Hi!
Thank you for open-sourcing this great work! I have a question about the overlap value during inference. In the code comments, overlap is set to 0.0 for inference (non-training), which doubles the data volume and results in TTE values 60 and 45. However, the paper mentions evaluating TTE between 1–2 seconds (30–60), which seems inconsistent with overlap=0. Additionally, during my reproduction of the original code, the inference overlap remained the same as training. Could you clarify the correct overlap setting for inference?
overlap = opts['overlap'] # if data_type == 'train' else 0.0 olap_res = obs_length if overlap == 0 else int((1 - overlap) * obs_length) olap_res = 1 if olap_res < 1 else olap_res for k in d.keys(): seqs = [] for seq in d[k]: start_idx = len(seq) - obs_length - time_to_event[1] end_idx = len(seq) - obs_length - time_to_event[0] seqs.extend([seq[i:i + obs_length] for i in range(start_idx, end_idx + 1, olap_res)]) d[k] = seqs
Hi!
Thank you for open-sourcing this great work! I have a question about the overlap value during inference. In the code comments, overlap is set to 0.0 for inference (non-training), which doubles the data volume and results in TTE values 60 and 45. However, the paper mentions evaluating TTE between 1–2 seconds (30–60), which seems inconsistent with overlap=0. Additionally, during my reproduction of the original code, the inference overlap remained the same as training. Could you clarify the correct overlap setting for inference?
overlap = opts['overlap'] # if data_type == 'train' else 0.0 olap_res = obs_length if overlap == 0 else int((1 - overlap) * obs_length) olap_res = 1 if olap_res < 1 else olap_res for k in d.keys(): seqs = [] for seq in d[k]: start_idx = len(seq) - obs_length - time_to_event[1] end_idx = len(seq) - obs_length - time_to_event[0] seqs.extend([seq[i:i + obs_length] for i in range(start_idx, end_idx + 1, olap_res)]) d[k] = seqs