{ "id": 330173, "title": "AutoEncoders?", "url": "/competitions/denoising-shabby-pages/discussion/330173", "author_name": "", "post_date": "2022-06-10T18:06:51.335478800Z", "votes": null, "comment_count": 6, "views": 0, "content": "

Anyone trying to use Auto Encoders for this problem? I have tried them but getting worst results. In an old and similar competition https://www.kaggle.com/competitions/denoising-dirty-documents many people were using Auto encoders.

", "messages": [ { "id": "1816974", "postDate": "06/10/2022 18:06:51", "content": "

Anyone trying to use Auto Encoders for this problem? I have tried them but getting worst results. In an old and similar competition https://www.kaggle.com/competitions/denoising-dirty-documents many people were using Auto encoders.

", "rawMarkdown": "Anyone trying to use Auto Encoders for this problem? I have tried them but getting worst results. In an old and similar competition https://www.kaggle.com/competitions/denoising-dirty-documents many people were using Auto encoders.", "votes": null }, { "id": "1816979", "postDate": "06/10/2022 18:13:12", "content": "

I'd be curious how U-NETs perform on this task. Because these document images are more realistic, perhaps the task is too challenging for a basic auto encoder?

", "rawMarkdown": "I'd be curious how U-NETs perform on this task. Because these document images are more realistic, perhaps the task is too challenging for a basic auto encoder?", "votes": null }, { "id": "1816993", "postDate": "06/10/2022 18:26:13", "content": "

Yes that can be tried as well. I have found an approach that seems promising. Let me try this approach specially made for this problem :)
\nDE-GAN: A Conditional Generative Adversarial Network for Document Enhancement

", "rawMarkdown": "Yes that can be tried as well. I have found an approach that seems promising. Let me try this approach specially made for this problem :)\n[DE-GAN: A Conditional Generative Adversarial Network for Document Enhancement](https://ieeexplore.ieee.org/document/9187695)", "votes": null }, { "id": "1831441", "postDate": "06/24/2022 07:08:23", "content": "

So the model is able to get a better result?

", "rawMarkdown": "So the model is able to get a better result?", "votes": null }, { "id": "1833047", "postDate": "06/25/2022 16:32:06", "content": "

I was not able to run it due to slow training. I will update soon.

", "rawMarkdown": "I was not able to run it due to slow training. I will update soon.", "votes": null }, { "id": "1833391", "postDate": "06/26/2022 00:52:06", "content": "

I see, is it due to low batch size ? And low batch size is due to memory constraint?

", "rawMarkdown": "I see, is it due to low batch size ? And low batch size is due to memory constraint?", "votes": null }, { "id": "1833777", "postDate": "06/26/2022 10:45:30", "content": "

There were 2 approaches actually:
\n1) https://github.com/dali92002/DocEnTR (Vit-pytorch library I am unable to install the GPU version, it is installing CPU version on my machine although torch GPU is working well)
\n2) https://github.com/dali92002/DE-GAN (Giving black images, https://github.com/dali92002/DE-GAN/issues/18)

", "rawMarkdown": "There were 2 approaches actually:\n1) https://github.com/dali92002/DocEnTR (Vit-pytorch library I am unable to install the GPU version, it is installing CPU version on my machine although torch GPU is working well)\n2) https://github.com/dali92002/DE-GAN (Giving black images, https://github.com/dali92002/DE-GAN/issues/18)", "votes": null } ], "comments": [ { "id": 1816979, "author_name": "jboarman", "author_url": "", "post_date": "06/10/2022 18:13:12", "content": "

I'd be curious how U-NETs perform on this task. Because these document images are more realistic, perhaps the task is too challenging for a basic auto encoder?

", "votes": null, "replies": [ { "id": 1816993, "author_name": "deepdreamx", "author_url": "", "post_date": "06/10/2022 18:26:13", "content": "

Yes that can be tried as well. I have found an approach that seems promising. Let me try this approach specially made for this problem :)
\nDE-GAN: A Conditional Generative Adversarial Network for Document Enhancement

", "votes": null, "replies": [] }, { "id": 1831441, "author_name": "weisinx7", "author_url": "", "post_date": "06/24/2022 07:08:23", "content": "

So the model is able to get a better result?

", "votes": null, "replies": [] }, { "id": 1833047, "author_name": "deepdreamx", "author_url": "", "post_date": "06/25/2022 16:32:06", "content": "

I was not able to run it due to slow training. I will update soon.

", "votes": null, "replies": [] }, { "id": 1833391, "author_name": "weisinx7", "author_url": "", "post_date": "06/26/2022 00:52:06", "content": "

I see, is it due to low batch size ? And low batch size is due to memory constraint?

", "votes": null, "replies": [] }, { "id": 1833777, "author_name": "deepdreamx", "author_url": "", "post_date": "06/26/2022 10:45:30", "content": "

There were 2 approaches actually:
\n1) https://github.com/dali92002/DocEnTR (Vit-pytorch library I am unable to install the GPU version, it is installing CPU version on my machine although torch GPU is working well)
\n2) https://github.com/dali92002/DE-GAN (Giving black images, https://github.com/dali92002/DE-GAN/issues/18)

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1816974": "Anyone trying to use Auto Encoders for this problem? I have tried them but getting worst results. In an old and similar competition https://www.kaggle.com/competitions/denoising-dirty-documents many people were using Auto encoders.", "1816979": "I'd be curious how U-NETs perform on this task. Because these document images are more realistic, perhaps the task is too challenging for a basic auto encoder?", "1816993": "Yes that can be tried as well. I have found an approach that seems promising. Let me try this approach specially made for this problem :)\n[DE-GAN: A Conditional Generative Adversarial Network for Document Enhancement](https://ieeexplore.ieee.org/document/9187695)", "1831441": "So the model is able to get a better result?", "1833047": "I was not able to run it due to slow training. I will update soon.", "1833391": "I see, is it due to low batch size ? And low batch size is due to memory constraint?", "1833777": "There were 2 approaches actually:\n1) https://github.com/dali92002/DocEnTR (Vit-pytorch library I am unable to install the GPU version, it is installing CPU version on my machine although torch GPU is working well)\n2) https://github.com/dali92002/DE-GAN (Giving black images, https://github.com/dali92002/DE-GAN/issues/18)" }, "source": "meta" }