![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F448347%2Fbd23cde9a0327a9937a69a61dbabed61%2Fflops_latency.png?generation=1582786285917679&alt=media\")
paper: https://arxiv.org/abs/1911.11907
\n\ngithub: https://github.com/huawei-noah/ghostnet
\n\n
paper: https://arxiv.org/abs/1911.11907
\n\ngithub: https://github.com/huawei-noah/ghostnet
\n\n
Seems like you are stuck at the top ;)
", "rawMarkdown": "Seems like you are stuck at the top ;)", "votes": 4, "replies": [ { "id": 758774, "postDate": "2020-02-28T06:19:16.620Z", "content": "😂
", "rawMarkdown": "😂 " }, { "id": 760940, "postDate": "2020-03-01T23:47:28.077Z", "content": "Not the worst place to be stuck at at all))
", "rawMarkdown": "Not the worst place to be stuck at at all))", "votes": 1 } ] }, { "id": 764010, "postDate": "2020-03-05T04:28:35.347Z", "content": "Great !!!! thanks for sharing @haqishen ....
", "rawMarkdown": "Great !!!! thanks for sharing @haqishen ....", "votes": 1 }, { "id": 760136, "postDate": "2020-02-29T22:23:03.327Z", "content": "This architecture trains very fast with the width parameter set to the default=1.0 (in the paper, they report on == 0.5, 1.0, and 1.5). By very fast, I mean 1-2 min/epoch on a single 2080ti.
\n\nWith the exact same seeded setup just swapping out the models, I'm getting roughly ~0.02 less than EffNetB3. Please note with the EffNet model I was using imagenet weights, with the GhostNet model I was training from scratch because I was too lazy to port over shared tensorflow weights. I don't think this guy will win any prizes on the Single Model Best CV thread, but it's great to experiment other aspects on and might make for a good low-cost ensemble.
\n\nThanks for sharing!
", "rawMarkdown": "This architecture trains very fast with the width parameter set to the default=1.0 (in the paper, they report on == 0.5, 1.0, and 1.5). By very fast, I mean 1-2 min/epoch on a single 2080ti.\n\nWith the exact same seeded setup just swapping out the models, I'm getting roughly ~0.02 less than EffNetB3. Please note with the EffNet model I was using imagenet weights, with the GhostNet model I was training from scratch because I was too lazy to port over shared tensorflow weights. I don't think this guy will win any prizes on the Single Model Best CV thread, but it's great to experiment other aspects on and might make for a good low-cost ensemble.\n\nThanks for sharing!", "votes": 1, "replies": [ { "id": 760297, "postDate": "2020-03-01T04:54:36.940Z", "content": "Thanks for your report!\nSeems that this guy is a good one and once pytorch ImageNet weight is released, it may be popular in future competition.
", "rawMarkdown": "Thanks for your report!\nSeems that this guy is a good one and once pytorch ImageNet weight is released, it may be popular in future competition." } ] }, { "id": 759763, "postDate": "2020-02-29T12:36:57.367Z", "content": "It's from HUAWEI Noah. They are also doing quite well in NAS, model compression and also optimization.
", "rawMarkdown": "It's from HUAWEI Noah. They are also doing quite well in NAS, model compression and also optimization.", "votes": 1 }, { "id": 758936, "postDate": "2020-02-28T11:15:07.283Z", "content": "Thanks for sharing. The competition is about to end soon; have you tried ghostnet already?
", "rawMarkdown": "Thanks for sharing. The competition is about to end soon; have you tried ghostnet already? ", "votes": 1, "replies": [ { "id": 758950, "postDate": "2020-02-28T11:38:53.627Z", "content": "No... I'm not going to try it on this dataset.\nJust sharing a new thing with you guys ;)
", "rawMarkdown": "No... I'm not going to try it on this dataset.\nJust sharing a new thing with you guys ;)", "votes": 1 }, { "id": 758972, "postDate": "2020-02-28T12:17:52.683Z", "content": "Hmm, I got that. Thanks. \nOur team runs that .97 scored kernel but in the final submission we won't select it. However, in our experiments, we've tried much stuff but couldn't score a minimum of 97 or 98 until now while some for others it's so doable. On top of that some high scored kernel gives a long gap in LB.
