Requirements

• Known and proven brands that have supported the hardware for at least a decade. And not known for dropping products. No Google, Apple, or Samsung.
• It’s meant for outdoor use, so at least semi-rugged. Buttons and not just touch screen. Good readable screen.
• Good battery life
• It should work without a phone for at least a few days. I.e., good offline functionality. It should work without a network for sure.
• GPS and Offline Maps. Ability to run GPS/Maps for at least 24 hours.
• Good community, openness
• Cost. I had a fairly high budget of 50K.

Solution

I looked into Garmin, Coros, and Casio.

I chose the Garmin Forerunner 965. Other watches I considered were the Garmin Instinct 3, the Garmin Fenix® 7X Pro, and the Coros Nomad. I considered the Casio ProTek (WSD-F20) series, but neither the community nor the OS seemed interesting enough. But I do love how rugged they are. I will keep an eye on them.

Garmin Instinct 3 has everything except offline maps. It has a cool screen and GPS. But it has no routeable maps. It’s also expensive. Garmin Fenix® 7X Pro is better in every way, but it’s double the cost and weight. Coros Nomad is the closest competitor; it has all the features I wanted, at a similar price, and it’s rugged and pretty. But I don’t know anything about Coros; no one I know has it. Also, I don’t know much about the community around it. It’s also the brand that is on my list.

I was able to load Indian OSM maps onto my Forerunner 965. I have used it while running and for casual walking. I haven’t done much routing on the watch, though it’s possible. I could follow a GPX track easily. I will blog more. Overall, these features seem to work.

Everything on the phone works offline and also works without the phone. Of course, you can send messages without the phone (it doesn’t have a network), but it can connect to Wi-Fi if needed. It can also store music and play it through my Bluetooth earbuds. So you can go out without the phone.

I have not charged the watch much. I am quite active every day, so it’s not bad. I haven’t taken it out for a full day of motorcycling or trekking with maps/routing on. When I do that, I will write more.

The community around it is good, though they are more focused on outdoor activities and mapping. I haven’t explored their apps yet.

As I said, this is not a review. This is just the list of things that I looked for while buying this watch, held up against them. For how it performs in those areas, you will have to wait. And it won’t be a single blog post. For now, there are no red flags with the Forerunner 965.

Download

1. Go to BBBikes Garmin Page
2. Choose the format Garmin OSM (UTF-8). All modern Garmin devices can use UTF-8.
3. Select the area you want (bbox)
4. Click on Extract
5. Once it’s ready, you will get an email.
6. You can also check the progress on the status page and download it from there.
7. Extract the zip file to get .img file

Load

1. Plug into Ubuntu with USB; it should appear as mass storage.
2. It’s in the Garmin Connect App; On the top right, click on the watch. Go to System > USB Mode; set it to Media Transfer (MTP). Reattach and check.
3. In my case (Forerunner 965), I had to block the `garmin_gps` kernel module from loading. You can try that too.
4. Once that is done. Once you connect, the watch will ask for your permission to connect as mass storage
5. Say yes
6. It should appear as mass storage
7. Go to /Garmin/
8. Paste the downloaded .img file
9. Safely unmount the Garmin
10. Now go to the watch and go to Maps Manager
11. Your map file should be there
12. Open it and see

Quantization

Weights are represented by numbers in a model. A normal model might store each weight using 16 bits, i.e FP16 (half-precision floating-point). Quantization reduces that dramatically, where weights are represented using INT8 (8-bit integer) or INT4 (4-bit integer). They reduce precision but increase inference speed and reduce hardware requirements. Most of the time, quantization is performed post-training, which reduces model accuracy. But there is INT4 QAT (4-bit Integer Quantization-Aware Training), which does this during training, thereby reducing the extreme accuracy drops common with standard 4-bit quantization.

1 bit weight

1-bit goes one step further and asks: what if the weights were just one bit, and if that were done during training? With exactly 1 bit per weight, each weight can have only 2 possible values. 0 or 1. 0 represents -1 and 1 represents +1. i.e., instead of storing numbers like 0.72, -0.13, 1.24, -0.56, we will store +1, -1, +1, -1. That is at least 16 times less storage for weights theoretically.

1.58 bit weight

In 1.58-bit weights (ternary bit weights), there are three possible values: {-1, 0, +1}. Theoretically, we need log2(3) bits of information to distinguish between 3 possible states : 2^x=3. i.e. log₂​(3)≈1.585. Hence the name.

