A branch name is like a pull request's title.

Don't use "main" (of course), "patch-1" or even "fix-379".

Use more descriptive.

If you feel lazy doing so, use "commit2branch" like I do:

git commit
git branch -m $(commit2branch HEAD)
git push
gh pr create

It converts the commit message to a branch name like:

% git log -n 1 | cat
commit 15148eac7c64e7fabb4ae7faf7a22f8ed4c10517
Author: Kentaro Wada www.kentaro.wada@gmail.com
Date: Tue Jan 9 23:58:53 2024 +0900

Build nuitka on CI job

% commit2branch HEAD
build_nuitka_on_ci_job

Replace the "HEAD" with:

• "main" with un-pushed commits
• "15148ea"
• "HEAD~5"
• "temporary_branch"
• ""  # equivalent to HEAD

My normal workflow:

git branch  # --> main
git checkout -b spam
git commit  # add descriptive commit message
git branch -m $(commit2branch)
git push origin
gh pr create

To use it, add this to your ~/.zshrc or .bashrc:

commit2branch () {
  local hash=$1
  local commit_message=$(git log -1 --format="%s" $hash)
  echo "$commit_message" | sed -e 's/[^a-zA-Z0-9]/_/g' | tr '[:upper:]' '[:lower:]'
}

If you prefer seeing on GitHub:
https://github.com/wkentaro/dotfiles/blob/8c1288975cc498bcfc3bd9f0831a6e5eb982fdb2/zsh/plugins/git.sh#L292-L296

Use Descriptive Branch Name with Git

Introducing a new Python app:

"imshow": Flexible and Customizable Image Display

I needed an image viewer that's:

• Flexible - cli &amp; code interface
• Customizable - visualize any data as an image
• Fast &amp; clean

Examples:

imshow *.jpg
imshow *.jpg --plugin tile --col 3… https://twitter.com/i/web/status/1744375875548631424 

1. show one-by-one (default)

imshow *.jpg  # press 'n': next, 'p': previous, 'q': quit 

2. tiling (a builtin plugin)

imshow *.jpg --plugin tile --col 3
imshow *.jpg --plugin tile --col 3 --row 3

Show multiple images as a grid: 

3. marking (a builtin plugin)

imshow *.jpg --plugin mark

Press 'm' and save filenames to a text file: 

4. a custom plugin

imshow *.jpg --plugin countdown_plugin.py

Implement your own custom plugin with Python.
This is just an example: 

How to install:

pip install imshow

Code→ https://github.com/wkentaro/imshow

imshow: Flexible and Customizable Image Display

Certain Python applications require runtime download of files e.g., dataset, model. We'd like to do that once until the file becomes out-of-date.

Gdown (https://github.com/wkentaro/gdown) offers a concise API of cached download via:

gdown.cached_download(<URL>, <PATH>, <MD5>)

Although I created Gdown originally to download files from Google Drive (because curl/wget doesn't work for large files),

it supports any URL to download files from.

With just 4 lines of code, we can treat a remote file as if it were a local file:

npz_file = gdown.cached_download(
    url="https://drive.google.com/uc?id=1l_5RK28JRL19wpT22B-DY9We3TVXnnQQ%22,
    md5="fa837a88f0c40c513d975104edf3da17",
)

vs.

npz_file = "/Users/wkentaro/Downloads/fcn8s_from_caffe.npz"

As you may notice, <PATH> is optional.

gdown.cached_download(<URL>, [<PATH>], [<MD5>])

Unless passed, the URL is converted to a filepath in ~/.cache/gdown.

For a real app example, check how I do at @labelmeai:

https://github.com/labelmeai/labelme/blob/5b692a78ce6c5beb150fd24b15196b5b720893b4/labelme/ai/__init__.py#L12-L15

How to Cached Download in Python

Highly recommend this file handling package in Python: "path"

It gives me more pleasant development experience than other alternatives: os.path, pathlib.

Its object-oriented approach simplifies code a lot:

- os.listdir(file) vs. file.iterdir()
- os.path.join(root_dir, file)… https://twitter.com/i/web/status/1742846482523521082 

Last year, I watched 100+ self-improvement videos on YouTube.

I implemented many but only remember a few of the words that appeared in the videos.

Here are a few quotes that I couldn't forget:

1. "Solve your problem and then solve the world's problem." ー Denmo (Denmosocial)

Denmo discusses social skills and self-development a lot in his channel. He said this when talking about dating problem.

This quote emphasizes the importance of finding personal fulfillment before being able to contribute value to others.

We can only fill others' glasses with water when we have enough water ourselves.

2. "Think of yourself as dead. You have lived your life. Now take what’s left and live it properly." ー Marcus

This quote is from a book "Meditation" by a roman emperor, Marcus Aurelius.

He encourages to adopt a mindset that acknowledges the impermanence of life while emphasizing the importance of making the most out of every moment.

This helped me to stop worrying about what other people think.

3. "Life is about self-improvement to provide more value to the society." ー David (howtobeast)

He talks about self-improvement and muscularity in his channel.

I find this simple sentence elegantly depict the mindset to inspire people who consider self-improvement is try-hard.

