Thank you everyone and congratulations!
\nI thought winning silver medal, but small shake down:(
\nI will continue to challenging other kaggle competitions from now on.
I don't have signal processing knowledge, so this is very interesting competition.
\nI ensembled below 21 models and weighted oof ridge coefficients.
\nEnsemble result LB:0.8807 PB:0.8783
\nI tried all models containg other model and ridge weights. In this case, some model weights are negative, so I normalize to [0, 1] forcely. (LB:0.8808 PB:0.8783)
\nHowever, ensemble using only positive weights is good for PB.
This is power solution.
\nMy trying and questions are below, so could you tell me your opinion?
\nFirstly, I tried CQT, but CV score was not improved. Therefore, I used CWT and getting CV 0.874~.
\nLarger model and larger input size gave me CV:8.765~ LB:0.879~.
\nex)
\nFor ensembling, I retried using CQT, but score was worse than CQT.
\nWhich did you use ?
\nPlease tell me your tricks for CQT and CWT.
I tried 1d-cnn and got CV:0.8669 LB:0.8759 .
\nI found larger kenrel size is better than stacking small kernel size, but CV was saturated. I tried 1d-conv transformer-encoder model, but score became worse.
\nIn this competittion, tensorflow could get higher score than pytorch model.
\nFor some reason, I implemented 1d-cnn using pytorch.
Does anyone get high score using 1d-cnn implemented with pytorch?
\nI used ViT and SWIN whch is lower CV but boost ensemble.
\nIf you use ViT & SWIN, please tell me your cv and lb.
\nI tried Auto-encoder, and tried to use it for anomaly detection.
\nHowever, this was not good.
Does anyone succeed in auto-encoder approach?
\nThanks
", "messages": [ { "id": "1529403", "postDate": "09/30/2021 11:10:26", "content": "Thank you everyone and congratulations!
\nI thought winning silver medal, but small shake down:(
\nI will continue to challenging other kaggle competitions from now on.
I don't have signal processing knowledge, so this is very interesting competition.
\nI ensembled below 21 models and weighted oof ridge coefficients.
\nEnsemble result LB:0.8807 PB:0.8783
\nI tried all models containg other model and ridge weights. In this case, some model weights are negative, so I normalize to [0, 1] forcely. (LB:0.8808 PB:0.8783)
\nHowever, ensemble using only positive weights is good for PB.
This is power solution.
\nMy trying and questions are below, so could you tell me your opinion?
\nFirstly, I tried CQT, but CV score was not improved. Therefore, I used CWT and getting CV 0.874~.
\nLarger model and larger input size gave me CV:8.765~ LB:0.879~.
\nex)
\nFor ensembling, I retried using CQT, but score was worse than CQT.
\nWhich did you use ?
\nPlease tell me your tricks for CQT and CWT.
I tried 1d-cnn and got CV:0.8669 LB:0.8759 .
\nI found larger kenrel size is better than stacking small kernel size, but CV was saturated. I tried 1d-conv transformer-encoder model, but score became worse.
\nIn this competittion, tensorflow could get higher score than pytorch model.
\nFor some reason, I implemented 1d-cnn using pytorch.
Does anyone get high score using 1d-cnn implemented with pytorch?
\nI used ViT and SWIN whch is lower CV but boost ensemble.
\nIf you use ViT & SWIN, please tell me your cv and lb.
\nI tried Auto-encoder, and tried to use it for anomaly detection.
\nHowever, this was not good.
Does anyone succeed in auto-encoder approach?
