Description
I am writing to request more detailed documentation and clarification regarding the pre-training process in this repository. Currently, there are several significant discrepancies between the provided codebase and the published paper, making it difficult to reproduce the results.
Could you please provide explanations for the following points?
1. Pre-training Hyperparameters and Methodology
Epochs: What is the specific number of epochs used for the pre-training stage? This is not explicitly detailed in the paper.
Quantile Approach: I noticed pretrain-quantile in the code. Was the pre-training conducted using a quantile-based approach? This is not explicitly detailed in the paper.
2. Architecture and Task Discrepancies
Masking vs. Prediction: The current code implementation appears to use a combination of Masking + Prediction. However, the paper suggests that only Prediction was utilized. Could you clarify which approach was used for the final results?
Data Representation (img_size): There is a parameter named img_size. Does the model internally process time-series data as images? If so, could you explain the transformation logic?
3. Dataset and Script Consistency
Pre-training Dataset: While pretrain_quantile is configured to use the entire pre-training dataset, the standard pretrain.py script does not seem to follow the same logic. Is there a reason for this inconsistency?
Ambiguous Scripts: There are multiple overlapping scripts and learners. It is unclear which one is the "official" entry point for reproduction. Could you specify the correct scripts for the following?
Main Scripts: pretrain.py vs. pretrain_quantile.py
Task Logic: patch_mask_2task_predict.py vs. patch_mask.py
Learners: learner_2task.py vs. learner.py vs. learner_2task_quantile.py
Conclusion
There are many major and critical differences between the paper and the repository. A clearer guide or a brief explanation of these implementation details would be greatly appreciated for the community to properly build upon your work.
I look forward to your guidance.
Description
I am writing to request more detailed documentation and clarification regarding the pre-training process in this repository. Currently, there are several significant discrepancies between the provided codebase and the published paper, making it difficult to reproduce the results.
Could you please provide explanations for the following points?
1. Pre-training Hyperparameters and Methodology
Epochs: What is the specific number of epochs used for the pre-training stage? This is not explicitly detailed in the paper.
Quantile Approach: I noticed pretrain-quantile in the code. Was the pre-training conducted using a quantile-based approach? This is not explicitly detailed in the paper.
2. Architecture and Task Discrepancies
Masking vs. Prediction: The current code implementation appears to use a combination of Masking + Prediction. However, the paper suggests that only Prediction was utilized. Could you clarify which approach was used for the final results?
Data Representation (img_size): There is a parameter named img_size. Does the model internally process time-series data as images? If so, could you explain the transformation logic?
3. Dataset and Script Consistency
Pre-training Dataset: While pretrain_quantile is configured to use the entire pre-training dataset, the standard pretrain.py script does not seem to follow the same logic. Is there a reason for this inconsistency?
Ambiguous Scripts: There are multiple overlapping scripts and learners. It is unclear which one is the "official" entry point for reproduction. Could you specify the correct scripts for the following?
Main Scripts: pretrain.py vs. pretrain_quantile.py
Task Logic: patch_mask_2task_predict.py vs. patch_mask.py
Learners: learner_2task.py vs. learner.py vs. learner_2task_quantile.py
Conclusion
There are many major and critical differences between the paper and the repository. A clearer guide or a brief explanation of these implementation details would be greatly appreciated for the community to properly build upon your work.
I look forward to your guidance.