In the summer of 13 years, the semester of Yenching Electric Power University is coming to an end.

Most of the students have finished their classes, and except for a few subjects that have not yet been completed, they are basically idle.

While most of the students were dejectedly lamenting in their dormitories about why the school was not air-conditioned, two people were working hard to prepare for the competition, which would start in two months.

The two have chatted a lot, and Meng Fanqi probably sorted out the basic concepts and understandings for them.

He has always believed that newcomers who are exposed to AI and artificial intelligence should not be obsessed with a certain part and concept, which is easy to gain more than it loses, and it has not been understood after watching it for a long time.

It is better to understand the overall principle in a simple way, combined with common scenes and concepts in life.

After that, you will first get in touch with a code example that can run smoothly, and running a beta version will greatly enhance your self-confidence and have a strong sense of achievement.

I have interest, a sense of achievement, confidence, and a general understanding of the overall process and meaning. People naturally explore the code spontaneously, trying to understand the meaning of each component, each parameter, and the impact it will have.

Maybe there's going to be some whimsy.

In this way, from the top down, it is the right way. Instead of gnawing on formulas for months, you can't write a single word of code in front of the computer, and you can't start at all.

Some time ago, Meng Fanqi has implemented a toy-level version of a simple three-layer neural network to do ten-class tasks based on Alix's CUDA-CONVNET framework.

Don Juan was thus familiar with the first task of the competition.

MENG FANQI INSTRUCTED HIM TO TAKE A CLOSER LOOK AT THE DATA STORAGE METHODS, DEVELOPMENT TOOLS AND METHODS USED BY IMAGENET.

I tried to implement the algorithm I had prepared for this period of time based on the framework, but it didn't go well.

In fact, Alix himself is aware of this problem on his homepage, saying that the description of the code on his homepage is very inadequate.

At this time, the very imperfect ecosystem further exacerbated the problem, Nvidia's CUDA did not start for too long, and the code between different versions needed to be changed.

Meng Fanqi changed the versions of several codebases, and the drivers of the graphics card GTX-690 seemed to be incompatible, and the compilation of some code was not smooth. He believes that the code itself is a more mature version released by Alix, and there will be no problems. It's just that it's a real headache with the environment and debugging.

At this time, AI technology cannot be said to be niche, but it has not reached the point where such details can be checked everywhere. Meng Fanqi browsed the relevant technology websites and looked at some of the relevant discussions.

Although most of the problems were solved, there were still some left.

"I have no choice but to write an email to Alix for help." Meng Fanqi is not very worried about the question of whether Alix is willing to reply, as long as he shows his intention and makes it clear that he wants to participate in IMAGENET-2013 with a deep neural network, Alix will definitely be willing to lend a hand.

As we all know, the AI triumvirate is Hinton, Lecun and Bengio. The 2018 Turing Award was jointly awarded to the three in recognition of their perseverance and contributions over the years.

The core technology of AI take-off in the new era is deep neural networks, but in addition to the two booms in the 60s and 80s, neural networks have not received enough attention.

Alix was a student of Hinton, and was the only team to use neural networks when the two of them competed in IMAGENET-2012 with AlexNet in 2012.

Meng Fanqi is well aware that Hinton and Alix are hungry for neural networks to be promoted, and they have already beaten other competitors in 2012.

At this time, they have not yet fully begun their very busy business in the future, and they are in the joy of "all nations coming to court". We are happy to provide technical support to non-profit fans.

The open source cuda-convnet on Alix's homepage is actually the epitome of TF and Pytorch in the future.

Dr. Jia Yangqing's later development of Caffe is the first love of many AI developers, and a lot of content is also based on cuda-convnet. Dr. Jia himself has had a lot of communication with Alix and has received a lot of help and support from Alix.

After many years, Dr. Jia still has a sense of nostalgia for Alix and cuda-convnet, which shows that Alix is indeed a hospitable person when it comes to AI technology.

Although I decided to write an e-mail to bother, the content was not so easy to write. Programmers are a group of people who have high requirements for the quality of their questions, especially the top brains like Alix, who are at the forefront of the times, and such people have a low tolerance for stupidity and arrogance.

Because open-source is a major feature of this world, the rapid rise of AI relies heavily on this atmosphere of selfless sharing. AlexNet, which won last year's championship and beat a number of leading universities and tech companies, will soon be able to download all of its code directly to anyone's computer.

There are no barriers and thresholds, and the core content is placed in front of you for free. In such a situation, it would be insulting to ask a stupid question.

Those who don't want to think, or don't do what they should do before asking. Those people are time killers – they just want to take, they never give, and they consume the time we can spend on more interesting questions, or people who are more worthy of answering.

"Bring it." "Give it to me." That's what their question sounds like.

Alix helped many people in his own personal time, and for the first time he had direct contact with the biggest names in the industry, Meng Fanqi didn't want to look very rude and look like a fool. Moreover, sloppy questions can only get hasty answers.

If a person cannot describe his problem in detail, clearly and logically, he will not be able to get the answer he wants. Most people always focus on describing their feelings when they ask for help, but no one actually cares.

Take the source of the picture as an example. "Ahhh!!hhh Hurry, hurry!! I'm going to explode!! "And" Japanese comics,There are color pages.,It's a short story.,A total of three characters.,The style of painting is a bit like XX and XX.。 ”

Which of these two descriptions is more likely to be reliably answered does not need much explanation.

There is also a very classic starting style, which is "Is it there?" ”。 If you see such a problem, run quickly, don't look back, there will be no good things in the future.

Opening the mailbox, Meng Fanqi described in detail his relevant hardware model, driver version, environment and library version, the specific error location, the error log, the search and methods he has done and the methods he has tried, and their results.

Click Send.

"It's supposed to be early in the morning in California, and it's not dawn yet." Meng Fanqi calculated the time difference, and it would be tomorrow at the earliest for him to get a reply.

At this time, the most important thing is to modify the framework based on Alix, implement the world-famous ResNet for 15 years, and debug it after solving the problem.

Not only the code and experimental results, but he also needed to complete the adversarial generation algorithm experiments based on the new deep neural network as soon as possible.

Meng Fanqi knew very well that when the results of the event were announced a few months later, his DreamNet, which was based on the idea of kaiming residuals in later generations, would definitely cause a big sensation and get considerable attention.

At that time, Meng Fanqi will take advantage of this rare opportunity to launch the adversarial generation algorithm in one fell swoop.

First, use the residual-based DreamNet to whet everyone's appetite, and publish the specific technical details and experimental results of the DreamNet-based adversarial generative network when everyone is eager to know the technical details of DreamNet.

What you can't get is what you want more, isn't it?