\nThanks
", "rawMarkdown": "Thank you everyone and congratulations!\n\nI thought winning silver medal, but small shake down:(\nI will continue to challenging other kaggle competitions from now on.\n\nI don't have signal processing knowledge, so this is very interesting competition.\n\n# Solution\n\nI ensembled below 21 models and weighted oof ridge coefficients.\n\nEnsemble result LB:0.8807 PB:0.8783\n\n- CWT(256x256) + EffNetB7 : CV:0.8750 LB:0.8785\n- CWT(384x384) + EffNetB7 : CV:0.87613 LB:0.8790\n- CWT(448x448) + EffNetB7 : CV: 0.87643 LB: 0.8789\n- CWT(384x384) + Input scaling to similar order + EffNetB7 : CV:0.87608 LB:0.8791\n- CWT(512x512) + EffNetV2M : CV:0.87614 LB:0.8788\n- CWT(576x576, wavelet_width=8) + EffNetV2M : CV:0.87650 LB:0.8792\n- CWT(586x586, wavelet_width=6) + EffNetV2M : CV:0.87647 LB:0.8791\n- CWT(420x420) + EffNetB5 : CV:0.8752 LB:0.8788\n- CWT(512x512, wavelet_width=6) + DenseNet201 : CV:0.87527 LB:0.8782\n- CWT(640x640) + EffNetB5 : CV:0.8710 LB:0.8774\n- CWT(576x576, wavelet_width=6) + EffNetV2M : CV:0.87652 LB:0.8790\n- bandpass + CWT(?x?, ) + ResNet34 : \n- bandpass + CWT(576x576, wavelet_width=6) + EffNetV2M : CV:0.8765 LB:0.8791\n- bandpass + CWT(512x512, wavelet_width=6) + EffNetB7\n- bandpass + CWT(384x384, wavelet_width=6, upper_freq=500) + VitB16\n- bandpass + CWT(448x448) + EffNetB7\n- 1dCNN version0 : CV 0.86508, LB:0.8710\n- 1dCNN version1 : CV:0.8667 Lb:0.8738\n- 1dCNN version2 : CV:0.8651 Lb:0.8728\n- 1dCNN version3 : CV:0.8663 Lb:0.8716\n- 1dCNN version4 : CV:0.8669 LB:0.8759\n- 1dCNN version5 : CV:0.8689 LB:0.8740\n\nI tried all models containg other model and ridge weights. In this case, some model weights are negative, so I normalize to [0, 1] forcely. (LB:0.8808 PB:0.8783)\nHowever, ensemble using only positive weights is good for PB.\n\nThis is power solution.\n\nMy trying and questions are below, so could you tell me your opinion?\n\n# Q1. CWT vs CQT\n\nFirstly, I tried CQT, but CV score was not improved. Therefore, I used CWT and getting CV 0.874~.\n \nLarger model and larger input size gave me CV:8.765~ LB:0.879~.\n\nex)\n- EfficientNet b7 , 512x512 + bpf CV:0.8765 LB:0.8792\n- EfficietnNet V2M, 576x576 + bpf CV:0.8765 LB:0.8792\n\nFor ensembling, I retried using CQT, but score was worse than CQT.\n\nWhich did you use ?\nPlease tell me your tricks for CQT and CWT.\n\n\n# Q2. 1d-cnn using raw signal\n\nI tried 1d-cnn and got CV:0.8669 LB:0.8759 .\n\nI found larger kenrel size is better than stacking small kernel size, but CV was saturated. I tried 1d-conv transformer-encoder model, but score became worse.\n\nIn this competittion, tensorflow could get higher score than pytorch model.\nFor some reason, I implemented 1d-cnn using pytorch.\n\nDoes anyone get high score using 1d-cnn implemented with pytorch?\n\n# Q3. ViT & SWIN \n\nI used ViT and SWIN whch is lower CV but boost ensemble.\n\nIf you use ViT & SWIN, please tell me your cv and lb.\n\n# Q4. Auto-encoder\n\nI tried Auto-encoder, and tried to use it for anomaly detection.\nHowever, this was not good.\n\nDoes anyone succeed in auto-encoder approach?\n\n\nThanks", "votes": null }, { "id": "1559882", "postDate": "10/27/2021 08:05:35", "content": "Hey All,
\nThank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey https://forms.gle/QP9L16niPexozyhu5.
\nThank you all,
\nRegards,
\nChris
Hi @yoshito
\nI realize the competition is long over. But I wanted to figure something out as I am doing some research. EffnetB7 for 128x128 image gives me a CV score 0.83. How do all of you get to 0.87 or higher? Is there some particular normalization I am missing?
\nI am doing bandpass +CWT (normalize image to 255) + EffnetB7
\nI am not using the Virgo channel but surely that cannot add that big of a gain? I am also using smaller image, but I doubt using 256x256 gives a huge boost. Let me know if I am missing something obvious.
\nThanks,
\nAkshay
Hey All,
\nThank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey https://forms.gle/QP9L16niPexozyhu5.
\nThank you all,
\nRegards,
\nChris
Hi @yoshito
\nI realize the competition is long over. But I wanted to figure something out as I am doing some research. EffnetB7 for 128x128 image gives me a CV score 0.83. How do all of you get to 0.87 or higher? Is there some particular normalization I am missing?
\nI am doing bandpass +CWT (normalize image to 255) + EffnetB7
\nI am not using the Virgo channel but surely that cannot add that big of a gain? I am also using smaller image, but I doubt using 256x256 gives a huge boost. Let me know if I am missing something obvious.
\nThanks,
\nAkshay