![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8448924%2F15564725296900e1f17b99106f400e06%2Fcmi_overview.png?generation=1701834611847219&alt=media\")
\nWe utilized @tubotubo 's excellent notebook and code, implementing the following improvements.
First of all, we express our gratitude to the hosting team for organizing this intriguing competition.
\nThis was my first chance to tackle the challenge of event detection, and I gained valuable insights from participating in this competition.
\n
\nWe utilized @tubotubo 's excellent notebook and code, implementing the following improvements.
I will explain these approaches in the order in which we tackled them in this competition.
\nIn this event detection task, we speculated that accurate predictions could be achieved by ensemble averaging the outputs of multiple models. Therefore, we applied the following ensemble approach:
\nWe trained models with input durations of 8 hours (5760 steps), 16 hours (11520 steps), and 24 hours (17280 steps). During inference, we took the average of the outputs from these models.\"
\nFor the encoder model, we adopted timm-gernet-l, resnext101, and se-resnet50, and averaged the outputs of these three models, similar to the duration ensemble.
\nWith the idea from @sugupoko, we inserted BiGRU into the skip connections of the 1D-Unet decoder. We believe that this has improved the prediction accuracy in the temporal dimension.
\nWhile conducting several experiments, we noticed that the parameters of the scipy.signal.find_peaks function, used in post-processing, have an impact on accuracy.
\nInitially, we lowered the height (score_th) to 0.001, as setting a lower threshold is advantageous in terms of the evaluation metric for this competition. Subsequently, we adjusted the distance to 65. This parameter determines the distance between peaks, and we are confident that setting an appropriate value helped reduce false positives.
While analyzing the data, we noticed that some input data contained unnatural periodic signals. In many cases, these signals lacked event labels and could potentially be a cause of false positives.
\nWe didn't understand the reason for the presence of such signals, but we found that they occurred approximately every 24 hours (17280 steps). Therefore, we attempted to remove these repeating signals using a rule-based approach.
\nWe created a function to detect repeating signals as follows. We segmented the anglez data at intervals of 17280 steps, calculating the correlation coefficient between the previous and current segments. If this correlation coefficient exceeded a certain threshold, it was determined to be a repeating signal. By invalidating the model output for the detected repeating segments, we were able to significantly reduce false positives.
def (series_id, cycle_length=, cor_th=, phase='test'):\n anglez = np.(f)\n \n valid_area = np.(anglez.shape[])\n x1_list = ((, anglez.shape[], cycle_length))\n pre_anglez = np.(cycle_length)\n for x1 in x1_list:\n x2 = x1 + cycle_length\n cur_anglez = anglez[x1:x2]\n \n if (cur_anglez) < cycle_length:\n cur_anglez = np.(cur_anglez, (, cycle_length - (cur_anglez)), constant_values=)\n \n correlation_coefficient_anglez = np.(pre_anglez, cur_anglez)[, ]\n \n if correlation_coefficient_anglez > cor_th:\n valid_area[x1-cycle_length:x2] = \n \n pre_anglez = cur_anglez\n \n return valid_area\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8448924%2F901ddaca907928fafc411dd08c7163e3%2Fdetect_repeating_signals.png?generation=1701834185197721&alt=media\")
First of all, we express our gratitude to the hosting team for organizing this intriguing competition.
\nThis was my first chance to tackle the challenge of event detection, and I gained valuable insights from participating in this competition.
\n
\nWe utilized @tubotubo 's excellent notebook and code, implementing the following improvements.
I will explain these approaches in the order in which we tackled them in this competition.
\nIn this event detection task, we speculated that accurate predictions could be achieved by ensemble averaging the outputs of multiple models. Therefore, we applied the following ensemble approach:
\nWe trained models with input durations of 8 hours (5760 steps), 16 hours (11520 steps), and 24 hours (17280 steps). During inference, we took the average of the outputs from these models.\"
\nFor the encoder model, we adopted timm-gernet-l, resnext101, and se-resnet50, and averaged the outputs of these three models, similar to the duration ensemble.
\nWith the idea from @sugupoko, we inserted BiGRU into the skip connections of the 1D-Unet decoder. We believe that this has improved the prediction accuracy in the temporal dimension.
\nWhile conducting several experiments, we noticed that the parameters of the scipy.signal.find_peaks function, used in post-processing, have an impact on accuracy.
\nInitially, we lowered the height (score_th) to 0.001, as setting a lower threshold is advantageous in terms of the evaluation metric for this competition. Subsequently, we adjusted the distance to 65. This parameter determines the distance between peaks, and we are confident that setting an appropriate value helped reduce false positives.
While analyzing the data, we noticed that some input data contained unnatural periodic signals. In many cases, these signals lacked event labels and could potentially be a cause of false positives.
\nWe didn't understand the reason for the presence of such signals, but we found that they occurred approximately every 24 hours (17280 steps). Therefore, we attempted to remove these repeating signals using a rule-based approach.
\nWe created a function to detect repeating signals as follows. We segmented the anglez data at intervals of 17280 steps, calculating the correlation coefficient between the previous and current segments. If this correlation coefficient exceeded a certain threshold, it was determined to be a repeating signal. By invalidating the model output for the detected repeating segments, we were able to significantly reduce false positives.
def (series_id, cycle_length=, cor_th=, phase='test'):\n anglez = np.(f)\n \n valid_area = np.(anglez.shape[])\n x1_list = ((, anglez.shape[], cycle_length))\n pre_anglez = np.(cycle_length)\n for x1 in x1_list:\n x2 = x1 + cycle_length\n cur_anglez = anglez[x1:x2]\n \n if (cur_anglez) < cycle_length:\n cur_anglez = np.(cur_anglez, (, cycle_length - (cur_anglez)), constant_values=)\n \n correlation_coefficient_anglez = np.(pre_anglez, cur_anglez)[, ]\n \n if correlation_coefficient_anglez > cor_th:\n valid_area[x1-cycle_length:x2] = \n \n pre_anglez = cur_anglez\n \n return valid_area\n