hey guys, I'm one of the kaggle begginers, and I have a question. Why is LSTM or GRU effective for this competition? Is this because datasets are 3D images? Someone already may have talked about this question, but if you guys have some opinions, please teach me for education.
", "messages": [ { "id": 1062182, "postDate": "2020-10-27T16:21:51.547Z", "content": "In addition or alternatively to LSTM and GRU, you could also employ a 1D-CNN in the Z direction (on the GAP embeddings from your stage 1 2D slice models) (Even a random forest here in the Z direction was better than nothing). The important thing is to use many slices (the entire Z direction) to predict the targets. (Even when predicting the target associated with 1 slice i.e. pe_present_on_image using more neighbor slices helps).
thank you for sharing this knowledge
", "rawMarkdown": "thank you for sharing this knowledge", "votes": 1 }, { "id": 1062472, "postDate": "2020-10-27T21:35:23.317Z", "content": "I will also add that many of the top teams used a transformer (which is another kind of NN) in addition (or instead of) LSTM/GRU. Again, the bottom line is to use something that takes all slices as input to help predict the targets.
", "rawMarkdown": "I will also add that many of the top teams used a transformer (which is another kind of NN) in addition (or instead of) LSTM/GRU. Again, the bottom line is to use something that takes all slices as input to help predict the targets.", "votes": 1 }, { "id": 1063398, "postDate": "2020-10-28T21:02:36.140Z", "content": "thank you for sharing your knowledge. I learnt a lot from the idea using more neighbor slices to predict pe_present_on_image. I found that LSTM and GRU were also used in last year RSNA competion, so many of the top teams used them and some teams used Transformer too.
I would also like to point out that 8th place solution didn't use LSTM nor GRU for stage 2. They used LGBM and achieved similar results. Great job team Jan @jpbremer , Yuji-san @yujiariyasu , Yama-san @lisosia
\nI think it is wonderful that we see a variety of model design in the top solutions.
", "rawMarkdown": "I would also like to point out that 8th place solution didn't use LSTM nor GRU for stage 2. They used LGBM and achieved similar results. Great job team Jan @jpbremer , Yuji-san @yujiariyasu , Yama-san @lisosia \n\nI think it is wonderful that we see a variety of model design in the top solutions.", "votes": 3, "replies": [ { "id": 1063423, "postDate": "2020-10-28T22:27:12.867Z", "content": "We used CNN. It also worked, but not so good.
", "rawMarkdown": "We used CNN. It also worked, but not so good.", "votes": 1 }, { "id": 1063436, "postDate": "2020-10-28T23:02:37.347Z", "content": "Nice Jonny. I also ensembled a GRU and 1D-CNN for my stage 2 model. It did better than just GRU.
", "rawMarkdown": "Nice Jonny. I also ensembled a GRU and 1D-CNN for my stage 2 model. It did better than just GRU." } ] }, { "id": 1062152, "postDate": "2020-10-27T15:55:38.920Z", "content": "RNN modules like LSTM, GRU and Transformers are used for sequential data such as time series or text data. In this competition, we predict exam-level labels with many images. So, images or features of images can be considered as sequence indexed in z-pos order. I think that's the reason.
", "rawMarkdown": "RNN modules like LSTM, GRU and Transformers are used for sequential data such as time series or text data. In this competition, we predict exam-level labels with many images. So, images or features of images can be considered as sequence indexed in z-pos order. I think that's the reason.", "votes": 3, "replies": [ { "id": 1063400, "postDate": "2020-10-28T21:04:49.807Z", "content": "thank you for sharing your idea. I agree with you.
", "rawMarkdown": "thank you for sharing your idea. I agree with you." } ] }, { "id": 1061983, "postDate": "2020-10-27T13:36:35.440Z", "content": "hey guys, I'm one of the kaggle begginers, and I have a question. Why is LSTM or GRU effective for this competition? Is this because datasets are 3D images? Someone already may have talked about this question, but if you guys have some opinions, please teach me for education.
