braindotai / watermark-removal-pytorch

🔥 CNN for Watermark Removal using Deep Image Prior with Pytorch 🔥.

Yo, welcome to the 🧠 BrainBust universe.

Hi there , I am Rishik, and I feel most alive 😃 in those 10 seconds 🕐, that go through me running that code and seeing the loss converge 📉.

👉🏻👈🏻🥺 Who am I?	💗 Things I love
💻 I'm a Full Stack 🤖 ML Engineer and 🌎 Web Developer.	🎶🎶 in background.
👷‍♂️ Working full time with a startup ⭐ SkylarkLabs.ai, where I have developed and deployed 5 products to far.	💯% filling VRAM.
🏆 Won Covid ML Hackathon with 1st rank in my college.	🍔 Double cheeseburger.
🌐 You can check out my website, 🔨 built with Angular and Django.	🥤 Chilled soda.
📺 I'm the creator of BrainBust Youtube channel. I don't like that subscribe button either, so destroy it, kill it, suck its life out 👻.	🎹 Playing Piano.
📰 Working on my very first research publication.	😗 Whistling favourite songs.
🔠 Currently mastering NLP.	🏡 And home sweet home.

👨🏻‍🏫 # Technical Skills

🤖 Machine Learning / Deep Learning	🏭 Backend	🌄 Frontend

📝 # Written Articles

• 📈 Averaging Weights Leads to Wider Optima and Better Generalization
• 🌊 Watermark Removal Using Deep Image Priors
• 👦🏻 Facial Landmarks Detection Using Xception Net
• 🎨 Fake Anime Characters Using Deep Convolutional GANs
• ⚡ Best Performance Tuning Practices for Pytorch
• 😮 How Not to Train a Deep Learning Model?
• ➕ Resolving CUDA Being Out of Memory With Gradient Accumulation and AMP
• 💡 Supercharging automation in terminal with python

🔨 # Deployed Projects

Facial Landmarks Detection	Fake Anime Generation

Watermark Removal	Violence Detection

Lie Detection	Night 2 Day

📊 # Profile Stats

Wubba lubba dub dub!!! You made it to the end!!!

watermark-removal-pytorch's Issues

[RuntimeError: CUDA out of memory] when max dimensions of final image are increased

I am getting the following error when I increase the --max-dim parameter above 700 on a 6GiB GPU.

RuntimeError: CUDA out of memory. Tried to allocate 208.00 MiB (GPU 0; 6.00 GiB total capacity; 5.29 GiB already allocated; 0 bytes free; 5.33 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

I tried reducing the --input-depth to see if this would reduce the memory usage, but it didn't help.

Are there any other possible tweaks that would allow for processing of larger images?

Default parameters

Hi,
Could you please give us some correct, default parameters for :

max_dim=MAX_DIM,
show_step=SHOW_STEPS,
reg_noise=REG_NOISE,
input_depth=INPUT_DEPTH,
lr=LR,
training_steps=TRAINING_STEPS

on video ?

would this model be available on videos ?

Need some data on performance of this application

First of all, Thank you very much for this program. It is working great for an image.

Currently, I am running this application on a windows laptop and was able to execute the program python inference.py.
I do have training-steps as 3000 which is the default. i have prepared a masked image to remove the watermark from the original image. it is taking nearly 6hrs for an image to be processed and remove the watermark.

I wanted to know if I am doing anything wrong or will it be better if I run on some computer with good CPU power.

retrain

is it required that every time retrain for each image?

Loss of resolution

Hello,

Thank you for implementing this great pipeline! I successfully ran it but I saw the resolution of output is much lower than the original input. The loss of resolution also happended to those unmasked areas without the watermark. I set the output dimention the same as the input dimention. Is that a way to improve the resolution and avoid changing the quality of unmasked regions?

Thanks,
Shujun

Processing was stopped out , Upsampling is deprecated

Starting training...

