People often compare tensorflow with pytorch in recent competitions, the most recent is the Google Brain competition (the init weight problem in the LSTM module). In this competition I also had a similar performance issue with Detectron2 and mmdetection's with the same model and the same (or nearly similar) hyperparameters or powerful models in mmdetection but not yet available in detectron2 they all have worse results than the models in detectron2 (example 0.27x and 0.28x with 0.29x and 0.30x). I know this comparison may not be accurate because my tests are not good enough or not enough (No Free Lunch Theorem), please let me know the results of your experiment.
", "messages": [ { "id": "1578397", "postDate": "11/11/2021 01:20:17", "content": "People often compare tensorflow with pytorch in recent competitions, the most recent is the Google Brain competition (the init weight problem in the LSTM module). In this competition I also had a similar performance issue with Detectron2 and mmdetection's with the same model and the same (or nearly similar) hyperparameters or powerful models in mmdetection but not yet available in detectron2 they all have worse results than the models in detectron2 (example 0.27x and 0.28x with 0.29x and 0.30x). I know this comparison may not be accurate because my tests are not good enough or not enough (No Free Lunch Theorem), please let me know the results of your experiment.
", "rawMarkdown": "People often compare tensorflow with pytorch in recent competitions, the most recent is the Google Brain competition (the init weight problem in the LSTM module). In this competition I also had a similar performance issue with Detectron2 and mmdetection's with the same model and the same (or nearly similar) hyperparameters or powerful models in mmdetection but not yet available in detectron2 they all have worse results than the models in detectron2 (example 0.27x and 0.28x with 0.29x and 0.30x). I know this comparison may not be accurate because my tests are not good enough or not enough (No Free Lunch Theorem), please let me know the results of your experiment.", "votes": null }, { "id": "1578417", "postDate": "11/11/2021 02:36:21", "content": "I use mmdetection only(0.323 LB)
\nI did not try detectron2
Good score, thank you for keeping me from giving up mmdetection now :D
", "rawMarkdown": "Good score, thank you for keeping me from giving up mmdetection now :D", "votes": null }, { "id": "1578468", "postDate": "11/11/2021 04:51:02", "content": "both packages reported results on COCO and can achieve similar results.
\nHence their implementation is correct.
you need to be familiar with the API to finetune the parameters for your use case.
", "rawMarkdown": "both packages reported results on COCO and can achieve similar results.\nHence their implementation is correct.\n\nyou need to be familiar with the API to finetune the parameters for your use case.", "votes": null }, { "id": "1578527", "postDate": "11/11/2021 06:30:36", "content": "Thanks Heng, I think I need to experiment more and review my settings.
", "rawMarkdown": "Thanks Heng, I think I need to experiment more and review my settings.", "votes": null }, { "id": "1579050", "postDate": "11/11/2021 13:51:16", "content": "Based on my limited few experiences, detectron2 seems a little bit better.
\nbut for this competition, postprocessing is quite important, or in other word, postprocessing optimization is important for a better PB.
From my experiment, mmdetection performed worse than detectron2, but as previous pointed, maybe we should dig deeper to compare their parameters in config files (I only use the same configs in training). The gap is about 0.2 - 0.3
", "rawMarkdown": "From my experiment, mmdetection performed worse than detectron2, but as previous pointed, maybe we should dig deeper to compare their parameters in config files (I only use the same configs in training). The gap is about 0.2 - 0.3", "votes": null }, { "id": "1579487", "postDate": "11/11/2021 23:25:43", "content": "Sooooo… judging by the comments here, no one from top teams are using Unets?
", "rawMarkdown": "Sooooo... judging by the comments here, no one from top teams are using Unets?", "votes": null }, { "id": "1580589", "postDate": "11/12/2021 23:11:57", "content": "All possible models will be used in a couple of weeks. ;)
", "rawMarkdown": "All possible models will be used in a couple of weeks. ;)", "votes": null } ], "comments": [ { "id": 1578417, "author_name": "tereka", "author_url": "", "post_date": "11/11/2021 02:36:21", "content": "I use mmdetection only(0.323 LB)
\nI did not try detectron2
Good score, thank you for keeping me from giving up mmdetection now :D
", "votes": null, "replies": [] } ] }, { "id": 1578468, "author_name": "hengck23", "author_url": "", "post_date": "11/11/2021 04:51:02", "content": "both packages reported results on COCO and can achieve similar results.
\nHence their implementation is correct.
you need to be familiar with the API to finetune the parameters for your use case.
", "votes": null, "replies": [ { "id": 1578527, "author_name": "duykhanh99", "author_url": "", "post_date": "11/11/2021 06:30:36", "content": "Thanks Heng, I think I need to experiment more and review my settings.
", "votes": null, "replies": [] } ] }, { "id": 1579050, "author_name": "dragonzhang", "author_url": "", "post_date": "11/11/2021 13:51:16", "content": "Based on my limited few experiences, detectron2 seems a little bit better.
\nbut for this competition, postprocessing is quite important, or in other word, postprocessing optimization is important for a better PB.
From my experiment, mmdetection performed worse than detectron2, but as previous pointed, maybe we should dig deeper to compare their parameters in config files (I only use the same configs in training). The gap is about 0.2 - 0.3
", "votes": null, "replies": [] }, { "id": 1579487, "author_name": "chankhavu", "author_url": "", "post_date": "11/11/2021 23:25:43", "content": "Sooooo… judging by the comments here, no one from top teams are using Unets?
", "votes": null, "replies": [ { "id": 1580589, "author_name": "tolgadincer", "author_url": "", "post_date": "11/12/2021 23:11:57", "content": "All possible models will be used in a couple of weeks. ;)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1578397": "People often compare tensorflow with pytorch in recent competitions, the most recent is the Google Brain competition (the init weight problem in the LSTM module). In this competition I also had a similar performance issue with Detectron2 and mmdetection's with the same model and the same (or nearly similar) hyperparameters or powerful models in mmdetection but not yet available in detectron2 they all have worse results than the models in detectron2 (example 0.27x and 0.28x with 0.29x and 0.30x). I know this comparison may not be accurate because my tests are not good enough or not enough (No Free Lunch Theorem), please let me know the results of your experiment.", "1578417": "I use mmdetection only(0.323 LB) \nI did not try detectron2", "1578424": "Good score, thank you for keeping me from giving up mmdetection now :D", "1578468": "both packages reported results on COCO and can achieve similar results.\nHence their implementation is correct.\n\nyou need to be familiar with the API to finetune the parameters for your use case.", "1578527": "Thanks Heng, I think I need to experiment more and review my settings.", "1579050": "Based on my limited few experiences, detectron2 seems a little bit better.\nbut for this competition, postprocessing is quite important, or in other word, postprocessing optimization is important for a better PB.", "1579364": "From my experiment, mmdetection performed worse than detectron2, but as previous pointed, maybe we should dig deeper to compare their parameters in config files (I only use the same configs in training). The gap is about 0.2 - 0.3", "1579487": "Sooooo... judging by the comments here, no one from top teams are using Unets?", "1580589": "All possible models will be used in a couple of weeks. ;)" }, "source": "meta" }