- Data structure + operation
- Data structure + algorithm
- Object (as an encapsulation of data structure and operations) + algorithm
- 将官方wheel安装包下载到本地： wget https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow-0.8.0-cp34-cp34m-linux_x86_64.whl
- 重命名wheel安装文件: mv ~/username/tensorflow-0.8.0-cp34-cp34m-linux_x86_64.whl ~/username/tensorflow-0.8.0-cp35-cp35m-linux_x86_64.whl
- 使用pip从命令行手动安装wheel: pip install ~/username/tensorflow-0.8.0-cp35-cp35m-linux_x86_64.whl
- GPU版TF的安装除了上述步骤外，还需要提前确认用于计算的GPU(显卡)是否有CUDA Compute Capability 3.0以上的运算能力，具体档位和显卡型号对应表请查阅下面的链接： CUDA Compute Capability 官方链接
- CUDA Driver的安装， CUDA Driver包含在CUDA Toolkit里，请直接安装CUDA Toolkit; 请参照Nvida官方的文档, 如果安装环境以前装过非官方的Nvida显卡驱动，需要手动卸载。CUDA Driver Instructions 官方链接。CUDA Toolkit (含Driver和工具)的安装请使用Standalone安装方式(即从runfile离线安装)
- Cudnn （GPU运算库）的安装：官方下载需要注册和审核，这里给出一个临时下载地址：Cudnn Library 百度网盘链接下载解压候使用如下命令安装Cudnn文件
tar xvzf cudnn-7.0-linux-x64-v4.0-prod.tgz sudo cp cuda/include/cudnn.h /usr/local/cuda/include sudo cp cuda/lib64/libcudnn* /usr/local/cuda/lib64 sudo cp cuda/lib64/libcudnn* /usr/lib/x86_64-linux-gnu/ sudo chmod a+r /usr/local/cuda/lib64/libcudnn* sudo chmod a+r /usr/lib/x86_64-linux-gnu/libcudnn*
a. Cannot find CUDA cuda.xx.so.4.5. (for example): sudo ldconfig /usr/local/cuda/lib64
(Update the cache file of linker targets so that linker can find shared libraries)
b. Cannot Open CUDNN cudnn.xx. (for example): sudo ldconfig /usr/local/cuda/lib64
lspci | grep -i nvidia
nvidia-smi -L 显示GPU硬件信息
nvidia-smi -a 显示GPU使用情况
nvidia-smi 显示GPU综合情况(含Current Running GPU Process)
nvidia-smi -L 命令显示的GPU编号不一定是GPU运算库配置文件里所需要的编号， 配置文件有时需要按GPU个数编号如0, 1, 2。
nvidia-smi -L 命令显示出来的GPU编号可以从4, 5, 6这些开始
nvidia-smi -a 显示GPU使用情况
Ambiguous occurs when multiple options are open for consideration.
Java and C++ has the main method as the sole entry point, python has no official definition or requirement for entry point. In python, it is common to use a script called main.py as the entry point of the program (the interpreter starts running the main.py script first).
Arithmetic Evaluation (Division)
Python3 will by default evaluate integral division in a float division manner ( 1/2 = 0.5). Java, C++ and Scala will evaluate integral division in an integral division manner ( 1/2 = 0 )
- Python's real equivalence of java package is called module (which is actually a .py file). Python's package is equivalent to Java's package of package (nested package). Unlike a Python package, a Java package (a namespace) needs to be explicitly declared. Python allows programming at the package level. In Java, package is just a namespace.
By default, in Java the import statement applies to class names in the package name space. Import is followed by the full name of a class (package.package.....class). However, the import statement in python applies to not only class names (from package.module import class), but also other variables defined in the module (functions, constants, etc). In addition, the import mechanism of python allows the import of only package name (while in case the module names can not be referenced via the package name if no corresponding imports are found in the "init.py" file of the package).
After a full name import of module in Python, the imported module has to be referenced by the full name only. In Java, only the class name should be used for reference after the imports. To alleviate this problem, Python supports the use of alias by "as" keyword in import statement.
Python supports defining variables and introducing names at the package level, which makes package level API design available.
