Hans – Public Notes

My custom CSS for Vivaldi

2023-07-22T19:35:16.302730+00:00

The theme has to be set to 'Subtle'

My tweaks are available on Github: https://github.com/hanspinckaers/vivaldi-theme

Major limitation:

Tabs sidebar needs to be exactly 250px, otherwise you get misalignments of the menubar / tabs
Dark mode doesn't work

Before

After

Before

After

pandoc-eqnos on macOS

2023-07-14T12:55:53.001770+00:00

RuntimeError

RuntimeError: Cannot understand pandocversion=3.1.3

Relevant issues: https://github.com/tomduck/pandoc-eqnos/issues/68

What I did:

conda install -c conda-forge pandoc==2.19.2

Error running filter

Error running filter pandoc_eqnos:
Could not find executable pandoc_eqnos

Solution for me:

pip show pandoc_eqnos # get Location
pandoc --filter <>/python3.9/site-packages/pandoc_eqnos.py ...

Use Grammarly Premium with Grammarly LSP in Neovim

2023-07-14T08:32:34.028735+00:00

Basically, I needed to login via VSCode, then I got it working.

Download / install VSCode (the horror..)
Install the extension Grammarly
Shift-Command-p --> grammarly.login
Follow the login flow
Quit VSCode
Check ls ~/.config/grammarly-languageserver
Copy the folder name, this is your clientID

Configure the LSP like this:

require'lspconfig'.grammarly.setup{
     on_attach = on_attach,
     init_options = { clientId = "<>" }
}

Getting TypeError: Failed to parse URL ? You may need to run using node v16.40:

Install nvm (https://github.com/nvm-sh/nvm / https://github.com/jorgebucaran/nvm.fish)
nvm install v16.14.0
nvm use 16
nvim

Use LanguageTool Premium with LTeX LSP in Neovim

2023-07-13T14:32:22.465290+00:00

Find your API key here: https://languagetool.org/editor/settings/access-tokens

Install ltex (I'm on a Mac): brew install ltex-ls

require'lspconfig'.ltex.setup{
     on_attach = on_attach,
     settings ={
      ltex = {
        languageToolHttpServerUri='https://api.languagetoolplus.com',
        languageToolOrg = {
          username='<email>',
          apiKey = '<API key>',
        }
      }
    }
 }

Matplotlib backend for Kitty over SSH

2023-07-11T08:51:03.456484+00:00

I couldn't get kitty +kitten icat working: Terminal does not support reporting screen sizes via the TIOCGWINSZ ioctl

Workaround using tpix

wget https://github.com/jesvedberg/tpix/releases/download/v1.0.0/tpix-1.0.0-x86_64-linux.tar.gz
tar xzf tpix-1.0.0-x86_64-linux.tar.gz
mv tpix /usr/local/bin/

Install my fork of matplotlib-backend-kitty that uses tpix:

pip install git+https://github.com/HansPinckaers/matplotlib-backend-kitty-tpix.git

Recommended: put this in your bashrc (and source it):

export MPLBACKEND='module://matplotlib-backend-kitty'

Or use this in Python:

import matplotlib
matplotlib.use('module://matplotlib-backend-kitty')
import matplotlib.pyplot as plt

Asynchronous Black for Python in Neovim

2023-11-01T19:24:44.237421+00:00

Update: please use this awesome plug-in instead of this post: GitHub - stevearc/conform.nvim: Lightweight yet powerful formatter plugin for Neovim

I searched a bit on the internet for a non-blocking formatter using Black in Neovim, while keeping the jumplist intact. Since I couldn't find it, I coded this with help from a LLM. It's not perfect, but it does the job for me.

function async_black()
  local bufnr = vim.api.nvim_get_current_buf()

  local stdout = uv.new_pipe(false)
  local stderr = uv.new_pipe(false)
  local stdin = uv.new_pipe(false)

  local handle = uv.spawn('black', {
    args = {'--fast', '-q', '-'},
    stdio = {stdin, stdout, stderr},
    cwd = vim.loop.cwd(),
    detached = true,
  }, function(code, signal)
    -- print("black process exited with code: " .. code .. ", signal: " .. signal)
  end)

  local data = {}
  uv.read_start(stdout, function(err, chunk)
    assert(not err, err)
    if chunk then
      table.insert(data, chunk)
    else
      local all_data = table.concat(data)
      vim.schedule(function()
        local save_view = vim.fn.winsaveview()
        local curpos = vim.fn.getcurpos()
        local save_jumplist = vim.fn.getjumplist()[1]
        vim.api.nvim_buf_set_lines(bufnr, 0, -1, false, vim.split(all_data:sub(1, -2), '\n'))
        for i = 1, #save_jumplist do
          if save_jumplist[i].bufnr == bufnr then
            vim.cmd('normal! ' .. save_jumplist[i].lnum .. 'G' .. save_jumplist[i].col .. '|')
          end
        end
        vim.cmd('normal! ' .. curpos[2] .. 'G' .. curpos[3] .. '|')
        vim.fn.winrestview(save_view)
      end)
    end
  end)

  uv.read_start(stderr, function(err, data)
    assert(not err, err)
    if data then
      print("stderr data: " .. data)
    end
  end)

  local content = vim.api.nvim_buf_get_lines(bufnr, 0, -1, false)
  local all_text = table.concat(content, '\n')
  stdin:write(all_text)
  stdin:shutdown()
end

vim.cmd([[
  augroup async_black
    autocmd!
    autocmd FileType python autocmd BufWritePost <buffer> lua async_black()
  augroup END
]])

