Depending on your question, maybe you have a lemmy group (sublemmy?) for the tech you are working with.
Or you can try ask here.
Myself and others can try and help you.
this post was submitted on 24 Jan 2026
69 points (96.0% liked)
Ask Lemmy
36916 readers
1920 users here now
A Fediverse community for open-ended, thought provoking questions
Rules: (interactive)
1) Be nice and; have fun
Doxxing, trolling, sealioning, racism, and toxicity are not welcomed in AskLemmy. Remember what your mother said: if you can't say something nice, don't say anything at all. In addition, the site-wide Lemmy.world terms of service also apply here. Please familiarize yourself with them
2) All posts must end with a '?'
This is sort of like Jeopardy. Please phrase all post titles in the form of a proper question ending with ?
3) No spam
Please do not flood the community with nonsense. Actual suspected spammers will be banned on site. No astroturfing.
4) NSFW is okay, within reason
Just remember to tag posts with either a content warning or a [NSFW] tag. Overtly sexual posts are not allowed, please direct them to either !asklemmyafterdark@lemmy.world or !asklemmynsfw@lemmynsfw.com.
NSFW comments should be restricted to posts tagged [NSFW].
5) This is not a support community.
It is not a place for 'how do I?', type questions.
If you have any questions regarding the site itself or would like to report a community, please direct them to Lemmy.world Support or email info@lemmy.world. For other questions check our partnered communities list, or use the search function.
6) No US Politics.
Please don't post about current US Politics. If you need to do this, try !politicaldiscussion@lemmy.world or !askusa@discuss.online
Reminder: The terms of service apply here too.
Partnered Communities:
Logo design credit goes to: tubbadu
founded 2 years ago
MODERATORS
Generally people just call them communities here
although „sublemmy“‘s a great name
I still refer to them as "subs" for brevity
Comms works for that. I like to leave the Reddit baggage behind
I could roll with that
I would love a strong moderated community of people to answer technical questions.
Like a "no stupid questions" style of community with the opposite of StackOverflow's style of communication.
Could I have a more detailed view of what that would look like? When I want to explore a problem for the layman I go to Wikipedia first and later onto SO if I still can't find the answer in relevant literature. Making my question precise is often enough for me to accept an answer.
SO has a pretty great standard for communication. I known it can be frustrating if you don’t know the rules at first, but if one asks a clear question, states what they have attempted, and provides a minimal reproducible example, with code as text not an image, the support is pretty damned good.
The vpn block is probably trying to diminish what they can of bot scraping, to salvage what’s left of their husk of relevance.
I am trying to run an OCR program for handwriting to process some large PDFs of old journals that are scanned into PDF. Doing it by hand will take a very long time. I have a amd gpu and have rocm installed. I tried to configure pip with rocm and failed. I was considering pulling a docker of PyTorth and then configuring gradio in it, then trying to get gradio to run TrOCR. I have never run gradio. I have "easier" LLM programs like LM Studio and Ollama but I don't know if they can run TrOCR. There is AMD documentation on running OCR (https://rocm.docs.amd.com/projects/ai-developer-hub/en/latest/notebooks/inference/ocr_vllm.html) but it's not clear if it works well with handwriting. TrOCR is just trained for handwriting. It's also on huggingface, which i don't know how to use that well.
Ok excellent.
Let’s go step by step.
You say you tried to configure pip but failed.
What was the error? Any logs? Did you follow the steps from the link you provided?
I don't remember exactly, but I have rocm 7.2 installed, and there was something I was trying to install inside pip for rocm and it just wouldn't work, it was like 7.2 rocm wasn't out or the link didn't work. The LLM tried multiple suggestion and they all failed, then I gave up. When I said "inside" pip, I don't know if that's accurate. I am very knew to pip and am decent at linux and only know a small amount of coding and lack python familiarity.
So if you follow the instructions from the link again, can you make it work?
that's not for TrOCR, it's just for OCR, which may not work for handwriting
I did try some of the GPT steps:
pip install --upgrade transformers pillow pdf2image
getting some errors:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╺━━━━━━━━━ 3/4 [transformers] WARNING: The scripts transformers and transformers-cli are installed in '/home/user/.local/bin' which is not on PATH.
Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
mistral-common 1.5.2 requires pillow<11.0.0,>=10.3.0, but you have pillow 12.1.0 which is incompatible.
moviepy 2.1.2 requires pillow<11.0,>=9.2.0, but you have pillow 12.1.0 which is incompatible.
this is what GPT said to run, but it makes no sense because I don't have TrOCR even downloaded or running at all.
Install packages: pip install --upgrade transformers pillow pdf2image
Ensure poppler is installed:
Ubuntu/Debian: sudo apt install -y poppler-utils
macOS: brew install poppler
Execute: python3 trocr_pdf.py input.pdf output.txt
That's the script to save and run.
#!/usr/bin/env python3
import sys
from pdf2image import convert_from_path
from PIL import Image
import torch
from transformers import TrOCRProcessor, VisionEncoderDecoderModel
def main(pdf_path, out_path="output.txt", dpi=300):
device = "cuda" if torch.cuda.is_available() else "cpu"
model_name = "microsoft/trocr-base-handwritten"
processor = TrOCRProcessor.from_pretrained(model_name)
model = VisionEncoderDecoderModel.from_pretrained(model_name).to(device)
pages = convert_from_path(pdf_path, dpi=dpi)
results = []
for i, page in enumerate(pages, 1):
page = page.convert("RGB")
# downscale if very large to avoid OOM
max_dim = 1600
if max(page.width, page.height) > max_dim:
scale = max_dim / max(page.width, page.height)
page = page.resize((int(page.width*scale), int(page.height*scale)), Image.Resampling.LANCZOS)
pixel_values = processor(images=page, return_tensors="pt").pixel_values.to(device)
generated_ids = model.generate(pixel_values, max_length=512)
text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
results.append(f"--- Page {i} ---\n{text.strip()}\n")
with open(out_path, "w", encoding="utf-8") as f:
f.write("\n".join(results))
print(f"Saved OCR text to {out_path}")
if __name__ == "__main__":
if len(sys.argv) < 2:
print("Usage: python3 trocr_pdf.py input.pdf [output.txt]")
sys.exit(1)
pdf_path = sys.argv[1]
out_path = sys.argv[2] if len(sys.argv) > 2 else "output.txt"
main(pdf_path, out_path)
Ok so from the error, you have a version of pillow that is incompatible.
You have to downgrade pillow to version 11.
That’s the first step.
EDIT: Sorry just saw the rest of your comment. Do you really have to use that tech?
You have other alternatives. Amazon AWS has a service for handwriting ocr, can’t remember the name though.
You can also have a look at this, but it’s paid: https://www.handwritingocr.com/
More ocr alternatives: https://github.com/michaelben/OCR-handwriting-recognition-libraries
+1 for tesseract. I knew about this one a while ago. It may not recognise all handwriting, but you can train it to get better at it.
I don't trust big tech to not extract data and metadata and save it. Many companies get served with government requests to save data and keep it secret. Even if handwritingocr.com doesn't have such an agreement, it could run on AWS and that has an agreement. I would much rather do this locally. Some of the writings are confidential. Handwritingocr.com says data is encrypted in transit and at rest, but it's not open source and even if it were I can't verify the server code.
also Tesseract is CPU only, right? It will be so slow.
Fair point.
So what about Tensor flow and some local LLM to do the job?
You just need to find a reliable LLM in HuggingFace, for example.
That's exactly what I am trying to do, I'm just not that sure how to do it. I have the hardware needed, I just need to set up a docker with PyTorch and then find a way to set up Gradio inside that and then add TrOCR from hugging face, and then I'm good. I just am not totally sure how to do that and it seems hard, and when I ask AI for advice, it often is like "just run the following" and it's wrong, and I'm not skilled enough to know why.
Excellent. So let's try to do that instead.
From what I can see from the docs, Gradio is used to build a web interface and have a nice UI to visualise things.
Let's put Gradio aside for now and sort out Pytorch in Docker.
Select your LLM and make sure Pytorch works well with that.
If you run into trouble or get stuck let me known.
