Have Hisako's guitar replaced with her own guitar by machine learning -Execution-

Preparation (previous)

I wanted to have a good feeling with CycleGAN, so I first collected images of the guitar. I downloaded and resized the image. Click here for details (https://qiita.com/nozomi254/private/2fc55cc8ae8e6359f8a7)

Execution

My environment

• macOS High Sierra 10.13.6
• Anaconda3
• Python3.7.2
• AWS didn't understand at all and just created an instance as instructed
  • Deep Learning AMI (Ubuntu 16.04) Version 26.0 (？)
• Instance type p2.xlarge (?)

Upload images using FileZilla

The folder name is the same so that it follows the procedure described in the README. After all, I didn't really have a black Jazzmaster, and I didn't have enough energy to do my best.

maps
├── trainA(Sunburst for learning:284 sheets)
├── trainB(Black for learning:94 sheets)
├── testA(Test sunburst:151 sheets)
└── testB(Test black:76 sheets)

It became a very unmotivated and unbalanced feeling. It doesn't look good, but thank you.

I thought that FileZilla used in the university class might be easy to use, so I connected it. You can't just type in the square above, but you should look it up like "AWS FileZilla" (http://support.amimoto-ami.com/en/articles/934126-filezilla- % E3% 82% 92% E4% BD% BF% E3% 81% A3% E3% 81% 9F-sftp-% E6% 8E% A5% E7% B6% 9A)

• Open a window from the mysterious icon in the upper left
• Press "New Site"
• It seems to make something with SFTP
• The host name was ʻIPv4 Public IP`

The nice thing about this is that you can upload folders and files by dragging and dropping, or I don't know how to do it other than doing this.

Before uploading the entire directory maps created in the first part, it is necessary to make sure that there is no maps on the AWS side now. It took me a light year trying to remove it with FileZilla.

mv maps maps_real

Solved with. mv can also be used to change the name. I think that even operations on FileZilla can be done in 1/200 million seconds

Execute

The folder names are also available, so [README on the original GitHub]( Follow the instructions in pytorch-CycleGAN-and-pix2pix). First learn.

python train.py --dataroot ./datasets/maps --name maps_cyclegan --model cycle_gan

Takes time

Image created in check_point after learning

It seems that the code is really well made and it makes a web page and shows the result, but I can not see it because it keeps loading forever Also, for some reason, the images (aerial photographs and map photographs) that were in the original maps are also mixed. For the time being, the one I saw in Atom's Preview that seems to work HM. rec and idt seem to mean reconstruction reconstruction and identity identity rec_A = F(G(A)) rec_B = G(F(B)) idt_A = G(B) idt_B = F(A) Reference: different outputs (images) # 326

Once the map is mixed, as far as the creation date and time is seen, it seems that only the result (I feel that it is not mixed with the one for learning), so just run the test as it is

python test.py --dataroot ./datasets/maps --name maps_cyclegan --model cycle_gan

result

The one that worked

The one who couldn't

It seems that it was difficult for the background to be black or for live video What if the background is transparent? I was advised, but I wanted people to do their best on the guitar they are playing ...

That Jazzmaster has changed like that

There are still scratches and it looks like it

Is this black and white? About

This is just black and white

Summary and reflection

Summary

• Guitar images of musical instrument store promotional material work relatively well
• YouTube trial video screenshots work pretty well
• Live footage of J Mascis and Hisako Tabuchi didn't work at all.

Reflection

• The image with a black background doesn't seem to know where the guitar is.
• Sometimes I got a bug with rec after making it black by myself
• Images related to guitars and bandmen, such as live images, clothes, and places to put them, tend to be black.
• It might have been beautiful if I did it only with a guitar image with a transparent background without humans
• But then you can't change it ...
• It is a mystery why the trial system went well. Because there were many similar images
• I wanted to succeed with human inclusion, but there were very few images with human inclusion

Execution version ver.2

You can connect Pycharm and AWS, so I want to try that too

to be continued···