Add atrous_conv2d_transpose python function

[guotong1988]

Issue description : #4668

This is the first commit of the issue.
I will commit the unit test later.
I will also add the comment for the method.

Thank you for any advice.

Now I test it with:

# Input, output: [batch, height, width, depth]
x_image = tf.placeholder(tf.float32,shape=[1])
x = tf.reshape(x_image,[1,1,1,1])

#Filter: W [kernel_height, kernel_width, output_depth, input_depth]
W_cpu = np.array([[1,-1,1],[1,1,1],[-1,1,-1]],dtype=np.float32)
W = tf.Variable(W_cpu)
W = tf.reshape(W, [3,3,1,1])

strides=[1, 1, 1, 1]
padding='VALID'

y = tf.nn.atrous_conv2d_transpose(x, W, [1,5,5,1], 2, strides, padding)

x_data = np.array([1],dtype=np.float32)
with tf.Session() as sess:
    init = tf.initialize_all_variables()
    sess.run(init)

    x = (sess.run(x, feed_dict={x_image: x_data}))
    W = (sess.run(W, feed_dict={x_image: x_data}))
    y = (sess.run(y, feed_dict={x_image: x_data}))

    print "The shape of x:\t", x.shape, ",\t and the x.reshape(1) is :"
    print x.reshape(1)
    print ""

    print "The shape of x:\t", W.shape, ",\t and the W.reshape(3,3) is :"
    print W.reshape(3,3)
    print ""

    print "The shape of y:\t", y.shape, ",\t and the y.reshape(5,5) is :"
    print y.reshape(5,5)
    print ""

and

  def testAtrousConv2DTransposeSingleStride(self):
    with self.test_session():
      strides = [1, 1, 1, 1]

      # Input, output: [batch, height, width, depth]
      x_shape = [2, 6, 4, 3]
      y_shape = [2, 6, 4, 2]

      # Filter: [kernel_height, kernel_width, output_depth, input_depth]
      f_shape = [3, 3, 2, 3]

      x = tf.constant(1.0, shape=x_shape, name="x", dtype=tf.float32)
      f = tf.constant(1.0, shape=f_shape, name="filter", dtype=tf.float32)

      output = tf.nn.atrous_conv2d_transpose(x, f, y_shape, 2, strides=strides,
                                      padding="SAME")

      value = output.eval()
      print(value)

[tensorflow-jenkins]

Can one of the admins verify this patch?

[mention-bot]

@guotong1988, thanks for your PR! By analyzing the history of the files in this pull request, we identified @tensorflower-gardener, @keveman and @yuefengz to be potential reviewers.

[vrv]

Looks nice at a high level! Let's add the tests in the same PR.

[guotong1988]

I'm working.

[gpapan]

Thanks @guotong1988 . Please make sure the test to check that both the output tensor shape and values are correct.

[sbrodehl]

What's the current status here @guotong1988 ?
Need some help? What need's to be done to get this merged?

[guotong1988]

I got a little busy to change my job in the past three weeks. And I just finished it.
I am working on it now.

[sbrodehl]

Nice! Ping me if you need help.

[guotong1988]

@sbrodehl In fact I'm not sure that my only test case can cover that much . Please provide more test case, if you have. Thank you .

[guotong1988]

I write another example , and I think my commit may be right.

import numpy as np
import tensorflow as tf

# Input, output: [batch, height, width, depth]
x_image = tf.placeholder(tf.float32,shape=[2,2])
x = tf.reshape(x_image,[1,2,2,1])