I'll grab my laptop and try it myself.
nice starting point: https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/install/installrad/wsl/install-pytorch.html
Terminal error after running GPT code:
python3 trocr_pdf.py small.pdf output.txt
Traceback (most recent call last):
File "/home/user/.local/lib/python3.10/site-packages/transformers/utils/hub.py", line 479, in cached_files
hf_hub_download(
File "/home/user/.local/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn
return fn(*args, **kwargs)
File "/home/user/.local/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 1007, in hf_hub_download
return _hf_hub_download_to_cache_dir(
File "/home/user/.local/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 1124, in _hf_hub_download_to_cache_dir
os.makedirs(os.path.dirname(blob_path), exist_ok=True)
File "/usr/lib/python3.10/os.py", line 215, in makedirs
makedirs(head, exist_ok=exist_ok)
File "/usr/lib/python3.10/os.py", line 225, in makedirs
mkdir(name, mode)
PermissionError: [Errno 13] Permission denied: '/home/user/.cache/huggingface/hub/models--microsoft--trocr-base-handwritten'
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/home/user/Documents/trocr_pdf.py", line 39, in <module>
main(pdf_path, out_path)
File "/home/user/Documents/trocr_pdf.py", line 11, in main
processor = TrOCRProcessor.from_pretrained(model_name)
File "/home/user/.local/lib/python3.10/site-packages/transformers/processing_utils.py", line 1394, in from_pretrained
args = cls._get_arguments_from_pretrained(pretrained_model_name_or_path, **kwargs)
File "/home/user/.local/lib/python3.10/site-packages/transformers/processing_utils.py", line 1453, in _get_arguments_from_pretrained
args.append(attribute_class.from_pretrained(pretrained_model_name_or_path, **kwargs))
File "/home/user/.local/lib/python3.10/site-packages/transformers/models/auto/image_processing_auto.py", line 489, in from_pretrained
raise initial_exception
File "/home/user/.local/lib/python3.10/site-packages/transformers/models/auto/image_processing_auto.py", line 476, in from_pretrained
config_dict, _ = ImageProcessingMixin.get_image_processor_dict(
File "/home/user/.local/lib/python3.10/site-packages/transformers/image_processing_base.py", line 333, in get_image_processor_dict
resolved_image_processor_files = [
File "/home/user/.local/lib/python3.10/site-packages/transformers/image_processing_base.py", line 337, in <listcomp>
resolved_file := cached_file(
File "/home/user/.local/lib/python3.10/site-packages/transformers/utils/hub.py", line 322, in cached_file
file = cached_files(path_or_repo_id=path_or_repo_id, filenames=[filename], **kwargs)
File "/home/user/.local/lib/python3.10/site-packages/transformers/utils/hub.py", line 524, in cached_files
raise OSError(
OSError: PermissionError at /home/user/.cache/huggingface/hub/models--microsoft--trocr-base-handwritten when downloading microsoft/trocr-base-handwritten. Check cache directory permissions. Common causes: 1) another user is downloading the same model (please wait); 2) a previous download was canceled and the lock file needs manual removal.
LLMs are so bad at code sometimes. This happens all the time time with LLMs and code for me, the code is unusable and it saves no time because it's a rabbit hole leading to nowhere.
I also don't know if this is the right approach to the problem. Any sort of GUI interface would be easier. This is also hundreds of pages of handwritten stuff I want to change to text.
This error looks like it is saying a previous attempt aborted, and it needs you to clean up some file that was only partly downloaded.
Edit: The "please wait" makes me think I would try again in a couple hours.
So try again... in a couple of hours...
Why would that make a difference? It's a local model right?
If it is local only, then waiting probably won't help.
Another thought for you: pip behaves much better inside a virtual environment - using the Python venv module, or uv.
The instructions you have shared so far look more compatible with venv.
I don't understand what venv is or why this would work better. Will this make the compatibility issues go away? I could also just create a virtual Ubuntu environment that's fresh if that would be easier and try to give that environment access to my GPU but I don't know if that would work.
I don't understand what venv is or why this would work better.
Venv is a Python module that helps isolate sets of Python modules from the system installed Python version.
Will this make the compatibility issues go away?
No guarantees, but it often does.
I could also just create a virtual Ubuntu environment that's fresh if that would be easier and try to give that environment access to my GPU but I don't know if that would work.
I'm not sure either. But you've got the idea. Python packages install better when they're allowed to exist separately from the underlying operating system.
You could try codidact. I'm not sure what kind of gatekeeping they have around account creation, but the project basically started up with "Stackoverflow sucks now, let's make our own!"
