remove large file from commit

 git filter-branch --tree-filter 'rm -rf file_path' HEAD

markov process material

 https://www.stat.auckland.ac.nz/~fewster/325/notes/325book.pdf

yolov3 material

https://github.com/YunYang1994/tensorflow-yolov3

and his blog 

https://yunyang1994.gitee.io/2018/12/28/YOLOv3/

code analysis:

https://blog.csdn.net/sxlsxl119/article/details/103027960

Sample Code to use Keras Model

https://keras.io/api/applications/

Remove Sensitive Environment Variable / File That is too Big from Remote Repo

In case mistakenly pushed a large file/sensitive environment data and git rm --cached does not help (pushed to the remote repo but just untracked locally):

git filter-branch --force --index-filter \
  "git rm --cached --ignore-unmatch <path to your file>" \
  --prune-empty --tag-name-filter cat -- --all

(cd to the top level of the repo first), add -r flag if you want to remove the whole directory. REF: 

https://docs.github.com/.../removing-sensitive-data-from... 

This error may subsequently follow: fatal: refusing to merge unrelated histories Then

git pull origin the-remote-branch --allow-unrelated-histories

and resolve conflicts.

Examine Output Size in Tensorflow

When we are uncertain the output size of our tensor processed by some layer, we can go through the following:

x = tf.constant([[1, 1., 1., 2., 3.],
                 [1, 1., 4., 5., 6.],
                 [1, 1., 7., 8., 9.],
                 [1, 1., 7., 8., 9.],
                 [1, 1., 7., 8., 9.]])

x = tf.reshape(x, [1, 5, 5, 1])

print(MaxPool2D((5, 5), strides=(2, 2),  padding="same")(x))
print(math.ceil(5/2))

which yields

print(MaxPool2D((5, 5), strides=(2, 2),  padding="same")(x))
tf.Tensor(
[[[[7.]
   [9.]
   [9.]]

  [[7.]
   [9.]
   [9.]]

  [[7.]
   [9.]
   [9.]]]], shape=(1, 3, 3, 1), dtype=float32)

3

For layer that has training weight, we may try the following for testing:

model = Conv2D(3, (3, 3), strides=(2, 2), padding="same", kernel_initializer=tf.constant_initializer(1.))


x = tf.constant([[1., 2., 3., 4., 5.],
                 [1., 2., 3., 4., 5.],
                 [1., 2., 3., 4., 5.],
                 [1., 2., 3., 4., 5.],
                 [1., 2., 3., 4., 5.]])

x = tf.reshape(x, (1, 5, 5, 1))
print(model(x))

which yields

x = tf.constant([[1., 2., 3., 4., 5.],...
tf.Tensor(
[[[[ 6.  6.  6.]
   [18. 18. 18.]
   [18. 18. 18.]]

  [[ 9.  9.  9.]
   [27. 27. 27.]
   [27. 27. 27.]]

  [[ 6.  6.  6.]
   [18. 18. 18.]
   [18. 18. 18.]]]], shape=(1, 3, 3, 3), dtype=float32)

In fact it can be proved in both MaxPooling2D and Conv2D that if stride $=s$ and padding$=$same, then 

\[\text{output_width} = \left\lfloor\frac{\text{input_width}-1}{s}\right\rfloor + 1 = \left\lceil\frac{\text{input_width}}{s}\right\rceil\]

The last equality deserves a proof as it is not highly trivial:

Fact. For any positive intergers $w,s$, we have \[ \left\lfloor \frac{w-1}{s}\right\rfloor + 1 = \left\lceil \frac{w}{s}\right\rceil. \] Proof. We do case by case study. If $w=ks$ for some positive $k\in \N$, then \[\text{LHS} = \left\lfloor k - \frac{1}{s}\right\rfloor +1 = (k-1)+1=k = \lceil k\rceil = \text{RHS}. \] When $w=ks+j$, for some $k\in\N$ and $j\in \N \cap (0, s)$, then \[ \text{LHS} = \left\lfloor k+\frac{j-1}{s}\right\rfloor + 1 = k+1 = \left\lceil k+\frac{j}{s}\right\rceil = \left\lceil \frac{ks+j}{s}\right\rceil = \left\lceil\frac{w}{s}\right\rceil=\text{RHS}.\qed \]

conda virtual environment command

conda create --name tensorflow python=3.7
conda env remove --name tensorflow

conda env export --name ENVNAME > envname.yml
conda env create --file envname.yml