![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5513638%2Fdc1279adf7cddb1b4ad05a32e108d175%2F149.png?generation=1606437813920943&alt=media\")
Thanks to Kaggle and Lyft for hosting the competition.
\nRule-base is only my team name, but not a solution.๐
In this competition, I tested two papers from waymo & google research:
\nvector net: https://arxiv.org/abs/2005.04259 (final solution)
\ntnt: https://arxiv.org/abs/2008.08294 (hard to train, = =!)
vector net:
\nbatch=1024, 1 * 1080 GPU + 8 core CPU, trained for 2 days
\n<=800 vectors
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5513638%2Fcb42ce425d35ec735107a2f2178e0985%2F12.png?generation=1606437866782123&alt=media\")
tnt:
\nbatch=128, 1 * 1080 GPU + 8 core CPU, trained for 0.5 day
\n<=800 vectors
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5513638%2Fde8650eb8c219883afa0c64f65c51911%2Ftnt1.png?generation=1606438391638966&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5513638%2F97dd2340242bbc2ff9a1a8492b882d2e%2Ftnt2.png?generation=1606438403511527&alt=media\")
Thanks to Kaggle and Lyft for hosting the competition.
\nRule-base is only my team name, but not a solution.๐
In this competition, I tested two papers from waymo & google research:
\nvector net: https://arxiv.org/abs/2005.04259 (final solution)
\ntnt: https://arxiv.org/abs/2008.08294 (hard to train, = =!)
vector net:
\nbatch=1024, 1 * 1080 GPU + 8 core CPU, trained for 2 days
\n<=800 vectors
tnt:
\nbatch=128, 1 * 1080 GPU + 8 core CPU, trained for 0.5 day
\n<=800 vectors
very good work. so you did not use image data at all?
\ndo you plan to release any code or kaggle kernel?
\ni would love to reimplement your solution. thanks
@bosskong I will also join the train of people asking for your code. Let us know if you plan to release this. I would like to test it out.
", "rawMarkdown": "@bosskong I will also join the train of people asking for your code. Let us know if you plan to release this. I would like to test it out.", "votes": 1 }, { "id": 1093165, "postDate": "2020-11-27T14:00:59.557Z", "content": "not use any image at all
", "rawMarkdown": "not use any image at all" }, { "id": 1204210, "postDate": "2021-02-15T23:45:41.463Z", "content": "Hello everyone! I wonder if someone else succeeded with implementation of this approach and if we can see the code what performs well
", "rawMarkdown": "Hello everyone! I wonder if someone else succeeded with implementation of this approach and if we can see the code what performs well" } ] }, { "id": 1221723, "postDate": "2021-03-01T08:02:55.267Z", "content": "I have created a YouTube video explaining the VectorNet method. Check it out if you are lazy to read the paper
\n๐ https://youtu.be/yJFtf-fz3WA
I was expecting a complete set of rules for driving here
", "rawMarkdown": "I was expecting a complete set of rules for driving here", "votes": 1 }, { "id": 1099697, "postDate": "2020-12-02T14:42:28.210Z", "content": "Wonderful approach, thank you for sharing it!!!
", "rawMarkdown": "Wonderful approach, thank you for sharing it!!! " }, { "id": 1098866, "postDate": "2020-12-01T23:04:36.123Z", "content": "Congrats @bosskong and thanks for sharing your solution.
", "rawMarkdown": "Congrats @bosskong and thanks for sharing your solution." }, { "id": 1096011, "postDate": "2020-11-30T06:35:41.243Z", "content": "Thanks for sharing interesting approach!
\nI guess the big advantage of not using image is its speed, no rasterization & no heavy CNN is necessary. How was the training speed of your approach? For example how many iter/sec speed is achieved with this approach?
about 3min per 1000 batch๏ผ batch_size = 1024
", "rawMarkdown": "about 3min per 1000 batch๏ผ batch_size = 1024", "votes": 1 }, { "id": 1096620, "postDate": "2020-11-30T16:42:04.110Z", "content": "oh my goodness, 1024*1000 ~ 1000000 (for calculation) in 3min and train_full dataset contains 220 million the time it takes 220 * 3 = 660 min which is almost 11 hours.
\nCan I ask a question please can you make any improvement in this from your current score. or it is the highest score you could achieve
I'm not sure. I met gradient exploding problem & gradient vanishing problem at the same time.
\nsome layers' gradient explodes easily but others' vanish ๐
I see, so this is indeed pretty faster than image training (My CNN need 40 GPU days to run train_full 1 epoch) thanks!
", "rawMarkdown": "I see, so this is indeed pretty faster than image training (My CNN need 40 GPU days to run train_full 1 epoch) thanks!", "votes": 1 } ] }, { "id": 1092607, "postDate": "2020-11-27T03:02:31.503Z", "content": "is this were your rasterized image you forwarded to the model or is this just a viisulization
", "rawMarkdown": "is this were your rasterized image you forwarded to the model or is this just a viisulization", "replies": [ { "id": 1092660, "postDate": "2020-11-27T04:32:40.507Z", "content": "it's my vis tool based on the matplotlib
", "rawMarkdown": "it's my vis tool based on the matplotlib" } ] }, { "id": 1092763, "postDate": "2020-11-27T06:32:04.353Z", "rawMarkdown": "", "votes": 3, "isDeleted": true, "replies": [ { "id": 1092938, "postDate": "2020-11-27T10:10:19.243Z", "content": "Yeah, it would be great to read source code
", "rawMarkdown": "Yeah, it would be great to read source code", "votes": 2 } ] } ], "comments": [ { "id": 1092772, "author_name": "hengck23", "author_url": "", "post_date": "2020-11-27T06:41:59.060000", "content": "very good work. so you did not use image data at all?