For our interpretation and understanding:

• 0: the input is effectively ignored for that connection. i.e., no influence at all, which also gives you built-in sparsity.
• +1: the input contributes positively to the output. i.e., an increase in the input pushes the weighted sum up. This is “same-direction” influence.
• -1: the input contributes negatively. i.e., an increase in the input pushes the weighted sum down. This is “opposite-direction” influence.

1.125-bit weight

1.125-bit applies different logic for compression and calculation. One can’t literally allocate 1.125 bits per weight. The way this works is that the average bit allocation is used when they are packed. For example, it takes 9 bits to store 8 weights; thus, the average number of bits per weight is 1.125. One important distinction here is that bits per weight doesn’t necessarily tell you how many values each individual weight can take. 1.58 has a clean theoretical connection to three states, while 1.125 bpw usually describes the overall encoding/storage scheme. The table below should make a bit clear

BitNet b1.58 2B4T

This is the first open-source, native 1-bit Large Language Model (LLM) at the 2-billion parameter scale, developed by Microsoft Research. You can find the model on the hugging face, it needs a specific inference code called bitnet.cpp.

Bonsai models

Bonsai 1-bit models are probably the first set of production-ready binary/ternary models. They have 1.7B, 4B, 8B, and 27B parameter models.

I have been running the 8B-parameter version locally. My experience with it has been positive. I have explored it for text summaries, code explanations, etc. I will try to install the 27B version and do some coding to see how it goes.

# Install llama
curl -LsSf https://llama.app/install.sh | sh
# Serve the model using OpenAI-compatible API and a web ui
llama serve -hf prism-ml/Bonsai-1.7B-gguf:Q1_0
# or 8B parameter version
llama serve -hf prism-ml/Bonsai-8B-gguf:Q1_0

Note: Yes, I have moved away from Ollama and now mostly use llama.cpp. Also I sometimes Jan.ai along with OpenWebUI

After that, I wanted an easy way to embed it on any web page. I chose to use web components. It converts the given list of exercises into muscle groups, then applies styles to them within the SVG and embeds the SVG. I have verified it to the best of my knowledge and from available sources. It took a lot of battling with Claude to make it the way I want, that is, to make it generic and reusable. I still have to map a bunch of muscle groups, I have listed all the short comings in the readme.

Finally, I created a WordPress shortcode to queue the JavaScript and embed the component. But that’s not really required. If you have a simple static site, then you can include SVG and a JS file and use the muscle-map tag in your HTML

Wednesday Workout

Run

Calf raise, Tibialis raise

Calf raise

Happy workout and happy logging.

In response to these aggressive moves by the platforms, I have started adding the data directly to the visible part of the image, like how your old film camera would add the date/time directly onto the photo itself. There is no danger of losing that information. As long as I have the picture, I have the date/time it was taken.

Similarly, now, when I take certain pictures, I add text directly to the bottom-right corner, usually in English. I am making sure it has good contrast so humans and OCRs can pick it up easily. I have scripts to extract this data easily if I want to process it. Of course, any cheap visual model can also do it, but it’s easy and cheap to make the model write code and reuse it.

I have also started playing around with adding text to a QR code and then placing it on top of a Photo/Image. It works very well. It makes it much easier for machines to read and can also be read by humans using almost any phone camera.

The Apps that I use are

1. OpenCamera (F-Droid, Play): The most flexible camera app I have used. It’s open source. I use the GPS and Text stamping features a lot. It adds GPS tags, a timestamp, and a line of text on top of the image. I wish it had some more text features, but it works. All my new OpenPostboxIndia project pictures are taken with this. So unlike my first set of pictures, however they are maintained or transmitted, that data will remain with them. It can also be configured to take a sequence of pictures (like a timelapse) that can be uploaded to Panoramax. I also love that it can record video along with the GPS tags as a separate subtitle text.
2. ImageToolBox (FDroid, Obtanium and Play): A FOSS Swiss Army knife of image editing and annotating images or pictures on Phone. It has a ton of tools; some aren’t even directly related to image manipulation, like PDF Tools. So take your time to check out all the tools. Recently I discovered OCR tools and some local AI-based tools. All useful. I generally use the built-in QR code generator, then the watermarking feature to add that QR code to the image. Sometimes I use it for hand drawn notes or markers on top of pictures. I am still exploring this app. I am sure I will fit it into my other workflows and uninstall some other apps.
3. BinaryEye (FDroid, Play) is a dedicated QR Code (Barcode) scanner and creator App. It works well and can also do other things, like forwarding the scanned content to a a web endpoint or working as a remote Bluetooth scanner.