#Filter: W [kernel_height, kernel_width, output_depth, input_depth]
W_cpu = np.array([[1,-1,1],[1,1,1],[-1,1,-1]],dtype=np.float32)
W = tf.Variable(W_cpu)
W = tf.reshape(W, [3,3,1,1])

strides=[1, 1, 1, 1]
padding='VALID'

y = tf.nn.atrous_conv2d_transpose(x, W, [1,5,5,1], 2, strides, padding)

x_data = np.array([[1,1],[1,1]],dtype=np.float32)
with tf.Session() as sess:
    init = tf.initialize_all_variables()
    sess.run(init)

    x = (sess.run(x, feed_dict={x_image: x_data}))
    W = (sess.run(W, feed_dict={x_image: x_data}))
    y = (sess.run(y, feed_dict={x_image: x_data}))

    print "The shape of x:\t", x.shape, ",\t and the x.reshape(2,2) is :"
    print x.reshape([2,2])
    print ""

    print "The shape of x:\t", W.shape, ",\t and the W.reshape(3,3) is :"
    print W.reshape(3,3)
    print ""

    print "The shape of y:\t", y.shape, ",\t and the y.reshape(6,6) is :"
    print y.reshape(6,6)
    print ""

[guotong1988]

Now that I'm sure that my commit is right , because I write this example which is similar to the above one.

import numpy as np
import tensorflow as tf

# Input, output: [batch, height, width, depth]
x_image = tf.placeholder(tf.float32,shape=[2,3])
x = tf.reshape(x_image,[1,2,3,1])

#Filter: W [kernel_height, kernel_width, output_depth, input_depth]
W_cpu = np.array([[1,-1,1],[1,1,1],[-1,1,-1]],dtype=np.float32)
W = tf.Variable(W_cpu)
W = tf.reshape(W, [3,3,1,1])

strides=[1, 1, 1, 1]
padding='VALID'

y = tf.nn.atrous_conv2d_transpose(x, W, [1,5,5,1], 2, strides, padding)

x_data = np.array([[1,1,1],[1,1,1]],dtype=np.float32)
with tf.Session() as sess:
    init = tf.initialize_all_variables()
    sess.run(init)

    x = (sess.run(x, feed_dict={x_image: x_data}))
    W = (sess.run(W, feed_dict={x_image: x_data}))
    y = (sess.run(y, feed_dict={x_image: x_data}))

    print "The shape of x:\t", x.shape, ",\t and the x.reshape(2,2) is :"
    print x.reshape([2,3])
    print ""

    print "The shape of x:\t", W.shape, ",\t and the W.reshape(3,3) is :"
    print W.reshape(3,3)
    print ""

    print "The shape of y:\t", y.shape, ",\t and the y.reshape(6,7) is :"
    print y.reshape(6,7)
    print ""

[guotong1988]

Then I think I could start to finish the unit test.

[Fenugreek]

I tried running your code on MNIST. The output shape looks wrong to me:

        print(inputs.get_shape(), W.get_shape())
        outputs = atrous_conv2d_transpose(inputs, W, [100, 28, 28, 1],
                                          2, [1,1,1,1], padding='SAME')
        print(outputs.get_shape())

gives:

(100, 28, 28, 8) (5, 5, 1, 8)
(100, 36, 36, 1)

I was expecting the output shape, i.e. the last line above, to read (100, 28, 28, 1).

[guotong1988]

@Fenugreek Yes , I fixed it.

[vrv]

@gpapan just sent me a full PR for implementing the transpose, btw.

[guotong1988]

@vrv I will finish the unit test.

[vrv]

@gpapan, since you wrote the internal one, I'll let you decide which one we should accept :)

(based on correctness / merits, of course.)

[guotong1988]

I nearly finish. But the code need to review and edit, of course.

[guotong1988]

Thank you for your code. I learn a lot. Close this PR please.

[gpapan]

Thanks @guotong1988 <https://github.com/guotong1988>! @vrv: Can you please close #4668 and #5300?

…

On Mon, Dec 5, 2016 at 12:40 AM, 郭同jet · 耐心 ***@***.***> wrote: Thank you for your code. I learn a lot. Close this PR please. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#5300 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AF8Y_V16SQBnwOD5zAEWa-thMmHqWVh4ks5rE836gaJpZM4Kk04b> .

-- Best, George