()
    plt.subplot(121)
    plt.imshow(test_img)
    plt.subplot(122)
    plt.imshow(test_img_result)

Result:

这篇关于Tensorflow中单应性的实现的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！