Hi @v.chung , I am really sorry to disturb you even after the competition deadline. I found a very critical issue when loading my model, although it can still generate the predictions file (the auc scores may be wrong). If the model loaded correctly, the prediction score for SC1 leaderboard would be around 0.39 based on training data only. Could you give me an opportunity to submit the docker again (I think I only need to submit once)? I really hope my model could represent some results on Dream Challenge competition, because it is very important for our team to make next exploration on research. The only difference in my current code is the commentend lines here: ``` checkpoint = torch.load(train_data_path + 'interaction-model-final.pt', map_location=torch.device('cpu')) load_states = list(interaction_model.state_dict().keys()) pretrained_interaction_model_dict = {key: checkpoint.get(key) for key in load_states} # interaction_model_dict = interaction_model.state_dict() # pretrained_interaction_model_dict.update(interaction_model_dict) interaction_model.load_state_dict(pretrained_interaction_model_dict) ``` Many thanks and sorry again, Tongtong