Diary 2022-09-16

img2prompt consumes 24GB of VRAM

A bug was discovered for images under seed 752.

• When I was reading multiple prompts from a file and generating them, I had to fix the seed for each prompt, but I only did it at the beginning.
• So, even if "seed=500" is recorded, that is "the state after several images are generated after seed is set to 500", so if seed=500 is set again, the same image will not appear
• I can manage if I try hard enough, but...
  • For example, after setting seed=500, you can draw several times until you get that image
  • Wife: "Not as hard as I try."
• Okay then.
  • I was going to print out the pattern of Diary 2022-09-13 again in a size that could be used for a book cover or a hand towel, but I was puzzled because the image I was expecting did not appear.
  • Now that we can output cyclic images, I thought I'd do that as far as the hand towels are concerned.

Speeded up img2prompt to work in 2-3 seconds per case.

• 638 img2prompt results were obtained

• Let's think about importing into Scrapbox in parallel with Instagram posts.

• Thompson sampling

• The Bernoulli distribution is used to determine whether an image is "good" or "not good".

• Its conjugate prior distribution is a beta distribution

• Thompson sampling to select the largest one by sampling

• Sampling from a beta distribution can be done with numpy.random.beta

• Let's say there are two slots here, one slot per 5 draws and one slot per 30 draws.

  • Although the latter is a greater expectation, the uncertainty of the former should be positively evaluated py

In [245]: [0 if beta(2, 5) > beta(31, 71) else 1 for i in range(20)]
Out[245]: [0, 0, 0, 1, 0, 1, 1, 0, 0, 1, 0, 1, 0, 0, 0, 1, 1, 0, 1, 1]

- The former was chosen 11 out of 20 times.

python-gyazo · PyPI

• Diary 2022-09-15 ← Diary 2022-09-16 → Diary 2022-09-17 100 days ago Diary 2022-06-08. 1 year ago Diary 2021-09-16.

This page is auto-translated from /nishio/日記2022-09-16 using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I'm very happy to spread my thought to non-Japanese readers.