Example: import tensorflow as tf (all APIs are introduced into the tensorflow namespace by using imports in "init.py" file of package tensorflow)
In contrast, a java version of tensorflow API would have to be imported by the following statements: import tensorflow.api.xxx.
(In java, package object can not have methods/functions. Instead, all methods are defined in classes and must be introduced through importing classes)
Python's import mechanism is pretty messed up with lots of inconsistency, partly due to its natural as an interpreting language.
- Wildcard imports in python can only be used when "all" magic variable is defined in the "init.py" file of the package. Only names defined by "all" will be imported by the wildcard import statement.In Java, no such definition of wildcard names is required.
Avoid using module attribute: "file" in any cases ( due to inconsistency and ambiguity), all module/package path issue should be resolved through entry point: sys.path or cwd() (1 in 2).
Note: After py3.4.3, "file" will return the absolute path of the module object when not invoked from "main". Using "file" would result in different behavior between running a script directly under cwd (using relative path) and running a script indirectly from cwd/abs path. Due to the fact that "file" is a module attribute, this different behavior can not be mitigated by patch wrapping.
Good python API design should collect all important APIs in the package namespace through programming "init.py" file. This avoids explicit imports of all sub-modules and class/function names
Class instantiation in python does not require the use of new keyword.
All if and loop "block" in python has global scope.
- In python, "is" operator is the identity test, "a" is "b" is the equivalent of id(a) == id(b). "==" operator is the value equality test. When in the equality test, if the right operand is a primitive type object (an expression), then the "==" operator is safe to be replaced by "is" operator.
当他们用（==）进行比较的时候，比较的是他们在内存中的存放地址，所以，除非是同一个new出来的对象，他们的比较后的结果为true，否则比较后结果为false。 JAVA当中所有的类都是继承于Object这个基类的，在Object中的基类中定义了一个equals的方法，这个方法的初始行为是比较对象的内存地 址，但在一些类库当中这个方法被覆盖掉了，如String,Integer,Date在这些类当中equals有其自身的实现，而不再是比较类在堆内存中的存放地址了。
pass by value v.s. pass by reference v.s. pass by object reference
Python and Java and C++ has its own language describes their object model and argument pass mechanism.
Note: Object reference means an object is bound to an identifier (called a variable), all changes to the variable applies to the object referenced as well. (Java: Pass the object reference by value).
A reference to a variable is treated exactly the same as the variable itself, any changes made to the reference are passed through to the argument.
Python is designed to be "pass by object reference", that is for primitive types and non-primitive types argument passing is done through passing of object reference (which is the id of this object, iid, memory address).
Java is designed to be "pass by value", for primitive types it is equal to value copying, for non-primitive types it is the copying and passing of the object's reference by value.
C++ has pass by value and pass by reference mechanism. Pass by value applies to both arithmetic values and object pointers (is itself an object) and class instances, it is essentially a copying of the argument (applies to class instance as well). Pass by reference applies to all types in that all changes to the parameters affects the arguments as well.
Note: The effect of C++ pass by reference cannot be produced in Java and Python (for both the primitive types and the non-primitive types), which is, a function can change the reference of the variable that acts as the argument of this function, iid, change the reference of outer scope variable via operation in the inner scope. (Needs verification).
The above mechanism reflects the justification of "pass reference by copying value".
In C++, a variable is a storage location. "int a = 1", assign value 1 to the address location of a.
In Python, a variable is a purely an identifier (bounded to certain objects). "a = 1", assign/change the object reference (not a reference in c++ context) of a to object 1.
Java requires that all functions been implemented as methods of a class. Variables take two forms, one is the class members (attributes in JS, property member in python), the other is temp variables that reside in the local scope of methods.
C++ can have variables and functions defined/used in the global scope.
Python technically can only have variables and functions defined/used in the module and all kinds of local scope. "Global Scope" in python usually refers the module scope of the "main" module.
In C++, the pointer is declared by adding a affix to a non-pointer type to construct a pointer type. int* a = &b
In golang, the pointer is declared by prefixing the non-pointer type to construct a pointer type. var a *int = &b
Single Threaded Concurrency
Due to the fact that CPython implementation has an embedded GIL, the mainstream python ecosystem is build on the single threading assumption. All major IO libraries are consequently written in a synchronous manner.
