search

Maelic / SGG-Benchmark

A New Benchmark for Scene Graph Generation, targeting real-world applications
MIT License
44 stars 5 forks source link

SGDET_on_cutom_images #32

Open keeganNull opened 1 month ago

keeganNull commented 1 month ago

image

keeganNull commented 1 month ago

how i import the model, config image the model was downloaded from your oneDrive

Maelic commented 1 month ago

Hi, you have to put the config.yml and the .pth file under a new ./checkpoints/VG150-penet-yolov8m/ directory at the root of the codebase otherwise it won't work. Alternatively, you can change the variable PATHS_CATALOG in your config.yml file to the absolute path of the file in your system (something like THE_PATH_YOU_INSTALL_THE_CODEBASE/SGG-Benchmark/sgg_benchmark/config/paths_catalog.py)

keeganNull commented 1 month ago

I did as you told, but it still didn't work. After check, I found out it caused by the path in paths_catalog.py. image And I'm new of SGG, I want to transfer random image to scene graph. Can you tell me what I need to do. thanks.

keeganNull commented 1 month ago

After trying, I found out that the weight download from MODEL_ZOO does not match the size of model predefined. size mismatch for roi_heads.relation.union_feature_extractor.feature_extractor.pooler.reduce_channel.0.weight: copying a param with shape torch.Size([192, 576, 3, 3]) from checkpoint, the shape in current model is torch.Size([192, 192, 3, 3]). It's my problem, or I need train the model on my own.

Maelic commented 1 month ago

Yes I am sorry but the weights in model_zoo are probably outdated. I will remove them to not confuse other people. I may retrain the models and put new weights instead when I have time. In the meantime, you can train your own model with the codebase!

keeganNull commented 1 month ago

Yes I am sorry but the weights in model_zoo are probably outdated. I will remove them to not confuse other people. I may retrain the models and put new weights instead when I have time. In the meantime, you can train your own model with the codebase! It still worked after I change some parameters in the code.Thanks for your code. Btw, I found out that the detection result is not stable, which means yolov8 can't detect all the people.