Completed: 0%| | 0/9 [00:00<?, ?it/s]C:\Program Files\Python36\lib\site-packages\torch\nn\modules\upsampling.py:122: UserWarning: nn.Upsampling is deprecated. Use nn.functional.interpolate instead.
warnings.warn("nn.Upsampling is deprecated. Use nn.functional.interpolate instead.")
Completed: 100%|█████████████████████████████████████████| 9/9 [00:12<00:00, 1.43s/it, Loss=0.0263]

Saving final output image to: "E:\Watermark-Removal-Pytorch\demo\frames\00000-output.jpg"

Broken can't install it even with build tools

gi\gimodule.c: fatal error C1083: Impossible d'ouvrir le fichier includeÿ: 'msvc_recommended_pragmas.h'ÿ: No such file or directory
      error: command 'C:\\Program Files (x86)\\Microsoft Visual Studio\\2022\\BuildTools\\VC\\Tools\\MSVC\\14.37.32822\\bin\\HostX86\\x86\\cl.exe' failed with exit code 2
      [end of output]

  note: This error originates from a subprocess, and is likely not a problem with pip.
  ERROR: Failed building wheel for PyGObject
Failed to build PyGObject
ERROR: Could not build wheels for PyGObject, which is required to install pyproject.toml-based projects 

Model learns to synthesize watermark too

Hi @braindotai ,

First of all, thank you!! This is definitely a nice wrapper for the original implementation, which is quite poorly documented.

I am in the second case: watermark not available; I am using a new file of my own. However, I am getting nowhere close to your results. The model learns to reproduce the watermark.

Am I doing sth wrong? Maybe the image format of the watermark?

Cheers.

Not an issue. wondering if the set up is correct.

after changing the image in the watermarked folder.
and masks in the masks folder.

then runpython3 inference.py

(base) x@xnoMacBook-Pro Watermark-Removal-Pytorch % python3 inference.py   
Setting device to "cpu", since torch is not built with "cuda" support...
It is recommended to use GPU if possible...
Visualizing mask overlap...

got this message for a while. Can you please confirm if the code is running as expected? Thank you.

How to use the script?

The script looks great! Unfortunately I am not a developer.
May I ask how to use it?

I first downloaded it to my Mac, with GitHub Desktop.
Then I run the following in the terminal

cd /Users/MyUserName/Documents/GitHub/Watermark-Removal-Pytorch/
python inference.py

It has the following error

Traceback (most recent call last):
File "inference.py", line 2, in
from api import remove_watermark
File "/Users/MyUserName/Documents/GitHub/Watermark-Removal-Pytorch/api.py", line 63
print(f'\nSaving final output image to: "{output_path}"\n')
^
SyntaxError: invalid syntax

Could you explain a little where I did it wrong?
Thanks in advance.
I would love to try this great script.
Thanks for sharing

Licence?

Hello,

really nice project!
Under which licence is it published?

Instructions

How do we use this?

CPU usage.

When set to CPU mode this sent a 13900k into 100% and made the cpu pulse/squeal with electronic noise.
I have never had this before even stress testing.
I thought a hard drive was getting read its that loud!

Using the same model for the batch of images

Is it possible to use the same model for removing the watermark from thousands of images if watermark mask is the same?

Inference time

Really nice project! Thank you for making it public!

The question is: what's the time of inference on image?
Actually my goal is to remove watermark in the bottom rigth corner of the video. I've already prepared the docker container and simple code to divide video into frames. How long will it take to run on 1k frames?

I can't actually check for now because video is not currently available. (I'll post my results when I finish)

Specs:

• i5 12600kf
• 16gb RAM 3200MHz
• RTX 3080 (10 GB version)

requirements.txt

Consider adding a requirements.txt file.
Dependencies I noticed:

torch
torch vision
numpy
tqdm
pillow
matplotlib

An iteration takes on my laptop dozens of seconds (cpu only)

How long should an iteration take?

Is there any chance this will work quickly on my i7-4600u CPU with 10gb of ram on windows 10?

Currently I have an 1800x1200 image

Does it work on CPU?

TypeError: 'NoneType' object does not support item assignment

So, after last update (5 days ago) does not work/
But before that update too didnt work