Do you annotate photographs or pictures? If yes, why? and What tools do you use?

I have had the same EPIC number since 2002, and it is still valid. I have kept my information updated over the years. Yet, I had to search for myself in a 2002 electoral roll, and I couldn’t find my name. I don’t exactly remember what happened in 2002 or why my name isn’t there. Or may be it was added after 2002 SIR². Ideally system should know it and guide accordingly.

Then I had to search for one of my parents and link their details. Thanks to my father’s good memory we were able to find the part no. Because the search dropdowns doesn’t always show the polling station address. One has to know it!

Of course, Aadhaar is not compulsory³; it’s just that you can’t submit the form online without e-signing it using Aadhaar + OTP. I have a valid digital signature certificate that I use for e-signing. Why can’t I use that? No one knows.

What happens after the submission, no one knows. There is acknowledgement shown on the screen after the signing, but I dont know what happens after that. You do get a SMS message of the acknowledgement. But that message also doesn’t show any acknowledgement or any details on how to check on it.

For someone with a tech background, it took considerable amount time and effort to submit the form. Imagine folks who are less comfortable with technology or those who have changed locations and doesn’t remember well. Imagine those who fall for things like “digital arrest” scams trying to navigate this.

What a mess this #SIR is. The #ECI website is probably one of the most user-unfriendly applications I have seen online recently. I thought IRCTC was bad.

Some coverage

• Citizen Matter’s has an How – To
• The Hindu – SIR: Voters, BLOs unsure what follows after filling online form
• TOI – SIR’s digital wall: 1 lakh try in Karnataka, only 0.2% get through
• The Hindu – SIR progeny mapping in Karnataka: Different BLOs, different instructions
• The Hindu – SIR: Missing 2002 ‘supplementary rolls’ leave voters unable to verify electoral records

Footnotes

SimulIDE is a simple real time electronic circuit simulator, intended for hobbyist or students to learn and experiment with analog and digital electronic circuits and microcontrollers. It supports PIC, AVR , Arduino and other MCUs and MPUs.

I installed it from Flathub (SimulIDE). It worked well. It is easy to understand and work with. It’s also easy for a beginner to pick up and start designing circuits in 5 minutes. I tried a few basic circuits. Then I tried the Blink circuit and Fade circuits based on Arduino. Unlike previous ones, this one included building a firmware and loading it. To make it complex, I was also using the Flatpak version of the Arduino IDE. Then I found two options

Load Firmware in SimulIDE

In this option, we will build the software in the Arduino IDE. Compile it. Once it generates a firmware (a HEX file), we will load it into SimulIDE. For this, we need to know where the compilation output is in the Arduino IDE, and then make sure SimulIDE has access to that path. Enable verbose output in the Arduino IDE Preferences, and then compile. The IDE will print the path. Take the path, go to Flatseal ¹, and give SimulIDE access to that path.

Now SimulIDE has access to the folders where firmware is built. Before you simulate the circuit, right-click the Uno, go to mega328-109, and then load firmware. Choose the hex file. Alternatively, right-click Uno, go to mega328-109 – Properties, enter the firmware path, and select Reload HEX at Simulation Start.

Now everything is set up. Code and Compile in Arduino IDE. Come to SimulIDE and restart the simulation. That will load the recent firmware and start the simulation.

Configure SimulIDE with Arduino Toolchain

For this, you can use the SimulIDE flatpak version, but you will need the native Arduino installed on the machine. For me, I downloaded the Arduino Linux ZIP file version. Then extracted it into a local directory, say /home/thej/code/arduino-ide_2.3.10_Linux_64bit. Then go to Flatseal and grant SimulIDE access to the folder /home/thej/code/arduino-ide_2.3.10_Linux_64bit .

Once this is done, open an INO file (the Arduino Sketch) on the right-hand panel of the SimulIDE. Go to compiler settings and set the Arduino installation folder path (/home/thej/code/arduino-ide_2.3.10_Linux_64bit) as the Tool path.

Now it’s all set. Use the integrated editor in the SimulIDE to edit code, compile, and load firmware. Then simulate.

Both setups were easy, even though Flatpak made it a bit difficult. If you are not using Flatpak, then you can ignore the Flatseal permission-giving part. The rest of the configuration will still work. Once set up, both flows worked flawlessly. For most people option 2, setting up Arduino toolchain is an easy option. See the demo below.

SimulIDE is an amazing piece of software for beginners. You will have fun building and simulating circuits in it. Try it and let me know.