Papers on Concept Activation Vectors

2023-07-10T09:10:11.331206+00:00

Theory

On completeness-aware concept-based explanations in deep neural networks. Yeh, C.-K., Kim, B., Arik, S., Li, C.-L., Ravikumar, P., & Pfister, T. (2020). https://arxiv.org/abs/1910.07969

Debugging Tests for Model Explanations. Adebayo, J., Muelly, M., Liccardi, I., & Kim, B. (2020). arXiv, 2011.05429v1. http://arxiv.org/abs/2011.05429v1

Concept Whitening for Interpretable Image Recognition. Chen, Z., Bei, Y., & Rudin, C. (2020). Nature Machine Intelligence, Vol 2, Dec 2020, 772-782, 2002.01650v5. http://arxiv.org/abs/2002.01650v5

Concept Bottleneck Models. Koh, P. W., Nguyen, T., Tang, Y. S., Mussmann, S., Pierson, E., Kim, B., & Liang, P. (2020). arXiv, 2007.04612v3. http://arxiv.org/abs/2007.04612v3

Towards Automatic Concept-based Explanations. Ghorbani, A., Wexler, J., Zou, J., & Kim, B. (2019). arXiv, 1902.03129v3. http://arxiv.org/abs/1902.03129v3

Explaining Classifiers with Causal Concept Effect (CaCE). Goyal, Y., Feder, A., Shalit, U., & Kim, B. (2019). arXiv, 1907.07165v2. http://arxiv.org/abs/1907.07165v2

Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCAV). Kim, B., Wattenberg, M., Gilmer, J., Cai, C., Wexler, J., & Viegas, F. (2018). In International conference on machine learning (pp. 2668-2677). PMLR. http://proceedings.mlr.press/v80/kim18d.html

Application (medical domain)

Robust and Interpretable Convolutional Neural Networks to Detect Glaucoma in Optical Coherence Tomography Images. Thakoor, K. A., Koorathota, S. C., Hood, D. C., & Sajda, P. (2020). IEEE Trans Biomed Eng, PP. https://doi.org/10.1109/TBME.2020.3043215

On interpretability of deep learning based skin lesion classifiers using concept activation vectors. Lucieri, A., Bajwa, M. N., Braun, S. A., Malik, M. I., Dengel, A., & Ahmed, S. (2020). International Joint Conference on Neural Networks (IJCNN) (pp. 1-10). IEEE. https://ieeexplore.ieee.org/document/9206946

Concept-based model explanations for Electronic Health Records. Baur, S., Hou, S., Loreaux, E., Mincu, D., Mottram, A., Protsyuk, I., Tomasev, N., Seneviratne, M. G., Karthikesanlingam, A., & Schrouff, J. (2020). arXiv, 2012.02308v1. http://arxiv.org/abs/2012.02308v1

Notes on dialing in an espresso

2023-07-10T09:09:50.885943+00:00

First keep recipe the same (amount of ground coffee in and amount of extracted coffee out) and tweak grind size to get brew time in the right range (25-30 seconds)
When numbers are in the right range use taste to tweak extraction: sourness is mostly under-extraction, bitterness is over-extraction.
Lighter roasts: more work to extract the coffee, which means use less dose or coarser grind. If you start with a big dose, impossible amount of work to do. Sour, weak body as a result. Here underextraction can thus be because of too fine grinds as well!
Darker roast: easy to extract, soluble get out quickly, is OK to use a higher dose. Can benefit from finer grind.
Instead of changing grind size you can also tweak dose a bit, add more dose: is more work to extract, slower flow, so longer contact time.
You can increase extraction with lower dose, also maybe combined with finer grind
Brew time: 25-30 seconds is a good target, but not so important.
You can also increase brew time for an extra little bit extraction, you trade this for strength (because you dilute), but can add clarity and sweetness
Grams out is very important to monitor, to keep the ratio the same.
Espresso ratio: 1.5-2.5, meaning ratio between grams coffee in and grams of liquid out.
Source: mainly James Hoffman's fantastic YouTube channel
Barista Hustle has some good blogs post as well:
- https://www.baristahustle.com/blog/coffee-extraction-and-how-to-taste-it/
- https://www.baristahustle.com/blog/espresso-recipes-analyzing-dose/
- https://www.baristahustle.com/blog/espresso-recipes-strength/
- https://www.baristahustle.com/blog/espresso-recipes-understanding-yield/
- https://www.baristahustle.com/blog/espresso-recipes-time/
- One of their tips: only dose more or less coffee to make more or less coffee.

Using a Sony alpha camera as a webcam in Ubuntu

2023-07-10T09:10:53.594892+00:00

Make sure you have ffmpeg: sudo apt-get install ffmpeg
Install latest release gphoto2: https://github.com/gonzalo/gphoto2-updater
Install https://github.com/umlaeute/v4l2loopback from source (see README.md)
On Sony alpha camera: change USB connection in settings on camera to PC Remote.
Enable kernel extension: sudo modprobe v4l2loopback exclusive_caps=1 max_buffers=2
Launch webcam: gphoto2 --stdout --capture-movie | ffmpeg -i - -vcodec rawvideo -pix_fmt yuv420p -threads 8 -f v4l2 /dev/video2
- Change /dev/video2/ to device mentioned in v4l2-ctl --list-devices

It doesn't work anymore

Did you start the kernel module? sudo modprobe v4l2loopback exclusive_caps=1 max_buffers=2
When after upgrading the v4l2loopback kernel module will not load: build from source again, but start with a make clean make sudo make install sudo depmod -a

Main source:

https://www.crackedthecode.co/how-to-use-your-dslr-as-a-webcam-in-linux/