Problems about the code

[SMZCC]

Dear @bilylee

Hello

Thanks for your contributions :tada:

I have encountered a problem in the code which lies in line .

• my problem is that how does this function work ?

• to understand it , I have written a code snippet(mimic your function) as the following:

  
  def get_mean(values):
      with tf.name_scope('mean'):
          mean, update_op = tf.metrics.mean(values)  # values = [1, 2, 3, 4, 5]
          with tf.control_dependencies([update_op]):
              mean = tf.identity(mean)
          return mean
  
  g = tf.Graph()
  g.as_default()
  values = tf.convert_to_tensor([1, 2, 3, 4, 5], dtype=tf.float32)
  mean= get_mean(values)
  init_global_variables = tf.global_variables_initializer()
  init_local_variables = tf.local_variables_initializer()
  g.finalize()
  
  sess = tf.InteractiveSession(graph=tf.get_default_graph())
  sess.run(init_global_variables)
  sess.run(init_local_variables)
  print sess.run(values)
  for i in range(5):
      print 'mean:', sess.run(mean)

out:

[1. 2. 3. 4. 5.]

mean: 0.0

mean: 6.0

mean: 4.5

mean: 4.0

mean: 3.75

not corresponding to my expectation

expected values are:

[1. 2. 3. 4. 5.]

mean: 3.0

mean: 3.0

mean: 3.0

mean: 3.0

mean: 3.0

reasons as following：

iter: 1

total = 1+2+3+4+5 = 15

count = 1+1+1+1+1 = 5

mean = 15 / 5 = 3

iter: 2

total = 15 + 15 = 30

count = 5 + 5 = 10

mean = 30 / 10 = 3

...

- Would you please told me I did anything wrong ? And how to understand it ?
#### Thanks in advance 
#### Best regards 

[pmixer]

Hi, @SMZCC Here's what I got, different from yours, as a personal opinion, the function aims to calculate how many response maps has a peak outside the manually labels center which makes it simple and intuitive, your question may lay on mechanisms of tensorflow. As for this issue, I suggest reading the official document and report your experiment settings here(os, tf version, python version, etc.), while I also got 0.0 in the first round (tf=1.4,cpu, py=2.7), I will be better to double check the step in execution.

[bilylee]

Sorry for the late reply, I have been very busy lately : (

It seems that the graph is evaluated differently when using tf.InteractiveSession. Change it to tf.Session, you should get expected outputs.

[SMZCC]

Wow ! You are right ! Thank you very much ! @bilylee :+1: of course for the warm hearts of @PeterHuang2015

• Do you know where this line plays a role which looks so abrupt ?
  • Why we need this parameter ?

    Best regards :muscle:

[bilylee]

This parameter is the stride of the feature extractor. It is used for computing ground truth labels during training and target positions during testing.

I used convolutional_alexnet.stride = 8 in the early days of implementing SiamFC. Its usage is now replaced by https://github.com/bilylee/SiamFC-TensorFlow/blob/17f51563f081281fc7d966602d393715e7ecb469/configuration.py#L36

You can safely delete this line without affecting anything.

[SMZCC]

You can safely delete this line without affecting anything.

I have tried and found nothing changed , which made me confused, but now , it's OK.Thanks a lot !

Good job :+1: :muscle: @bilylee

Best regards