Using pre-trained model without prompts

[agoodge]

Hello, I have a few questions:

if args.prompt_ST==1: if args.file_load_path != '': model.load_state_dict(torch.load('{}'.format(args.file_load_path),map_location=device), strict=False)
I see that if using prompt_ST, then the model weights will be loaded from args.file_load_path. Why must we use prompts in order to load model weights? What if we want to test the pre-trained model without using prompts?

What is model_SinCos.pkl for in the model weights? It is not used anywhere in the code.

Do you have the pre-trained prompt pool available so that we can reproduce your results?

Can you explain what X_period is used for? How is it different from the regular data X?

Thanks very much!

[YuanYuan98]

Thank you for your questions!

Q1: The reason prompts are necessary for loading model weights is that using prompts introduces additional model parameters, which alters the model's weights.

Q2: The model_SinCos.pkl file contains the trained model parameters. You can download it and place it in the directory specified by args.file_load_path.

Q3: For reproducing results, please refer to the model_SinCos.pkl file, which serves as the pre-trained prompt pool.

Q4: The X_period variable aggregates the corresponding timestamps from the past several days, capturing the periodic patterns inherent in the spatio-temporal data.

Thanks again for your inquiries!

[agoodge]

Thank you for your very fast response!

Then what are the model weights in the model_Emb.pkl file?

Thanks again

[YuanYuan98]

We provide two model parameters: model_SinCos.pkl and model_Emb.pkl. The key difference lies in the positional encoding used; model_SinCos.pkl employs Sine-Cosine positional encoding typical of Transformers, while model_Emb.pkl utilizes learnable positional encoding. Please ensure to specify the correct positional encoding approach in the hyper-parameter with args.pos_emb=SinCos

[agoodge]

It's clearer now. Thank you! It seems their performance is similar.