", "rawMarkdown": "hey guys, I'm one of the kaggle begginers, and I have a question. Why is LSTM or GRU effective for this competition? Is this because datasets are 3D images? Someone already may have talked about this question, but if you guys have some opinions, please teach me for education.", "votes": 3 } ], "comments": [ { "id": 1062182, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-10-27T16:21:51.547000", "content": "In addition or alternatively to LSTM and GRU, you could also employ a 1D-CNN in the Z direction (on the GAP embeddings from your stage 1 2D slice models) (Even a random forest here in the Z direction was better than nothing). The important thing is to use many slices (the entire Z direction) to predict the targets. (Even when predicting the target associated with 1 slice i.e. pe_present_on_image using more neighbor slices helps).
thank you for sharing this knowledge
", "votes": 1, "replies": [] }, { "id": 1062472, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-10-27T21:35:23.317000", "content": "I will also add that many of the top teams used a transformer (which is another kind of NN) in addition (or instead of) LSTM/GRU. Again, the bottom line is to use something that takes all slices as input to help predict the targets.
", "votes": 1, "replies": [] }, { "id": 1063398, "author_name": "ymgw55", "author_url": "", "post_date": "2020-10-28T21:02:36.140000", "content": "thank you for sharing your knowledge. I learnt a lot from the idea using more neighbor slices to predict pe_present_on_image. I found that LSTM and GRU were also used in last year RSNA competion, so many of the top teams used them and some teams used Transformer too.
I would also like to point out that 8th place solution didn't use LSTM nor GRU for stage 2. They used LGBM and achieved similar results. Great job team Jan @jpbremer , Yuji-san @yujiariyasu , Yama-san @lisosia
\nI think it is wonderful that we see a variety of model design in the top solutions.
", "votes": 3, "replies": [ { "id": 1063423, "author_name": "Johnny Lee", "author_url": "", "post_date": "2020-10-28T22:27:12.867000", "content": "We used CNN. It also worked, but not so good.
", "votes": 1, "replies": [] }, { "id": 1063436, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-10-28T23:02:37.347000", "content": "Nice Jonny. I also ensembled a GRU and 1D-CNN for my stage 2 model. It did better than just GRU.
", "votes": 0, "replies": [] } ] }, { "id": 1062152, "author_name": "Yohan Lee", "author_url": "", "post_date": "2020-10-27T15:55:38.920000", "content": "RNN modules like LSTM, GRU and Transformers are used for sequential data such as time series or text data. In this competition, we predict exam-level labels with many images. So, images or features of images can be considered as sequence indexed in z-pos order. I think that's the reason.
", "votes": 3, "replies": [ { "id": 1063400, "author_name": "ymgw55", "author_url": "", "post_date": "2020-10-28T21:04:49.807000", "content": "thank you for sharing your idea. I agree with you.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1062182": "In addition or alternatively to LSTM and GRU, you could also employ a 1D-CNN in the Z direction (on the GAP embeddings from your stage 1 2D slice models) (Even a random forest [here][1] in the Z direction was better than nothing). The important thing is to use many slices (the entire Z direction) to predict the targets. (Even when predicting the target associated with 1 slice i.e. `pe_present_on_image` using more neighbor slices helps).\n\n[1]: https://www.kaggle.com/c/rsna-str-pulmonary-embolism-detection/discussion/193402", "1063420": "I would also like to point out that 8th place solution didn't use LSTM nor GRU for stage 2. They used LGBM and achieved similar results. Great job team Jan @jpbremer , Yuji-san @yujiariyasu , Yama-san @lisosia \n\nI think it is wonderful that we see a variety of model design in the top solutions.", "1062152": "RNN modules like LSTM, GRU and Transformers are used for sequential data such as time series or text data. In this competition, we predict exam-level labels with many images. So, images or features of images can be considered as sequence indexed in z-pos order. I think that's the reason.", "1061983": "hey guys, I'm one of the kaggle begginers, and I have a question. Why is LSTM or GRU effective for this competition? Is this because datasets are 3D images? Someone already may have talked about this question, but if you guys have some opinions, please teach me for education." } }