In modern python, concurrency is achieved through the use of coroutine, which is implemented via generators and the yield statement (previously iterators).
A function which implements a yield statement become a coroutine, the context switch between the coroutine and the master routine (the caller) is triggered whenever the generator function is called. The yield statement allows the local variables, states, and execution progress of the coroutine to be retained, which makes single threaded concurrency available.
Single-threaded languages have urgent need for asynchronous IO libraries because in a multi-threaded environment, even when the IO of a specific thread is blocked, the overall program is still responsive. The extra burden of single threaded libraries such as python and js is the requirement of asynchronous implementation of IO library. The lack of IO std library is the reason why JS is chosen by Ryan Dahl as the implementation language of NodeJS, largely to achieve asynchronous IO from the bottom. To be noted, NodeJS is a run-time environment packaged with JavaScritpV8 engine rather than a new language. In NodeJS, all IO is performed in the default event-loop (with no event-loop called to be defined/declared explicitly).
The Limitation of Single-threaded Concurrency
Single-threaded concurrency can only solve one type of Blocking: IO-blocking, this is the emphasis of NodeJS and Python tornado. In an environment where all IO is non-blocking, a computationally intensive operation will block the entire server. For the above reason, a single threaded non-blocking IO server must and can only handle computationally intensive operation through Restful Http/RPC call to other service providers (which should be non-blocking). In this chain of service request / providing, except for the first call from client, all the following services should be non-blocking, otherwise the overall service is blocked.
Single-threaded non-blocking IO concurrency can only save CPU resources for the Http Server (The nodejs/tornado server that only handles http requests) in a manner that one thread v.s. multiple threads (may contain suspended threads). This server can only handle tasks that (besides IO operation) will not be CPU intensive enough to block the whole IO chain (a light task).
Blocking/Non-blocking v.s. Synchronous/Asynchronous
Synchronous/Asynchronous describes the identity of the thread who process the IO. In synchronous mode, the main thread process the IO, so all IO must be processed one after another (refers as "blocking") and the main thread cannot do anything until the IO request returns. In asynchronous mode, the main thread assign the IO request to a child thread (or coroutine), so all IO can be processed simultaneously (non-blocking). Synchronous/Asynchronous is used to describe the mechanism of the request processing logic, this terminology is commonly applied from a service provider's viewpoint.
Blocking/Non-blocking is used to describe how requests are processed, this terminology is commonly applied from a IO request caller's viewpoint. Blocking means the service operates in a blocked manner.
All kinds of Blocking
Blocking for single-threaded language includes all executions (computationally or non-computationally) that would suspend the current (main and only) thread. Even sleep(1) in different coroutines (python) will block each other because there is only one real thread running at a given time. In this sense, single-threaded languages tend to be very sensitive to all "blocking" operations.
Single threaded concurrency can not handle cpu blocking (deployment of dl network, syntactic parsing, etc).
When dealing with blockings from other sources: network-io, disk-io, etc, event-loop is more efficient in terms of cpu resources. Event-loop is most suitable for light-weighted http server.
- Blocking/Non-blocking Http Client -> Async Tornado Http Server (Thrift: AsyncClient) -> (Thrift) TThreadPoolServer
- Blocking/Non-blocking Http Client -> Actor-based Concurrency Playframework Http Server (Thrift: SyncClient) -> (Thrift) TThreadPoolServer
Note: service requester can have blocking io, only service providers should be able to handle concurrency request (through multi-threading or event-loop).
本地测试的入口由tf.app.run()改为run_inference_on_image(), 测试图片的路径在if name == "main"下修改，避免更改模块代码。
The creation of tensor object under tensorflow framework.
- The generation of tensors as the input of a computation graph is through tf.constant() and tf.Variable() constructor.
- The basic arithmatic operation between tensors in a given computation node (ops) must be done through specialized operations such as tf.add, tf.mul, etc
- The operations (and the entire graph) on tensors will be executed on a specified device (CPU or GPU).
- The return of the execution of the computation graph is a ndarray of numpy
The programming style of the tensorflow framework.
A complete computation process is defined by a series of computation steps, the output of one step become the input of the next step.
In short, TF adopts functional programming.
