Hello, this was my first Kaggle competition, and I have to admit I underestimated it…
\nI wanted to test the RNNs and this challenge looked a good fit to apply them, but my final score was really bad. I used TensorFlow to construct a couple of models inspired by the CarNet and SOPHIE net from Sadeghian, I was not able to fully train them, but their score was much worst than the CNN approaches (~190 vs ~20).
\nI was wondering if anyone has successfully applied RNNs to this task, and if so, how did you do it.
My score is not the best but I have shared my code and results, I like to think that ML grows from open code =)
\nMy humble effort repo
Hello, this was my first Kaggle competition, and I have to admit I underestimated it…
\nI wanted to test the RNNs and this challenge looked a good fit to apply them, but my final score was really bad. I used TensorFlow to construct a couple of models inspired by the CarNet and SOPHIE net from Sadeghian, I was not able to fully train them, but their score was much worst than the CNN approaches (~190 vs ~20).
\nI was wondering if anyone has successfully applied RNNs to this task, and if so, how did you do it.
My score is not the best but I have shared my code and results, I like to think that ML grows from open code =)
\nMy humble effort repo
One of my experiments is resnet3d + GRU. Resnet3d output frame-embedding which is input and hidden state of GRU. Then for-loop GRU model to get 50 axes. This approach gets 17.x LB with training one night.
", "rawMarkdown": "One of my experiments is resnet3d + GRU. Resnet3d output frame-embedding which is input and hidden state of GRU. Then for-loop GRU model to get 50 axes. This approach gets 17.x LB with training one night.", "votes": null }, { "id": "1092842", "postDate": "11/27/2020 08:01:03", "content": "I tried the similar approach and received worse result comparing to predicting 50 points directly from CNN.
", "rawMarkdown": "I tried the similar approach and received worse result comparing to predicting 50 points directly from CNN.", "votes": null }, { "id": "1093927", "postDate": "11/28/2020 05:59:53", "content": "we tried both, GRU on the input frames as well as on the output. But neither helped to improve the score
", "rawMarkdown": "we tried both, GRU on the input frames as well as on the output. But neither helped to improve the score", "votes": null } ], "comments": [ { "id": 1092171, "author_name": "axotcc", "author_url": "", "post_date": "11/26/2020 15:23:40", "content": "One of my experiments is resnet3d + GRU. Resnet3d output frame-embedding which is input and hidden state of GRU. Then for-loop GRU model to get 50 axes. This approach gets 17.x LB with training one night.
", "votes": null, "replies": [ { "id": 1092842, "author_name": "dmytropoplavskiy", "author_url": "", "post_date": "11/27/2020 08:01:03", "content": "I tried the similar approach and received worse result comparing to predicting 50 points directly from CNN.
", "votes": null, "replies": [] } ] }, { "id": 1093927, "author_name": "christofhenkel", "author_url": "", "post_date": "11/28/2020 05:59:53", "content": "we tried both, GRU on the input frames as well as on the output. But neither helped to improve the score
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1092070": "Hello, this was my first Kaggle competition, and I have to admit I underestimated it...\nI wanted to test the RNNs and this challenge looked a good fit to apply them, but my final score was really bad. I used TensorFlow to construct a couple of models inspired by the CarNet and SOPHIE net from Sadeghian, I was not able to fully train them, but their score was much worst than the CNN approaches (~190 vs ~20).\nI was wondering if anyone has successfully applied RNNs to this task, and if so, how did you do it.\n\nMy score is not the best but I have shared my code and results, I like to think that ML grows from open code =)\n[My humble effort repo](https://github.com/RawthiL/Lyft_Agent-Motion-Prediction)", "1092171": "One of my experiments is resnet3d + GRU. Resnet3d output frame-embedding which is input and hidden state of GRU. Then for-loop GRU model to get 50 axes. This approach gets 17.x LB with training one night.", "1092842": "I tried the similar approach and received worse result comparing to predicting 50 points directly from CNN.", "1093927": "we tried both, GRU on the input frames as well as on the output. But neither helped to improve the score" }, "source": "meta" }