Open mrdbourke opened 1 year ago
Hi, for recognize open-set of tags, you can refer to "RAM Inference on Unseen Categories (Open-Set)" in README.md. And thank you for providing such a concise and amazing code reference, and we will consider implementing this in the feature. Best regards!
Hi there,
Firstly, fantastic work and thank you for sharing!
Second, do you mind providing a small code example to recognize an open-set of tags?
Eg.
from ram import get_transform, inference_ram, inference_tag2text from ram.models import ram, tag2text_caption ram_model = ram(pretrained=ram_checkpoint, image_size=image_size, vit='swin_l').eval().to(device) # Setup image image = ... # Custom tags tags = ["house", "car", "pig"] # Perform inference result = inference_ram(model, image, custom_tags)
Hi, with PR #63 now you can set up custom dataset, whether for open-set or not. Please refer to the "Batch Inference and Evaluation" section of readme, and set up dataset following datasets/openimages-rare-200
.
Here are main steps of setting up datasets:
mkdir datasets/my_dataset/
.
ln -s /path/to/image_folder datasets/my_dataset/imgs
.
Create tag list datasets/my_dataset/my_dataset_ram_taglist.txt
. This is where you put any custom tags. (But for closed-set, these tags should be subset of the label system, i.e., subset of ram/data/tam_tag_list.txt
.)
my_tag_1
my_tag_2
my_tag_3
Create image list datasets/my_dataset/my_dataset_ram_annots.txt
. Ignore annots if you don't care about metrics.
subfolder1/subfolder2/aaa.jpg
subfolder1/subfolder2/bbb.jpg,my_tag_2,my_tag_3
hi @majinyu666, @xinyu1205,
I want to create my custom dataset for tag2text batch inference. But I saw that it has different format, tagidlist.txt
and idannots.txt
. Can you please show an example for each, what they refer to?
Hi there,
Firstly, fantastic work and thank you for sharing!
Second, do you mind providing a small code example to recognize an open-set of tags?
Eg.