TF does not require a "graph" object to be explicitly constructed, rather, "graph" are built through a series of function calls. All functions(ops) defined in the "graph" is carried implicitly with the tensor returned. The run method of session method in the execution phrase will call all function calls carried with the tensor input.
In tensorflow, matrix is represented by a two dimensional list or np array. Row vector is represented by a list, column vector is represented by the transpose of row vector. The vector must be represented in matrix form ([[1, 2]]) in stead of a plain list.
Under the notation of tf, a matrix (tensor) is a column vector which is composed of several row vectors.
N-row is determined by the length of the outer list, N-column is determined by the length of the inner list.
Input/Output of TF Operations
The input of an TF operation is typically tensors (matrix), and the output is typically tensors (matrix) also. Even scalars should be wrapped in tensors (matrices). This implies the flow of datasets.
The output of a tf.reduce_mean() could be a scalar or tensor (ndarray), the output of a tf.matmul() is a tensor (even as a tensor with rank 0).
Indexing and Item assignment on Tensor Object
Indexing on a tensor object is allowed while item assignment is not supported.
- Identifying for the base case: What is the easiest case ? What case(s) can I solve trivially ?
- Redefining the difficult cases
- How can I build a path to solve a difficult case based on the easy case ?
Shared Characteristics Between Recursion and DP
Reasoning is about walking from the starting point, usually some known facts/existing materials, through a path and reach the predefined target.
The solution, which is the goal of problem solving, is the definition or description of such a path. The routine that is necessary to solve the problem is sequential in a forwarding manner because of the temporal nature of computing, however, the measure that takes to find the exact steps of such solutions may require backward reasoning. This is because the threads that leads to the target is often limited in comparison to the possible threads that can be pulled from the start point.
The logic behind recursion and dynamic programming when applied to solve a problem are both based on the inference chain, either through forwarding reasoning or backward reasoning.
When applied in the form of backward reasoning, the logic chain is built through the form of "if the goal is to reach E, then D has to be reached, etc, etc, etc".
When applied as a forward reasoning, the form changed into "From A, we can reach B, the from B we can reach C, etc, etc, etc".
The path can be found by either forwarding or backward reasoning, which are referred to recursive thinking and deduction thinking.
Recursion and DP actually refers more to the how the path is implemented than how it is discovered. Recursion is a style of software implementation as a direct translation from the backward reasoning mental procedure to the source code. DP, in comparison, is a style of algorithm implementation that reflects a way of natural forwarding reasoning.
Algorithms are the representations of logic to solve problems in the form of discrete mathematics in the context of computer science. Algorithm implementation can be strict or loose.
Why Recursion is Useful in Programming ?
Recursive programming avoids the need to write nested loops which is both difficult to understand and debug. It also improves code simplicity, Besides, recursion (make a function call in the body of itself) can implement logic that is nearly impossible to be done in loop structure by human beings.
In recursive programming, the control flow is forwarded through the change of argument values of recursive function calls. This beats loop control structure in cases when it requires that the loop counter changes in a manner other than natural incremental. Additionally, the mechanism of loop counter determines that it is not the best tool in traversalling ADTs.
Install executable package into a directory (example: /apps/my-app) instead of site-packages
1. Keep all statements, functions, classes in one module (your script !) . No relative imports is allowed.
2. function main() is not needed.
Package Python (package name maintained)
1. Keep the overall package layout flat. The package name is the only entry point of all module imports, use absolute imports only, sub-packages are generally accepted but not encouraged.
from mypackage.mymodule import myfunction
2. Expand sys.path in the module that contains main()
Package Python (package name not maintained)
1. Keep the overall package layout flat. Use relative imports (import from the same folder) only, sub-packages are generally accepted but not encouraged.
from mymodule import xxx
Coding scripts (temporary):
Put all definitions, variables, classes into one script; Or split them into several scripts if really needed (using relative import)
Coding non-trivial packages:
Split definitions, variables, classes into separate scripts; using absolute import, expand sys.path if needed.
- Interfaces cannot have typed parameters (dynamic typing nature), which raises troubles and confusions for the users (partial due to the lack of IDE support).
- Interfaces with a generics nature in dynamic typing system accepts arguments with different primitive types (root types). This complicates interface calls.