\ndo you plan to release any code or kaggle kernel?
\ni would love to reimplement your solution. thanks
@bosskong I will also join the train of people asking for your code. Let us know if you plan to release this. I would like to test it out.
", "votes": 1, "replies": [] }, { "id": 1093165, "author_name": "bosskong", "author_url": "", "post_date": "2020-11-27T14:00:59.557000", "content": "not use any image at all
", "votes": 0, "replies": [] }, { "id": 1204210, "author_name": "sdrnr", "author_url": "", "post_date": "2021-02-15T23:45:41.463000", "content": "Hello everyone! I wonder if someone else succeeded with implementation of this approach and if we can see the code what performs well
", "votes": 0, "replies": [] } ] }, { "id": 1221723, "author_name": "Artem.Sanakoev", "author_url": "", "post_date": "2021-03-01T08:02:55.267000", "content": "I have created a YouTube video explaining the VectorNet method. Check it out if you are lazy to read the paper
\n๐ https://youtu.be/yJFtf-fz3WA
I was expecting a complete set of rules for driving here
", "votes": 1, "replies": [] }, { "id": 1099697, "author_name": "Sergio Manuel Papadakis", "author_url": "", "post_date": "2020-12-02T14:42:28.210000", "content": "Wonderful approach, thank you for sharing it!!!
", "votes": 0, "replies": [] }, { "id": 1098866, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2020-12-01T23:04:36.123000", "content": "Congrats @bosskong and thanks for sharing your solution.
", "votes": 0, "replies": [] }, { "id": 1096011, "author_name": "corochann", "author_url": "", "post_date": "2020-11-30T06:35:41.243000", "content": "Thanks for sharing interesting approach!
\nI guess the big advantage of not using image is its speed, no rasterization & no heavy CNN is necessary. How was the training speed of your approach? For example how many iter/sec speed is achieved with this approach?
about 3min per 1000 batch๏ผ batch_size = 1024
", "votes": 1, "replies": [] }, { "id": 1096620, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2020-11-30T16:42:04.110000", "content": "oh my goodness, 1024*1000 ~ 1000000 (for calculation) in 3min and train_full dataset contains 220 million the time it takes 220 * 3 = 660 min which is almost 11 hours.
\nCan I ask a question please can you make any improvement in this from your current score. or it is the highest score you could achieve
I'm not sure. I met gradient exploding problem & gradient vanishing problem at the same time.
\nsome layers' gradient explodes easily but others' vanish ๐
I see, so this is indeed pretty faster than image training (My CNN need 40 GPU days to run train_full 1 epoch) thanks!
", "votes": 1, "replies": [] } ] }, { "id": 1092607, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2020-11-27T03:02:31.503000", "content": "is this were your rasterized image you forwarded to the model or is this just a viisulization
", "votes": 0, "replies": [ { "id": 1092660, "author_name": "bosskong", "author_url": "", "post_date": "2020-11-27T04:32:40.507000", "content": "it's my vis tool based on the matplotlib
", "votes": 0, "replies": [] } ] }, { "id": 1092763, "author_name": "", "author_url": "", "post_date": "2020-11-27T06:32:04.353000", "content": "", "votes": 3, "replies": [ { "id": 1092938, "author_name": "Vladislav Tretyak", "author_url": "", "post_date": "2020-11-27T10:10:19.243000", "content": "Yeah, it would be great to read source code
", "votes": 2, "replies": [] } ] } ], "raw_markdown_by_id": { "1092552": "Thanks to Kaggle and Lyft for hosting the competition.\n**Rule-base** is only my team name, but not a solution.๐\n\nIn this competition, I tested two papers from waymo & google research:\nvector net: https://arxiv.org/abs/2005.04259 (final solution)\ntnt: https://arxiv.org/abs/2008.08294 (hard to train, = =!)\n\nvector net:\nbatch=1024, 1 * 1080 GPU + 8 core CPU, trained for 2 days\n<=800 vectors\n\n\n\n\n\ntnt:\nbatch=128, 1 * 1080 GPU + 8 core CPU, trained for 0.5 day\n<=800 vectors\n\n\n\n", "1092772": "very good work. so you did not use image data at all?\n\ndo you plan to release any code or kaggle kernel?\ni would love to reimplement your solution. thanks", "1221723": "I have created a **YouTube video** explaining the VectorNet method. Check it out if you are lazy to read the paper \n๐ https://youtu.be/yJFtf-fz3WA", "1092864": "I was expecting a complete set of rules for driving here", "1099697": "Wonderful approach, thank you for sharing it!!! ", "1098866": "Congrats @bosskong and thanks for sharing your solution.", "1096011": "Thanks for sharing interesting approach! \nI guess the big advantage of not using image is its speed, no rasterization & no heavy CNN is necessary. How was the training speed of your approach? For example how many `iter/sec` speed is achieved with this approach?", "1092607": "is this were your rasterized image you forwarded to the model or is this just a viisulization", "1092763": "" } }