xiongyiheng/prior-radgraphformer

[GRAIL @ MICCAI 2023] Prior-RadGraphFormer: A Prior-Knowledge-Enhanced Transformer for Generating Radiology Graphs from X-Rays

prior-radgraphformer's Issues

Loading checkpoint weights

Hello,
I am trying to perform inference for an image tensor using the checkpoint weights you provided. My code can be reduced to this

from models.relationformer_2D import build_relationformer as build_model

with open(config_file) as f:
config = yaml.load(f, Loader=yaml.FullLoader)
config = dict2obj(config)

model = build_model(config)

checkpoint = torch.load(checkpoint_path, map_location='cpu')
model.load_state_dict(checkpoint)

img_tensor = torch.load(img_tensor_path)
with torch.no_grad():
output = model(img_tensor)

However, I am getting this error:

Traceback (most recent call last):
File "my_inference.py", line 65, in
main()
File "my_inference.py", line 56, in main
model = load_checkpoint(model, args.checkpoint)
File "my_inference.py", line 41, in load_checkpoint
model.load_state_dict(checkpoint)
File "C:\Users\josen\anaconda3\envs\pradgraph\lib\site-packages\torch\nn\modules\module.py", line 1604, in load_state_dict
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for RelationFormer:
Missing key(s) in state_dict: "backbone.0.body.conv0.weight"(...) "input_proj.3.1.bias".
Unexpected key(s) in state_dict: "net", "optimizer", "scheduler".

How do you recommend loading the weights and performing inference (if I understood the code correctly, inference.py already takes the model outputs(embeddings) as inputs)?

Intention of Converting Entities in RadGraph into Non-sensitive Numbers

Hi, I have a question about the code in preprocessing of RadGraph. What's the intention of converting all entities' names into non-sensitive numbers? I am also working on the same dataset and wondering if there are some special concerns.

Thank you and look forward to hearing from you!

Training weights

Hello,
Could you please provide some training weights of the model?

mapping_delete_meaningless

Could you provide mapping_delete_meaningless.json?