This is the link to the official website:JiShi Developer Platform-Computer Vision Algorithm Development Platform-JiShi Technology (cvmart.net)
There were many problems when I used this platform for the first time. First of all, before using this platform, I mostly used Baidu’s Feipian platform, which is BML, to train some deep learning models.
First, let me talk about the biggest difference between the two platforms of Yiha:
First, the BML platform is more customized, and it is not just jupylab, but also includes packages for rapid environment construction. This may be due to Baidu's own improved and developed deep learning framework, the paddle series, which includes paddledetection, A series of commonly used basic deep learning fields such as paddleocr and nlp provide convenience for quickly building a pain point that is extremely painful for developers, which is to restore the open source warehouse python environment.
The second one is about Jishi’s platform. What gives me the biggest impression is that in addition to using python for deep learning development, it also provides ev_sdk encapsulation using C++. I have detailed this function I have to complain, there are really few tutorials for this ev_sdk in the official documents. This may be because of the "immature platform". Just kidding, back to the topic, its sdk still has a lot of merits. For example, Baidu does not. I specialize in sdk development, but Baidu’s open source suites basically all have C++ deployed, so when I first came into contact with ev_sdk, I thought I could just implement C++ and put it on the platform to compile so files ( The .so file is a shared object file (Shared Object file), which is a binary file format used to store the executable code and data of a program. Different from the static link library dll, the .so file is connected with other programs or libraries in the form of dynamic link when the program is running.), later it was found that this is not the case. Developers need to use it in a given environment, such as tensorflow. It specifies the version 8.3, or specifies the version pytorch1.13, etc. Therefore, if you want to use some dependencies outside the framework, you need to install them and modify related files such as cmkaelist.txt.
Third, compared with the projects in Baidu's BML, the project applications of the Jishi platform seem to be more practical (leaning towards industrialization and smart scenarios), while BML's projects have a lot of playability and diversity. Of course, it is undeniable that I may still like BML more, which can indeed reduce the development cycle of my own. In addition, in recent years, Baidu’s Flying Plasma team has launched paddlex, which is a rapidly developed module. You only need to handle it yourself. Put the data set in the designated location and follow the three-step process to train the model you want without even writing the deep learning code yourself! Absolutely awesome.
The fourth is the deep learning of Jishi platform, which is closed-loop. Maybe it’s because I chose a competition case, but I can’t see what the training data set looks like. The reference example pictures given are also Desensitization, such as coding license plate numbers and mosaic avatars! I have to complain, there are really many bugs in Jishi platform. Several times I suspected that it was a problem with the platform, so the model could not be loaded, so I went to submit a question form. The first time I raised it, the staff responded that I did not follow the instructions. The document operation asked me to read the document. I asked again, but there was no reply. I had to open a new form (asked for the second time) before they would reply to me. Finally, I found that it was a problem with the platform and the model could not be loaded (... at that moment I do have some opinions in my heart), and I use BML a lot, and all the pictures in it can be seen. This may be helpful when choosing a framework. Then there is a baseline project in the BML competition, but the Jishi platform does not seem to have one. . . Although there are some help documents, I have to complain, the writing style is indifferent, and it also contains ambiguity and controversy! Then when it comes to model evaluation, in BML’s cat classification competition, you only need to submit the results to get scores, but on the Jishi platform, you need to write the ji.py interface to implement it. In fact, to be honest, it is not difficult to implement. , the trouble is whether the test can be implemented correctly, because every time the test set is put in, it is a long process. . . . Environment preparation, and there is also a bug, that is, the output on the local computer is obviously correct, but when it comes to the designated test link to judge the results, the display is inexplicably wrong. For example, each line of output in the result.txt file represents a box selection result. It was specified that seven parameters should be output. When I successfully loaded the model, the output was seven parameters, which was consistent with what was described in the document. However, when it came to the model verification stage, the error message showed that there were only six parameters. . .