Unforgettable quotes from self-improvement videos

Four months ago, I created my first digital product.

It was a simple guide, tailored for the user of my software.
I put it up for sales on Gumroad for free.

In 7 days, I made my first $1.

It blew my mind.
People were paying for a digital product that was FREE!

Over the few… https://twitter.com/i/web/status/1741804460815622230 

Btw, I was inspired by @ModestMitkus to start doing this.
He talks about digital product and personal branding.
Check it out if interested.

My first digital product

Migrated the linter/formatter of @labelmeai to ruff.

Why? because

1. ruff replaces all of flake8, isort, black
2. faster
3. configurable by other than pyproject.toml (i.e., ruff.toml)

#3 was also important for my project, since it uses pyinstaller to build executable and… https://twitter.com/i/web/status/1740961449021968651 

Journey of integrating EfficientSAM into Labelme.

It wasn't straightforward.

1. Convert torchscript to onnx-> Not supported❌
2. Reuse `export_onnx.py` in segment-anything-> Cannot get original results❌
3. Implement new `export_onnx.py`-> Slow runtime❌
4. ✅

#buildinpublic

https://twitter.com/wkentaro_/status/1739620648190927096?s=20

1. Convert torchscript to onnx -> Not supported❌

EfficientSAM authors provided Torchscript (*.jit) as their primary example, so I tried to convert them into Onnx first. However it turned out to be unsupported, and needed to use PyTorch model and checkpoint.

https://twitter.com/wkentaro_/status/1736778489301479712?s=20

2. Reuse `export_onnx.py` in segment-anything -> Cannot get original results❌

I couldn't just run `torch.onnx.export` on EfficientSAM, since Labelme needs its 1) image-embedding and 2) prompt-mask-decoding to be different onnx files, so that image-embedding can be cached.

Since I did the similar export for SAM before (https://github.com/labelmeai/segment-anything), I tried to reuse the same script and the `SamOnnxModel` but failed due to the difference in the architecture and design.

Main differences were:

- SAM uses 0-255 RGB whereas E-SAM uses 0-1 RGB.
- SAM supports text and mask prompt, which E-SAM doesn't.
- SAM resizes longest image size, whereas E-SAM just resizes.

*E-SAM = EfficientSAM

Even though I managed to export and run with onnx files, couldn't get the results as good as original.

And the above differences made it difficult to track down what's causing the difference.

So I decided to implement new `export_onnx.py`.

3. Implement new `export_onnx.py` -> Slow runtime❌

I implemented those, and the qualitative results matched to the original,

- `OnnxEfficientSam_(Encoder|Decoder)`
- `export_onnx_esam_(encoder|decoder)`

however, the decoder runtime felt really slow.

Actually, the prompt-mask-decoding part of EfficientSAM was much slower than SAM.

Was around 0.03s vs. 0.12s (4x slower).

This didn't happen on Pytorch/Torchscript inference, so it must have been due to Onnxruntime specific.

After profiling, it turned out that Onnxruntime resize is much slower (>150x) than Pytorch/Torchscript with some mode ("bicubic").

Since the results looked compelling, I replaced with "bilinear".


https://twitter.com/wkentaro_/status/1739229293438554598?s=20

4. Successful Integration✅

After confirming its performance and speed, I integrated with Labelme.

The image-embedding computation is much faster, and the freeze time feels much more acceptable. (25s -> 2.5s)

The code, onnx models, PR:

- Code to export→ https://github.com/labelmeai/efficient-sam
(see export_onnx, onnx_models, predict_onnx.py)

- Exported models→ https://github.com/labelmeai/efficient-sam/releases/tag/onnx-models-20231225

- PR to Labelme→ https://github.com/labelmeai/labelme/pull/1375
(still draft, but will be merged soon)

Journey of integrating EfficientSAM into Labelme

Integrated EfficientSAM into Labelme.

Timing comparison of the image embedding computation on my M2 mac:

- SAM: 25.5s
- EfficientSAM: 2.5s

EfficientSAM seems more conservative in region selection, but that may be easier to annotate an intended region.

Released our work on object reorientation: ReorientBot #ICRA2022! This robotic system plans to *dynamically* orient objects for specific-posed placement. Check out the project page: http://reorientbot.wkentaro.com
Work with @stepjamUK @AjdDavison (1/n) 

ReorientBot: Learning Object Reorientation for Specific-Posed Placement
project page: http://reorientbot.wkentaro.com
paper: http://arxiv.org/abs/2202.11092
video: http://youtu.be/ahWN84sWWJU
code: http://github.com/wkentaro/reorientbot (2/n)

We train learning models to select the best grasp points and reorientation poses, which are then fed into motion planning to generate a trajectory. (3/n) 

The learned models allow the robot to efficiently find the motion waypoints that give a short, collision-free trajectory for reorientation. (4/n) 

The *hybrid* approach of learned and non-learned motion generation makes the robot both efficient by learning (fast inference speed) and general by planning (various placement configurations e.g., shelf and box). (5/n) 

To demonstrate the capability and preciseness, we tasked the robot with *reverse rearrangement* that reproduces the original pile configuration. (6/n) 