", "rawMarkdown": "Hmm, I got that. Thanks. \nOur team runs that .97 scored kernel but in the final submission we won't select it. However, in our experiments, we've tried much stuff but couldn't score a minimum of 97 or 98 until now while some for others it's so doable. On top of that some high scored kernel gives a long gap in LB. " }, { "id": 758977, "postDate": "2020-02-28T12:24:57.427Z", "content": "if you find your code not working so well, why not try to use public training pipeline and then add tricks on it.
", "rawMarkdown": "if you find your code not working so well, why not try to use public training pipeline and then add tricks on it." }, { "id": 758986, "postDate": "2020-02-28T12:33:09.910Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 757909, "postDate": "2020-02-27T08:30:54.410Z", "content": "Thank you! Does it have a model pre-trained on ImageNet??
", "rawMarkdown": "Thank you! Does it have a model pre-trained on ImageNet??", "votes": 1, "replies": [ { "id": 757954, "postDate": "2020-02-27T09:27:41.207Z", "content": "yes for tensorflow, but no for pytorch, for now.
", "rawMarkdown": "yes for tensorflow, but no for pytorch, for now.", "votes": 1 } ] }, { "id": 933609, "postDate": "2020-07-17T20:21:29.113Z", "content": "PyTorch implementation of GhostNet and Official pretrained weights
\n\n", "rawMarkdown": "PyTorch implementation of GhostNet and Official pretrained weights \n\n- https://www.kaggle.com/ipythonx/ghostnetpretrained/settings" }, { "id": 761142, "postDate": "2020-03-02T07:24:19.180Z", "content": "Cool, Thanks for sharing.
", "rawMarkdown": "Cool, Thanks for sharing.", "votes": 3 }, { "id": 760078, "postDate": "2020-02-29T20:42:38.950Z", "content": "Thanks for sharing.
", "rawMarkdown": "Thanks for sharing.", "votes": 1 }, { "id": 758084, "postDate": "2020-02-27T12:25:29.757Z", "content": "Thanks for share 💯
", "rawMarkdown": "Thanks for share 💯 ", "votes": 1 } ], "comments": [ { "id": 757879, "author_name": "Manoj", "author_url": "", "post_date": "2020-02-27T07:55:03.507000", "content": "Seems like you are stuck at the top ;)
", "votes": 4, "replies": [ { "id": 758774, "author_name": "Santosh", "author_url": "", "post_date": "2020-02-28T06:19:16.620000", "content": "😂
", "votes": 0, "replies": [] }, { "id": 760940, "author_name": "Kostya Atarik", "author_url": "", "post_date": "2020-03-01T23:47:28.077000", "content": "Not the worst place to be stuck at at all))
", "votes": 1, "replies": [] } ] }, { "id": 764010, "author_name": "SAIF UDDIN", "author_url": "", "post_date": "2020-03-05T04:28:35.347000", "content": "Great !!!! thanks for sharing @haqishen ....
", "votes": 1, "replies": [] }, { "id": 760136, "author_name": "عثمان", "author_url": "", "post_date": "2020-02-29T22:23:03.327000", "content": "This architecture trains very fast with the width parameter set to the default=1.0 (in the paper, they report on == 0.5, 1.0, and 1.5). By very fast, I mean 1-2 min/epoch on a single 2080ti.
\n\nWith the exact same seeded setup just swapping out the models, I'm getting roughly ~0.02 less than EffNetB3. Please note with the EffNet model I was using imagenet weights, with the GhostNet model I was training from scratch because I was too lazy to port over shared tensorflow weights. I don't think this guy will win any prizes on the Single Model Best CV thread, but it's great to experiment other aspects on and might make for a good low-cost ensemble.
\n\nThanks for sharing!
", "votes": 1, "replies": [ { "id": 760297, "author_name": "Qishen Ha", "author_url": "", "post_date": "2020-03-01T04:54:36.940000", "content": "Thanks for your report!\nSeems that this guy is a good one and once pytorch ImageNet weight is released, it may be popular in future competition.
", "votes": 0, "replies": [] } ] }, { "id": 759763, "author_name": "syoya", "author_url": "", "post_date": "2020-02-29T12:36:57.367000", "content": "It's from HUAWEI Noah. They are also doing quite well in NAS, model compression and also optimization.