Word2Vec and FastText

Embeddings are learned numerical representations. The actual values depend on the model architecture, training data, and its training objective. So the same object (say a word) gets different embeddings in Word2Vec and GPT. Think of an embedding as a model’s custom learned coordinate system for meaning.

This coordinate system lets us do useful things. Find semantically similar items, power search, feed downstream ML models, or even perform arithmetic on meaning. A general example from Word2Vec is king − man + woman ≈ queen. There are mainly two types

• SkipGram: Where the model predicts context words given a target word
• CBOW (Continuous Bag of Words): Where the model predicts the target word from its context.

I wanted to build a simple Word2Vec model for Kannada just to try it out and see if it’s useful. In the process I also built a FastText model. FastText is the same as Word2Vec with character n-grams. By default, I have built Skipgram based models. But you can try CBOW too.

With n-grams support, FastText makes it easy to handle out of vocabulary words and capture subword information, which Word2vec is not good at. For example, given the settings minn=3, maxn=8, FastText breaks each word into character n-grams of lengths 3 to 8 and learns embeddings for those n-grams as well. There is also a wordNgrams setting, set to 2 by default. We can play with this setting. wordNgrams is used to learn embeddings about the sequences of adjacent words.

For example consider the sentence:

"ಬೆಂಗಳೂರು ಕರ್ನಾಟಕದ ರಾಜಧಾನಿ"

With:

wordNgrams=1 (default) Learns only individual words:

ಬೆಂಗಳೂರು
ಕರ್ನಾಟಕದ
ರಾಜಧಾನಿ

wordNgrams=2 Also learns bigrams:

ಬೆಂಗಳೂರು ಕರ್ನಾಟಕದ
ಕರ್ನಾಟಕದ ರಾಜಧಾನಿ

wordNgrams=3 Also learns trigrams:

ಬೆಂಗಳೂರು ಕರ್ನಾಟಕದ ರಾಜಧಾನಿ

Data

For the corpus, I went with the Kannada text of Kasturi Magazine OCRed by Sanchaya and available on Archive.org. My repository has this text; you can use it instead of redownloading from Archive.org. I chose Kasturi content because it’s the most generic Kannada content available, small but good enough to do quick tests. It’s a monthly family magazine published from the house of Samyukta Karnataka, a very popular newspaper. You can think of it as Reader’s Digest of Kannada. I have not given the script used to download, as it’s not required. But it’s not that difficult to write if you want to. My bots were identified by headers={"User-Agent": "thejeshgn.com/1.0"} and all the downloaded files are named kananda_kasturi_{identifier}.txt where identifier is archive..org identifier.

Build and Test

You can run the trainer, change parameters and test them. I have included all the code and steps for both Fastext and Word2Vec.

Project is available on Codeberg, Github and HF.

# Prepare the clean corpus.txt from the raw inputs
uv run prepare_corpus.py prepare ../data/ ./corpus.txt

# Train the embeddings with default parameters 
uv run fasttext_embeddings.py train ./corpus.txt --recursive --out ../models/kasturi_fasttext_v4

# Try the embeddings
uv run fasttext_embeddings.py test ../models/kasturi_fasttext_v4.bin --word ಕನ್ನಡ
uv run fasttext_embeddings.py test ../models/kasturi_fasttext_v4.bin --word ರಾಜಧಾನಿ
uv run fasttext_embeddings.py test ../models/kasturi_fasttext_v4.bin --sentence "ಬೆಂಗಳೂರು ಕರ್ನಾಟಕದ ರಾಜಧಾನಿ"
uv run fasttext_embeddings.py compare ../models/kasturi_fasttext_v4.bin

For comparison and evaluation, I used the same method that I had used before for testing other embeddings. Take a bunch of Kannada words (I have used the same bunch), get the cosine similarity between them and create a comparison matrix. I also compared my matrix against the word vectors trained by FastText team using Wikipedia and Common Crawl data. You can see them below.

TODO

I have called it kannada-kasturi-embeddings because it’s a good name and we started with Kasturi. But it wont be limited to Kasturi magazine data alone. I will add other newspaper and magazine data to it and build a comprehensive embedding that’s useful. I will also keep comparing to see how it performs. The current test and compare methods are very basic. I will come up with a better method. So the TODO list is

• Add more quality data
• Improve the clean_text function
• Add more automated quality testing and comparisons
• Build different embedding models and not just Word2Vec or Fastext

In the meantime, try it and let me know what you think.