Since nothing else to do today, simply to write about six months to learn summary of it, by the way also write about some of the tragic experience of the encounter is really a type of bitter tears.
First, entering a new school
After determining eligibility push free election boss, set a direction from October 2018, felt the life of another downturn. You can only blame yourself, each selected direction. Here advise to ask you, direction, and the owner of RP is very important , not the teacher's past reputation and value the name of the first school, and they to you, and not so much help. The direction of interest is ML, DL, but the direction of the teacher is to do what a mess of things industrial control, you move came in talking about the hype, after you come to know, is what GP ah! No way, can not change. . .
Second, look for opportunities
In order to seek their own way out, only to discover themselves, explore a good direction to go to learn, and no one with someone with a really big difference, and now I do not have a clear direction for their future to be engaged, really difficult!
July 20, 2019 to a new school, started the postgraduate level, the more do the task, the more miserable feeling, whenever someone asked my research, my heart is always a burst of low model, which point to make their own adjustments. During this time, we are given an algorithm to do after the post, although I heard good direction, but the platform is still, brothers and senior sister apprentice into manufacturers pretty much, one will be determined according to the Internet by the company. This time will be in the eating book algorithms, because most undergraduate game is to do partial embedded control, with the basic direction of future do not take sides, turned into the equivalent of a white, know how bad. Look at some data structures and algorithms related knowledge, this time the first Python re-live it again, it took about one week's time, then took about a month's time to read some knowledge of data structures, because not C ++, so look for some exercise algorithm data structure of the book of this python, the python also this continues to consolidate a bit. Or because no base, so after finished the list, behind the temporary things can not go on, miserable. After reading this phase, also in October, and began to turn the machine learning related things. I still clearly remember the time, 10.11 start.
Third, start doing machine learning and data mining
Here we began doing things related to machine learning, or to understand, to a entry. Here is strongly recommended a series of introductory tutorial Ali Tianchi, this phase of the study probably took me less than a month's time, each tutorial over again.
Tianchi AI
probably in November, I started a novice tournament, scientific data entry.
By doing this race, I learned a lot of machine learning some practical knowledge, not to mention stay in the tutorial above empty, to improve their game by playing regarded as an efficient way to the bar. Finally scored the first 10 .. . . (In fact, when patchwork forum Gangster open source bseline up, lgb final single-mode, single-mode convergence not drink well ...) but found in the back, a good single-mode lgb line effect, it is because over-fitting the line for a data set, there is no fusion of the stable.
Fourth, the official race
In mid-January, when my roommate took the start for the first time to play official matches, reference baseline big brother and then do their own characteristics, the first to the last official test set testA we use single-mode results were very good, but when the second test set to give testB, the effect is obvious a lot of difference, indicating a severe over-fitting line, in fact, cause your model is not good.
Too much garbage, but barely entered the semi-finals.
Met a lot of pit beginning we have been doing feature, reading papers, watching baseline characteristics plus others open source, a lot of points up, she was still playing it spent a lot of time on it, I feel cost is not very high, because the school into something not many. In fact, there is a very sucker thing is tm test scores do not always go online and try so much every day, the line did not move or drop effect, good line, but so what? The equivalent of being told you do not know a test machine, and then you try crazy, if this tells you is not the same operation, so it feels a bit of a waste of time.
We are also the lgb integration with xgb do, in fact, the line did not rise, but probably will go up another set of data, so fused with the benefits of integration. For the training output characteristics do lgb experience, it seems, no P use.
The last greatest achievement is the big brother and another senior team together, with a lot of things to learn, here or sigh, Coban Coban is really not the same. Is a turning point in the group, because it did not move, so go and ask someone else, to do a big brother with Textcnn lgb fusion method and the effect of non-ordinary good, behind the stage followed by the brothers I learned a wavelet nlp, feeling harvest is still quite large. In addition brothers also told some directions after, let us how to do it, I feel big harvest.
Fifth, the future continue to fuel it
Then they have to improve efficiency, parallel play the game!
