Deployment : HuggingFace.co/spaces/mya-mya/HimitsudoguSearch
Code : HuggingFace.co/spaces/mya-mya/HimitsudoguSearch/tree/main

Doraemon is a character in  the Japanese famous anime, Doraemon. He is a robot came from the future and has many superscientific tools, called Himitsu-dogu.

Doraemon has a thousand unique tools, so it is hard for us ( and maybe him, too ) to find out the best tool for each situation and purpose. So, I made the recommendation engine that selects the tool which is most relevant to the sentences provided by the user.

The key part of the engine is the nlp model, Sentence BERT. It calculates the latent vector of the provided sentences. And similar vectors are calculated when each other's sentences are similar. By using this, we can search a tool whose description is similar to what the user said. Sentence BERT calculates the context, meaning of the sentence, so you can describe the tool freely, with your own words.

Code : https://editor.p5js.org/_Y.M_/sketches/a8x7nIy_1
Game : https://editor.p5js.org/_Y.M_/full/a8x7nIy_1
Blog : Qiita.com/Mya-Mya/items/162f44b2b1af2a2630b3 

The player ( orange circle ) in this game is under control using CBF. You basically can set the velocity of the player, but when the player is near the enemies ( blue circle ), your input is rejected. 

In this game, the player and each enemy have CBF, which aims to avoid a collision. And your input is filtered not to violate the constraints.

Code : GitHub.com/Mya-Mya/Novel2VecWeb
Deployment : HuggingFace.co/spaces/mya-mya/Novel2VecConsole 

In the Twitter hashtag #名刺代わりの小説10選, meaning 10 selected novels as a substitute for a business card, many readers tweeted 10 of their favorite novel titles.

I made the following assumption; If the same reader lists novel A and novel B, then A and B are similar. In contrast, if the same reader does not list novel A and novel B, then A and B are not similar.

By collecting and analyzing the tweets, we are able to assign a feature vector to each novel, which represents its "direction" and can be used for cosine similarity calculation...? This idea is as same as Word2Vec. so I took one novel as a token, one tweet as a sentence, and trained the word2vec model.

As a result, a novel calculator was born. Each novel is represented as a vector, so we can ADD AND SUBTRACT NOVELS !, but I'm not sure what this means...

Code : Github.com/Mya-Mya/UnsiOptimizer 
Blog : Qiita.com/Mya-Mya/items/5255247abc3b162aea59 

When playing musical instruments, we need to think about fingering, the selection of which finger to use. Fingering is an important process for both amateur and professional, but it is difficult for amateur to do that; it's already hard enough to read score !!

Therefore, I introduced the method to automatically calculate the best fingering of the score. In each transition of notes, there are easy fingerings and difficult fingerings. For example, if we have notes D→E, playing with 2→3 is easy, while 5→4 is difficult. So, the 5×5 matrix representing the difficulty of each fingering can be defined on each transition.

Then, the fingering problem is expressed as follows; find fingering ( blue line in the fig ) which minimizes the sum of each difficulty of transition. And this problem can be solved by the Viterbi algorithm, one of the dynamic programmings.

Code : GitHub.com/Mya-Mya/c2c
Deployment : https://mya-mya-c2c.netlify.app/

How to visualize complex → complex functions ? Let z = f ( v ), where v, z be the complex number and f be a complex function.

One of the ideas is to change color according to the coordinate of  z and display it in the space of v ; the color on the point (Re v, Im v) is the color that is uniquely determined by the coordinates of (Re z, Im z). However, this method is still difficult to easily understand the correspondence between v and z.

In c2c, I visualize the f as a wind direction map - like stuff. A wind direction map can be seen in the weather forecasting program. The map shows the direction of the wind at each location ; that means it shows 2D→2D mapping, which can be also used in complex→complex function visualization !

Code : https://github.com/Mya-Mya/intelligent-fill-preview
Demo :  https://keio.box.com/s/76f2y89040n1gok8pow0lgxy5khft4y9
Paper : https://keio.box.com/s/1zum09np0eb7wgk56qsi47uyr75phmgy
Poster : https://keio.box.com/s/gjndyzz3so3aiiu18vqmqwyhqzzzs067

At the beginning of drawings, we first need to do the line drawing.  After the line drawing, the next comes filling. When filling, we are often in trouble ; OH, I GOT FILLED IN PLACES I DIDN'T EXPECTED ! It occurs because of the very small hole in the line drawing. The normal filling method would not work well unless we perfectly complete the line drawing, without leaving a 1px of missing. It is stressful for many illustrators.

To address this issue, I proposed a novel filling method, intelligent fill. It uses the line-closing pix2pix model to predict the line drawing image with no missing, and then perform the normal filling method. In this way, we can get our line drawing filled correctly even if it has some missings.

Paper : https://keio.box.com/s/t6x8znkstf5o865o2dprmtnzbbb2nlsh 

Score : https://musescore.com/user/50113001/scores/9445294

Deployment : omiyayimo.starfree.jp/latexitweb/

Enables you to write, check, and retry latex math much quicker than latexit.

Deployment : omiyayimo.starfree.jp/ffmaker/

Just fill in your request and it will automatically generates ffmpeg command for you.