python, C++, Caffe, TensorFlow

1. numpy.argsort()函数是将x中的元素从小到大排列，提取其对应的index(索引号)

>>> x = np.array([3, 1, 2])
>>> np.argsort(x) #按升序排列
array([1, 2, 0])
>>> np.argsort(-x) #按降序排列
array([0, 2, 1])

2. C++ constant string initialization - “Incomplete type not allowed” error:
   You need to explicitly specify the size of the arrays (if you want to use a C-style array) in your class initialization:

   Cause: When you declare an array without an explicit size, the compiler thinks that it is a flexible array, which is a C-style only feature. It seems to be allowed in C++ as a non-standard g++ extension.

class DataFilter
{
   // ...
   float firstVect[3] = {0.0f,0.0f,0.0f};
   int initialized = 0;
   int normalNumberOfDigits[3]= {0,0,0};
   // ...
};

3. vector of C-style array: first call [] then vector.

	const std::vector<std::pair<int, int>> mapIdx[2] = {
		{	//COCO
			{ 31,32 },{ 39,40 },{ 33,34 },{ 35,36 },{ 41,42 },{ 43,44 },
			{ 19,20 },{ 21,22 },{ 23,24 },{ 25,26 },{ 27,28 },{ 29,30 },
			{ 47,48 },{ 49,50 },{ 53,54 },{ 51,52 },{ 55,56 },{ 37,38 },
			{ 45,46 }
		},
		{	// BODY25
		}
	};
	//to get coco's mapIdx -> mapIdx[0][17]
	
	std::vector<std::pair<int, int>> pr = {
			{0,1}, {14,15}, {22,23}, {16,17}, {18,19},
			{24,25}, {26,27}, {6,7}, {2,3}, {4,5},
			{8,9}, {10,11}, {12,13}, {30,31}, {32,33},
			{36,37}, {34,35}, {38,39}, {20,21}, {28,29},
			{40,41}, {42,43}, {44,45},
			{46,47}, {48,49}, {50,51} 
			};
	for (int i = 0; i < pr.size(); i++)
	{
		std::pair<int, int> tmp = pr[i];
		std::cout << " {" << tmp.first + 26 << ", " << tmp.second + 26 << "},";
	}

4. mobile net TensorFlow 转 Caffe 要注意stride=2而且filter size=3时的情况，因为Caffe是symmetric pad，而TensorFlow是asymmetric pad。所以要使weight完全对应，有两种方法，一种pad 0 到 weights 的左边和上边，另一种是pad 0 到feature map的左边和上边，然后加2个slice layer去crop掉。

TensorFlow:padding:'same'(zero pad) or 'valid'(no pad)
padding = ‘VALID’
out_height = ceil((in_height – filter_height + 1) / strides[1]) （结果向上取整）
padding = ‘SAME'
out_height = ceil(in_height  / strides[1]) （结果向上取整）
Caffe
out_height  = floor((in_height - filter_height + 2pad)/strides) +1

5. C++ 创建函数，：后面，赋值变量时，可简写成（）

//Definition for singly-linked list.
 struct ListNode {
     int val;
     ListNode *next;
     ListNode(int x) : val(x), next(NULL) {}
 };

6. a->b 的含義是 (*a).b
   a 是指针，b是a指向的object的成员。（*a）则是指向的object。
   new 申请内存空间返回地址（malloc() 是 C standard library）。*p是p地址的值，prehead是值；p是pointer，&prehead是prehead的地址。
   直接使用&prehead时，&prehead是reference，必须初始化。
   太特么复杂了，记住别用&来做新建变量就行了…新建的都是reference

        // ListNode preHead = ListNode(0);
        // ListNode *p = &preHead;
        // ListNode *p = new ListNode(0);
        ListNode * p;
        p = new ListNode(0);
        ListNode &preHead = *p;
		cout << "p->next: " << p->next << "\n";
		cout << "(&preHead).next: " << (*p).next << "\n";
		cout << "preHead.next: " <<preHead.next << "\n";
		cout << "(&preHead)->next: " << (&preHead)->next << "\n";

7. int* a, b;
   你以为是int* a; int* b;
   但实际上是
   int* a; int b;

8. pip google module problem:
   真正缺的不是google module，而是protobuf module，所以：

>> python -c 'from caffe2.python import workspace; print(workspace.NumCudaDevices())'
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/home/icduser/anaconda3/lib/python3.7/site-packages/caffe2/python/__init__.py", line 2, in <module>
    from caffe2.proto import caffe2_pb2
  File "/home/icduser/anaconda3/lib/python3.7/site-packages/caffe2/proto/__init__.py", line 11, in <module>
    from caffe2.proto import caffe2_pb2, metanet_pb2, torch_pb2
  File "/home/icduser/anaconda3/lib/python3.7/site-packages/caffe2/proto/caffe2_pb2.py", line 6, in <module>
    from google.protobuf.internal import enum_type_wrapper
ModuleNotFoundError: No module named 'google'
>>conda install protobuf

9. Android assets problem:
   新版的Android assets folder需要右击new dir，location默认app/src/main/assets，否则会有java.io.FileNotFoundException
10. VS2015 加 CUDA 9.0 加 Python35 版的 Caffe 用 cmake 编译，提示：找不到boost lib，
    更改cmake的folder里的Dependencies.cmake：把所有boost的版本都改成1.61。
    原因：pre-build的1.55版boost，不支持 VS2015 加 CUDA 9.0 加 Python35，改成用pre-build的1.61版就OK。