Pfthb

I'm trying to implement unpool in tensorflow with tf.scatter_add, but I meet a strange bug, here is my code:

import tensorflow as tf
import numpy as np
import random

tf.reset_default_graph()

mat = list(range(64))
random.shuffle(mat)
mat = np.array(mat)
mat = np.reshape(mat, [1,8,8,1])
M = tf.constant(mat, dtype=tf.float32)
pool1, argmax1 = tf.nn.max_pool_with_argmax(M, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
pool2, argmax2 = tf.nn.max_pool_with_argmax(pool1, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
pool3, argmax3 = tf.nn.max_pool_with_argmax(pool2, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')


def unpool(x, argmax, strides, unpool_shape=None, batch_size=None, name='unpool'):
 x_shape = x.get_shape().as_list()
 argmax_shape = argmax.get_shape().as_list()
 assert not(x_shape[0] is None and batch_size is None), "must input batch_size if number of batch is alterable"
 if x_shape[0] is None:
 x_shape[0] = batch_size
 if argmax_shape[0] is None:
 argmax_shape[0] = x_shape[0]
 if unpool_shape is None:
 unpool_shape = [x_shape[i] * strides[i] for i in range(4)]
 x_unpool = tf.get_variable(name=name, shape=[np.prod(unpool_shape)], initializer=tf.zeros_initializer(), trainable=False)
 argmax = tf.cast(argmax, tf.int32)
 argmax = tf.reshape(argmax, [np.prod(argmax_shape)])
 x = tf.reshape(x, [np.prod(argmax.get_shape().as_list())])
 x_unpool = tf.scatter_add(x_unpool , argmax, x)
 x_unpool = tf.reshape(x_unpool , unpool_shape)
 return x_unpool 


unpool2 = unpool(pool3, argmax3, strides=[1,2,2,1], name='unpool3')
unpool1 = unpool(unpool2, argmax2, strides=[1,2,2,1], name='unpool2')
unpool0 = unpool(unpool1, argmax1, strides=[1,2,2,1], name='unpool1')


with tf.Session() as sess:
 sess.run(tf.global_variables_initializer())
 mat_out = mat[:,:,:,0]
 pool1_out = sess.run(pool1)[0,:,:,0]
 pool2_out = sess.run(pool2)[0,:,:,0]
 pool3_out = sess.run(pool3)[0,:,:,0]
 argmax1_out = sess.run(argmax1)[0,:,:,0]
 argmax2_out = sess.run(argmax2)[0,:,:,0]
 argmax3_out = sess.run(argmax3)[0,:,:,0]
 unpool2_out = sess.run(unpool2)[0,:,:,0]
 unpool1_out = sess.run(unpool1)[0,:,:,0]
 unpool0_out = sess.run(unpool0)[0,:,:,0]
 print(unpool2_out)
 print(unpool1_out)
 print(unpool0_out)

output:

[[ 0. 0.]
 [ 0. 63.]]
[[ 0. 0. 0. 0.]
 [ 0. 0. 0. 0.]
 [ 0. 0. 126. 0.]
 [ 0. 0. 0. 0.]]
[[ 0. 0. 0. 0. 0. 0. 0. 0.]
 [ 0. 0. 0. 0. 0. 0. 0. 0.]
 [ 0. 0. 0. 0. 0. 0. 0. 0.]
 [ 0. 0. 0. 0. 0. 0. 0. 0.]
 [ 0. 0. 0. 0. 0. 0. 0. 0.]
 [ 0. 0. 0. 0. 315. 0. 0. 0.]
 [ 0. 0. 0. 0. 0. 0. 0. 0.]
 [ 0. 0. 0. 0. 0. 0. 0. 0.]]

The location is right, but the value is wrong. unpool2 is right, unpool1 is double of expected value, and unpool2 is quintuple of expected value. I don't know what's wrong, can anyone tell me how to fix this bug?

Great thanks in advance.

In fact, the answer is simple. For convenience, I rename some variables, look this code:

def unpool(x, argmax, strides, unpool_shape=None, batch_size=None, name='unpool'):
 x_shape = x.get_shape().as_list()
 argmax_shape = argmax.get_shape().as_list()
 assert not(x_shape[0] is None and batch_size is None), "must input batch_size if number of batch is alterable"
 if x_shape[0] is None:
 x_shape[0] = batch_size
 if argmax_shape[0] is None:
 argmax_shape[0] = x_shape[0]
 if unpool_shape is None:
 unpool_shape = [x_shape[i] * strides[i] for i in range(4)]
 x_unpool = tf.get_variable(name=name, shape=[np.prod(unpool_shape)], initializer=tf.zeros_initializer(), trainable=False)
 argmax = tf.cast(argmax, tf.int32)
 argmax = tf.reshape(argmax, [np.prod(argmax_shape)])
 x = tf.reshape(x, [np.prod(argmax.get_shape().as_list())])
 x_unpool_add = tf.scatter_add(x_unpool , argmax, x)
 x_unpool_reshape = tf.reshape(x_unpool_add , unpool_shape)
 return x_unpool_reshape 

x_unpool_add is a op of tf.scatter_add, everytime we compute x_unpool_reshape, x_unpool_add will be called. So x_unpool will add x twice if we compute unpool2 twice. In my origin code, I compute unpool0, unpool1, unpool2 in order, x_unpool_add of unpool1 is called firstly, then when we compute unpool2, because of we need compute unpool1, x_unpool_add will be called again, so it's equal to call x_unpool_add twice, the value is wrong. If we compute unpool2 directly, we will get right result. So replacling tf.scatter_add with tf.scatter_update can avoid this bug.