", "votes": 1, "replies": [] }, { "id": 758936, "author_name": "Innat", "author_url": "", "post_date": "2020-02-28T11:15:07.283000", "content": "Thanks for sharing. The competition is about to end soon; have you tried ghostnet already?
", "votes": 1, "replies": [ { "id": 758950, "author_name": "Qishen Ha", "author_url": "", "post_date": "2020-02-28T11:38:53.627000", "content": "No... I'm not going to try it on this dataset.\nJust sharing a new thing with you guys ;)
", "votes": 1, "replies": [] }, { "id": 758972, "author_name": "Innat", "author_url": "", "post_date": "2020-02-28T12:17:52.683000", "content": "Hmm, I got that. Thanks. \nOur team runs that .97 scored kernel but in the final submission we won't select it. However, in our experiments, we've tried much stuff but couldn't score a minimum of 97 or 98 until now while some for others it's so doable. On top of that some high scored kernel gives a long gap in LB.
", "votes": 0, "replies": [] }, { "id": 758977, "author_name": "Qishen Ha", "author_url": "", "post_date": "2020-02-28T12:24:57.427000", "content": "if you find your code not working so well, why not try to use public training pipeline and then add tricks on it.
", "votes": 0, "replies": [] }, { "id": 758986, "author_name": "", "author_url": "", "post_date": "2020-02-28T12:33:09.910000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 757909, "author_name": "Jaebb", "author_url": "", "post_date": "2020-02-27T08:30:54.410000", "content": "Thank you! Does it have a model pre-trained on ImageNet??
", "votes": 1, "replies": [ { "id": 757954, "author_name": "Qishen Ha", "author_url": "", "post_date": "2020-02-27T09:27:41.207000", "content": "yes for tensorflow, but no for pytorch, for now.
", "votes": 1, "replies": [] } ] }, { "id": 933609, "author_name": "Innat", "author_url": "", "post_date": "2020-07-17T20:21:29.113000", "content": "PyTorch implementation of GhostNet and Official pretrained weights
\n\n", "votes": 0, "replies": [] }, { "id": 761142, "author_name": "MachineLP", "author_url": "", "post_date": "2020-03-02T07:24:19.180000", "content": "Cool, Thanks for sharing.
", "votes": 3, "replies": [] }, { "id": 760078, "author_name": "Johar M. Ashfaque", "author_url": "", "post_date": "2020-02-29T20:42:38.950000", "content": "Thanks for sharing.
", "votes": 1, "replies": [] }, { "id": 758084, "author_name": "dasmehdixtr", "author_url": "", "post_date": "2020-02-27T12:25:29.757000", "content": "Thanks for share 💯
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "757841": "paper: https://arxiv.org/abs/1911.11907\n\ngithub: https://github.com/huawei-noah/ghostnet\n\n\n", "757879": "Seems like you are stuck at the top ;)", "764010": "Great !!!! thanks for sharing @haqishen ....", "760136": "This architecture trains very fast with the width parameter set to the default=1.0 (in the paper, they report on == 0.5, 1.0, and 1.5). By very fast, I mean 1-2 min/epoch on a single 2080ti.\n\nWith the exact same seeded setup just swapping out the models, I'm getting roughly ~0.02 less than EffNetB3. Please note with the EffNet model I was using imagenet weights, with the GhostNet model I was training from scratch because I was too lazy to port over shared tensorflow weights. I don't think this guy will win any prizes on the Single Model Best CV thread, but it's great to experiment other aspects on and might make for a good low-cost ensemble.\n\nThanks for sharing!", "759763": "It's from HUAWEI Noah. They are also doing quite well in NAS, model compression and also optimization.", "758936": "Thanks for sharing. The competition is about to end soon; have you tried ghostnet already? ", "757909": "Thank you! Does it have a model pre-trained on ImageNet??", "933609": "PyTorch implementation of GhostNet and Official pretrained weights \n\n- https://www.kaggle.com/ipythonx/ghostnetpretrained/settings", "761142": "Cool, Thanks for sharing.", "760078": "Thanks for sharing.", "758084": "Thanks for share 💯 " } }