about new pretrained models about roadscene2vec

[malawada]

Hi, you can use the 5_fold_271_carla_sequence_classification_example_model.pt in the folder. It is the same model just named differently. The old version should still work as well, but the new one has updated model weights that may perform better.

[xinxinlv]

hi, thank you for your reply.
I modified the “use_case_2.py” and directly loaded the new pre-trained model “5_fold_271_carla_sequence_classification_example_model.pt” provided by you to predict “use_case_data/lanechange” (without training).
`def risk_assess():
scenegraph_extraction_config = configuration(r"use_case_2_scenegraph_extraction_config.yaml",from_function = True) #create scenegraph extraction config object
extracted_scenegraphs = extract_seq(scenegraph_extraction_config) #extracted scenegraphs for each frame for the given sequence into a ScenegraphDataset
training_config = configuration(r"use_case_2_learning_config.yaml",from_function = True) #create training config object
trainer = Scenegraph_Trainer(training_config) #create trainer object using config
#trainer.split_dataset() #split ScenegraphDataset specified in learning config into training, testing data
# trainer.build_model() #build model specified in learning config
# trainer.learn()

trainer.load_model() 


model_input = format_use_case_model_input(extracted_scenegraphs, trainer) #turn extracted original sequence's extracted ScenegraphDataset into model input
output, _ = trainer.model.forward(*model_input) #output risk assessment for the original sequence 
return output   `

The correct label is risky for “use_case_data/lanechange/22_lanchange”.
But, the direct prediction result is “safe”. This is incorrect. I don't know what went wrong

Model loaded from file. /home/liuxx/anaconda3/envs/av/lib/python3.9/site-packages/torch_geometric/deprecation.py:12: UserWarning: 'data.DataLoader' is deprecated, use 'loader.DataLoader' instead warnings.warn(out) tensor([-0.3435, -1.2354], device='cuda:0', grad_fn=<LogSoftmaxBackward0>)

[malawada]

It is a machine learning model so it will not always predict the correct result :). Our experiments show it predicts the correct result about 90% of the time.