In the summer of 2013, the semester of Yanjing Electric Power University is coming to an end.

Most students' courses have ended, and except for a few subjects that have not yet been tested, they basically have nothing to do.

While most students were dejectedly lamenting in the dormitory about why the school has not installed air conditioning yet, two people were very motivated and preparing for the competition that will start in two months.

The two have had a lot of chats, and Meng Fanqi probably sorted out the basic concepts and understanding for them.

He has always believed that when new people come into contact with AI and artificial intelligence, they should not be obsessed with a certain part or concept. It is easy to lose more than gain and fail to understand it after reading it for a long time.

It is better to first understand the overall principle in a simple way and combine it with common scenes and concepts in life.

Then, first come into contact with a code example that can run smoothly, and run a demo, which will greatly enhance your self-confidence and give you a strong sense of accomplishment.

With interest, a sense of accomplishment, confidence, and a general understanding of the overall process and meaning. People will naturally explore the code spontaneously and try to understand the meaning of each component, each parameter and the consequences it will produce.

Impact.

Maybe there will be some wonderful ideas.

This approach from top to bottom is the right way. Rather than chewing on formulas for several months, you end up not knowing how to write a line of code when you come to the computer, and you have no way to start.

Some time ago, Meng Fanqi has implemented a demo of a simple three-layer neural network for ten types of tasks based on Alex's cuda-conv framework.

Through this, Don Juan was basically familiar with the content of the first task of this competition.

Meng Fanqi then instructed him to take a closer look at the data storage method, development tools and usage methods provided by IMAGE.

I tried to implement the algorithm I had prepared during this period based on this framework, but it did not go well.

In fact, Alex himself is aware of this problem on his homepage. He said that the description of the code on his homepage is very 'skimpy' and relatively inadequate.

At this time, the very imperfect ecology has further aggravated this problem. NVIDIA's CUDA has not started for too long, and the code between different versions needs to be changed.

Meng Fanqi changed the versions of several code libraries. As a result, the driver of the graphics card GTX-690 seemed to be incompatible, and the compilation of some codes was not smooth. He believed that the code itself was a relatively mature version released by Alex, and there would be no

What's the problem? It's just a troublesome environment and debugging.

At this time, AI technology cannot be said to be a niche, but it is not to the point where such detailed issues can be found everywhere. Meng Fanqi browsed stackoverflow and other websites and checked some related discussions.

Although most of the problems have been solved, some are still left.

"I have no choice but to write an email to Alex for help." Meng Fanqi was not very worried about Alex's willingness to reply. As long as he expressed his intention and made it clear that he hoped to use deep neural networks to participate in IMAGE-2013, Alex would definitely be willing.

Lend a helping hand.

As we all know, the three giants of AI are Hinton, Lecun and Bengio. The 2018 Turing Award was jointly won by these three people in recognition of their persistence and contribution over the years.

The core technology for the take-off of AI in the new era is deep neural network. However, after the two booms in the 1960s and 1980s, neural networks did not actually receive enough attention.

Alex is a student of Hinton. When they and other students participated in IMAGE-2012 as Alex in 2012, they were the only team using neural networks.

Meng Fanqi knew very well that Hinton and Alex were eager to promote neural networks, and they had already defeated other competitors in a crushing manner in 2012.

At this time, they have not fully started their very busy career in the future, and are in the joy of "all nations coming to Korea". They are very happy to provide technical support to non-profit colleagues.

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

Caffe, which was later developed by Dr. Jia Yangqing, is the first love of many AI developers, and a lot of the content is also based on cuda-conv. Dr. Jia himself also communicated a lot with Alex, and received a lot of help and support from Alex.

After many years, Dr. Jia still remembers Alex and cuda-conv, which shows that Alex is indeed a warm and hospitable person in terms of AI technology.

Although I decided to write an email, the content is not that easy to write. Programmers are a group that have high requirements for the quality of questions, especially top brains like Alex who are at the forefront of the times. Such people are very concerned about stupidity and arrogance.

The tolerance level is very low.

Because Open Source is a major feature of this world, the rapid rise of AI relies heavily on this atmosphere of selfless sharing. Alex, who won the championship last year and defeated a number of top universities and technology companies, soon had all his codes

It can be downloaded directly to anyone's computer.

There are no barriers or thresholds, and the core content is placed in front of you for free. Under such circumstances, it would be a bit insulting to ask stupid questions.

People who don't want to think, or don't do what they need to do before asking a question. Those people are time killers - they only want to take, never give, and use up our time for more interesting questions, or people more worthy of answering.

Time on your body.

"Bring it." "Give it to me." That's what their questions sounded like.

Alex provided help to many people in his own time. It was the first time he had direct contact with influential figures in the industry. Meng Fanqi did not want to appear rude or like a fool. Besides, hasty questions can only get hasty answers.

If a person cannot describe his problem in detail, clearly and logically, he is unlikely to get the answer he wants. Most people always focus on describing their feelings when asking for help, but in fact

No one cares.

For example, please use the picture to find the source. "Ah ah ah ah!!! Urgent Urgent!! I'm going to explode!!" and "Japanese comic, with color pages, is a short story, with three characters in total, the style of painting is a bit like XX and

XX."

Which of these two descriptions is more likely to get a reliable answer probably does not require too much explanation.

There is also a very classic starting gesture, which is "Are you there?". If you see such a question, run quickly and don't look back. Nothing good will happen next.

Opening his mailbox, Meng Fanqi described in detail his relevant hardware models, driver versions, environment and library versions, specific error locations, error logs, searches he had done, methods he had tried, and their results.

Click to send.

"It should be early morning in California now, it's not yet dawn." Meng Fanqi calculated the time difference and found that the earliest he would get a reply would be tomorrow.

The most important thing at this time is to first make modifications based on Alex's framework to implement Res, which has been famous for 15 years, and then debug it after the problem is solved.

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

Meng Fanqi knew very well that when the results of the competition were announced in a few months, his Dream, which was based on the idea of ​​kaiming residuals from later generations, would definitely cause a great sensation and gain considerable attention.

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

First, use the residual-based Dream to whet everyone's appetite. When everyone is eager to know the technical details of Dream, we will announce the specific technical details and experimental results of the adversarial generation network based on Dream.
Chapter completed!