Wow, I had never heard of that before and thought it was a typo.
https://codidact.com/ also links to https://topanswers.xyz/ to check for communities they might not have.
Hard to tell how active they are but interesting to check out for sure.
Piefed iirc has question answer support like Stackoverflow. Where you can choose an accepted answer. So you could replicate stackoverflow with a piefed community, and something like that might already exist. I'm not aware of it if it does though.
Man, I need to get to switching to Piefed soon, it seems nice. Any cons vs Lemmy?
which community? i did look and didn't see one.
There soon will be. They are still trying to scrub all of the helpful information out there and, once done, all you need is an AI subscription to get you moving again. /s
I don't have an answer for you I'm sorry.
I think I asked a question on S/O once, like 10 years ago, and I got the usual "this question has been asked before" bullshit so I just never tried again.
These days I mostly just search for technical answers, yes often the answers are on S/O but often they're elsewhere.
I also use a self hosted instance of openwebui to chat with LLMs hosted by huggingface. You're correct that often technical answers are insufficient, but often it's helpful. For example today I was trying to measure contrast between 2 colors. The chat bot gave me theory, explanation, and a working spreadsheet formula in a few seconds. It didn't complain about how I'd formatted my question or that someone else may have asked the same question.
i got that a lot when i first started using linux-based distros. "why aren't you typing man?" or "just type sudo rmdir /"
Why not turn off the VPN to make the account then use the VPN like normal afterwards?
hell fucking no. I do not need some large corporate company taking a) my browser fingerprint b) my real IP address and ping time and c) selling them to every data broker that exists. Fuck no. Buying a residential IP costs less than a dollar. Big data does not need to associate my online activities with who I am IRL.
It's also a terrible idea because a lot of my online protection is because no one know my initial ping time, so I would be giving big data information like "here's this user with browser hash ########### and they have a ping time of X ms and origination IP of X.X.X.X." so after that, even if I protected my origination IP later, they could guess who I am based on hops and ping time and browser fingerprint, because even privacy browsers have a browser fingerprint. Fuuuuuuuuuuuuuuck no.
What VPN provider are you using because I guarantee you are leaking far more data to them then to anywhere else.
How do you know your VPN is not just doing the same things? Anything that you think is secure is only so temporarily because it becomes a target for state and private security firms. E.g. a private security company from Sweden could infiltrate a VPN provider than that security firm sells the data to states and other private firms. These firms are owned by the same conglomerates that own everything else and they consolidate the data. That is giving the VPN provider the benefit of the doubt, that they aren't just lying to you.
because it's encrypted by https and my VPN has a decent reputation. Yes, it's possible that the VPN is secretly selling everything they can to big data, but if it were exposed then it would ruin their entire business.
https://www.youtube.com/watch?v=1opKW6X88og
"Inside this video, you’ll learn: • How Facebook turned a “privacy app” into a surveillance weapon • The Israeli cyber intel unit behind Onavo and why it matters • What Project Ghostbusters did to break HTTPS encryption • Why 20+ top VPNs are secretly owned by spyware vendors • The real story behind ExpressVPN, Kape Technologies, and fake “independent” review sites • The 7-point checklist every VPN must pass to be trusted • Better tools to protect yourself: DoH, hardened Firefox, Tor, browser isolation, and more"
Stack Overflow thrives on participation. If you do not want to participate in any way then SO is probably not for you.
I am happy to participate. I do not want my IP and ping data sold to data brokers to serve targeted ads, track me, and go to the police surveillance state. I'm sure you are a good citizen who always keeps location on and feels like life would be easier if everyone just complied while you proudly put ring cameras on every door. Not everyone is a tech bro neo-feudalist.
I like your disdain for mass surveillance, but don't call me a tech bro
Are these questions that an LLM can't answer? For better or for worse, SO has basically died because LLMs can answer a great majority of SO-type questions. So I guess my question to you is: what are the nature of your queries? Something some bullshit LLM can't answer? Well by golly, then let's see if we can't answer them here!
GPT 5 gave me a lot of code that returned errors. I really need help with the specific terminal code or knowing if I am even approaching the problem right.
No, sorry. Never. I believe in mind training and development, human discussions, attribution, contribution, and discoveries for human work, imagination, achievements, and scientific miracles.
You? You do you with your LLM.
It's a shame it's not around anymore. You sound like you would have fit in nicely.