search

microsoft / unilm

Large-scale Self-supervised Pre-training Across Tasks, Languages, and Modalities
https://aka.ms/GeneralAI
MIT License
19.08k stars 2.43k forks source link

About the textdiffuser dataset #1552

Open Ruby-He opened 1 month ago

Ruby-He commented 1 month ago

Describe the bug Model I am using textdiffuser, but I had problems when downloading the dataset laion-ocr

The problem arises when using:

img2dataset --url_list="XX/laion_ocr_image_url/laion-ocr-url.txt" --output_folder=laion_ocr --thread_count=64 --resize_mode=no

I found that the index of the downloaded image and the index of the label information in the compressed package laion-ocr.zip do not match.

As shown in the figure below: image

why the same key corresponds to different image url ?

JingyeChen commented 1 month ago

Hi, thanks for your interest in TextDiffuser. Does this issue exist in few cases? Sometimes the image owner may replace the image, resulting in the mismatch of downloaded image.

Ruby-He commented 1 month ago

Hi, thanks for your interest in TextDiffuser. Does this issue exist in few cases? Sometimes the image owner may replace the image, resulting in the mismatch of downloaded image.

Finally, I have to use the index you provided in “laion-ocr-index-url.txt” and write a new script to move and rename the image, so that can match the annotaions.

Another question is could I resize the image directly in the "train.py/inference.py", such as image = Image.open(image_path).convert("RGB")
image = image.resize((512,512))