This code can reproducible this intuitively:

import tensorflow as tf

t1 = tf.get_variable(name='t1', shape=[1], dtype=tf.float32, initializer=tf.zeros_initializer())
t2 = tf.get_variable(name='t2', shape=[1], dtype=tf.float32, initializer=tf.zeros_initializer())
d = tf.scatter_add(t1, [0], [1])
e = tf.scatter_add(t2, [0], d)

with tf.Session() as sess:
 sess.run(tf.global_variables_initializer())
 d_out1 = sess.run(d)
 d_out2 = sess.run(d)
 e_out = sess.run(e)
 print(d_out1)
 print(d_out2)
 print(e_out)

output:

[1.]
[2.]
[3.]

Use tf.scatter_update can avoid this.

import tensorflow as tf
import numpy as np
import random

tf.reset_default_graph()

mat = list(range(64))
random.shuffle(mat)
mat = np.array(mat)
mat = np.reshape(mat, [1,8,8,1])
M = tf.constant(mat, dtype=tf.float32)
pool1, argmax1 = tf.nn.max_pool_with_argmax(M, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
pool2, argmax2 = tf.nn.max_pool_with_argmax(pool1, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
pool3, argmax3 = tf.nn.max_pool_with_argmax(pool2, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')


def unpool(x, argmax, strides, unpool_shape=None, batch_size=None, name='unpool'):
 x_shape = x.get_shape().as_list()
 argmax_shape = argmax.get_shape().as_list()
 assert not(x_shape[0] is None and batch_size is None), "must input batch_size if number of batch is alterable"
 if x_shape[0] is None:
 x_shape[0] = batch_size
 if argmax_shape[0] is None:
 argmax_shape[0] = x_shape[0]
 if unpool_shape is None:
 unpool_shape = [x_shape[i] * strides[i] for i in range(4)]
 unpool = tf.get_variable(name=name, shape=[np.prod(unpool_shape)], initializer=tf.zeros_initializer(), trainable=False)
 argmax = tf.cast(argmax, tf.int32)
 argmax = tf.reshape(argmax, [np.prod(argmax_shape)])
 x = tf.reshape(x, [np.prod(argmax.get_shape().as_list())])
 unpool = tf.scatter_update(unpool, argmax, x)
 unpool = tf.reshape(unpool, unpool_shape)
 return unpool


unpool2 = unpool(pool3, argmax3, strides=[1,2,2,1], name='unpool3')
unpool1 = unpool(unpool2, argmax2, strides=[1,2,2,1], name='unpool2')
unpool0 = unpool(unpool1, argmax1, strides=[1,2,2,1], name='unpool1')


with tf.Session() as sess:
 sess.run(tf.global_variables_initializer())
 mat_out = mat[:,:,:,0]
 pool1_out = sess.run(pool1)[0,:,:,0]
 pool2_out = sess.run(pool2)[0,:,:,0]
 pool3_out = sess.run(pool3)[0,:,:,0]
 argmax1_out = sess.run(argmax1)[0,:,:,0]
 argmax2_out = sess.run(argmax2)[0,:,:,0]
 argmax3_out = sess.run(argmax3)[0,:,:,0]
 unpool2_out = sess.run(unpool2)[0,:,:,0]
 unpool1_out = sess.run(unpool1)[0,:,:,0]
 unpool0_out = sess.run(unpool0)[0,:,:,0]
 print(unpool2_out)
 print(unpool1_out)
 print(unpool0_out)

output:

[[ 0. 0.]
 [ 0. 63.]]
[[ 0. 0. 0. 0.]
 [ 0. 0. 0. 0.]
 [ 0. 0. 0. 63.]
 [ 0. 0. 0. 0.]]
[[ 0. 0. 0. 0. 0. 0. 0. 0.]
 [ 0. 0. 0. 0. 0. 0. 0. 0.]
 [ 0. 0. 0. 0. 0. 0. 0. 0.]
 [ 0. 0. 0. 0. 0. 0. 0. 0.]
 [ 0. 0. 0. 0. 0. 0. 0. 0.]
 [ 0. 0. 0. 0. 0. 0. 0. 63.]
 [ 0. 0. 0. 0. 0. 0. 0. 0.]
 [ 0. 0. 0. 0. 0. 0. 0. 0.]]