The LIVECell repository maintained by the organizers has models trained with detectron2-ResNeSt. I tried running it, but I'm getting Not compiled with GPU support even though I have pytorch with CUDA and the regular detectron2 works fine.
It looks like the detectron2-ResNeSt repo is a fork of an old version (0.1.1 vs current 0.5) and hasn't been changed in over a year, so I'm wondering if it's just broken, or am I doing something wrong?
", "messages": [ { "id": 1564553, "postDate": "2021-10-29T10:48:53.480Z", "content": "The LIVECell repository maintained by the organizers has models trained with detectron2-ResNeSt. I tried running it, but I'm getting Not compiled with GPU support even though I have pytorch with CUDA and the regular detectron2 works fine.
It looks like the detectron2-ResNeSt repo is a fork of an old version (0.1.1 vs current 0.5) and hasn't been changed in over a year, so I'm wondering if it's just broken, or am I doing something wrong?
", "rawMarkdown": "The [LIVECell](https://github.com/sartorius-research/LIVECell) repository maintained by the organizers has models trained with [detectron2-ResNeSt](https://github.com/chongruo/detectron2-ResNeSt). I tried running it, but I'm getting `Not compiled with GPU support` even though I have pytorch with CUDA and the regular detectron2 works fine.\n\nIt looks like the detectron2-ResNeSt repo is a fork of an old version (0.1.1 vs current 0.5) and hasn't been changed in over a year, so I'm wondering if it's just broken, or am I doing something wrong?", "votes": 12 }, { "id": 1569472, "postDate": "2021-11-03T13:20:30.047Z", "content": "I have tried detectron2-ResNeSt model….But it performs quite bad after I have trained it on the competition dataset…. (its quite easy to install the detectron2 0.1.1 since there was someone who uploaded the older version of detectron2's GitHub repo on Kaggle's dateset. If you need, I could send you my pip install code for how to install it.)
", "rawMarkdown": "I have tried detectron2-ResNeSt model....But it performs quite bad after I have trained it on the competition dataset.... (its quite easy to install the detectron2 0.1.1 since there was someone who uploaded the older version of detectron2's GitHub repo on Kaggle's dateset. If you need, I could send you my pip install code for how to install it.)", "votes": 1, "replies": [ { "id": 1569570, "postDate": "2021-11-03T14:55:47.360Z", "content": "Thanks for the kind offer @jerrykun, I got it to work eventually.
\nThe problem turned out to be that I needed to manually install CUDA toolkit on my machine for it to build. The regular detectron2 didn't require that and worked with just the CUDA libraries I had installed from conda.
", "rawMarkdown": "Thanks for the kind offer @jerrykun, I got it to work eventually. \n\nThe problem turned out to be that I needed to manually install CUDA toolkit on my machine for it to build. The regular detectron2 didn't require that and worked with just the CUDA libraries I had installed from conda.", "votes": 1 }, { "id": 1569778, "postDate": "2021-11-03T17:54:56.187Z", "content": "did you get good result using it?
", "rawMarkdown": "did you get good result using it?" }, { "id": 1570385, "postDate": "2021-11-04T05:17:48.677Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 1567004, "postDate": "2021-11-01T12:36:15.087Z", "content": "Hello,
\nDo you have any update on this?
\nI have tried the CenterMask2 model (without retraining it) and the loss is very bad.
Hi. I check with the livecell-anchor_based SH-SY5Y model but the loss does not satisfying me(((
", "rawMarkdown": "Hi. I check with the livecell-anchor_based SH-SY5Y model but the loss does not satisfying me(((", "votes": 1 }, { "id": 1565452, "postDate": "2021-10-30T14:12:39.370Z", "content": "@slawekbiel detectron2 repo is not updated with many things. I was trying to use new baselines but finding it difficult to implement with LazyConfig. There is no proper documentation.
", "rawMarkdown": "@slawekbiel detectron2 repo is not updated with many things. I was trying to use [new baselines](https://github.com/facebookresearch/detectron2/tree/main/configs/new_baselines) but finding it difficult to implement with LazyConfig. There is no proper documentation.", "votes": 1, "replies": [ { "id": 1565471, "postDate": "2021-10-30T14:40:51.640Z", "content": "Do you mean the original repo or the ResNeSt fork? In detectron2 I see code changes even from last week, while the fork hasn’t been touched since July 2020
", "rawMarkdown": "Do you mean the original repo or the ResNeSt fork? In detectron2 I see code changes even from last week, while the fork hasn’t been touched since July 2020" }, { "id": 1565480, "postDate": "2021-10-30T14:49:53.803Z", "content": "No I am saying about original repo of detectron2. I haven't tried ResNest. But will surely look into it.
", "rawMarkdown": "No I am saying about original repo of detectron2. I haven't tried ResNest. But will surely look into it.", "votes": 1 }, { "id": 1588880, "postDate": "2021-11-19T18:06:55.703Z", "content": "def main(args):\n cfg = LazyConfig.load(args.config_file)\n cfg = LazyConfig.apply_overrides(cfg, args.opts)\n\n default_setup(cfg, args)\n cfg.dataloader.train.dataset.names = \"sartorius_train\"\n cfg.dataloader.test.dataset.names = \"sartorius_val\"\n\n cfg.train.init_checkpoint = model_zoo.get_checkpoint_url(\"new_baselines/mask_rcnn_R_101_FPN_400ep_LSJ.py\")\n cfg.train.max_iter = 10000\n cfg.train.amp.enabled = False\n cfg.train.checkpointer.period = 500\n@sanchitvj i am using the LazyConfig like this and able to train , I have not yet done inference , but for the same settings , i didnt find any visible difference in result between two checkpoints - current vs newbaseline . Except for that the newbaseline converges probably a bit faster .. I use the LazyTrainer .. but frankly speaking , I find the changes in parameter names between old and new method quite irritating
", "rawMarkdown": "```\ndef main(args):\n cfg = LazyConfig.load(args.config_file)\n cfg = LazyConfig.apply_overrides(cfg, args.opts)\n \n default_setup(cfg, args)\n cfg.dataloader.train.dataset.names = \"sartorius_train\"\n cfg.dataloader.test.dataset.names = \"sartorius_val\"\n\n cfg.train.init_checkpoint = model_zoo.get_checkpoint_url(\"new_baselines/mask_rcnn_R_101_FPN_400ep_LSJ.py\")\n cfg.train.max_iter = 10000\n cfg.train.amp.enabled = False\n cfg.train.checkpointer.period = 500\n```\n@sanchitvj i am using the LazyConfig like this and able to train , I have not yet done inference , but for the same settings , i didnt find any visible difference in result between two checkpoints - current vs newbaseline . Except for that the newbaseline converges probably a bit faster .. I use the LazyTrainer .. but frankly speaking , I find the changes in parameter names between old and new method quite irritating ", "votes": 3 }, { "id": 1590877, "postDate": "2021-11-21T18:09:16.070Z", "content": "@phoenix9032 you are right, new_baselines are not helping
", "rawMarkdown": "@phoenix9032 you are right, new_baselines are not helping" }, { "id": 1617058, "postDate": "2021-12-13T20:02:02.043Z", "content": "@phoenix9032 do you have train script kernel somewhere? I was trying to get this to word with the SimpleTrainer but it had a lot of issues. I am wondering if you used that which is mentioned in the example or DefaultTrainer.
Same as Zhikun, maybe some parameters setting mistake due to me, the training loss is not very good
", "rawMarkdown": "Same as Zhikun, maybe some parameters setting mistake due to me, the training loss is not very good" } ], "comments": [ { "id": 1569472, "author_name": "Zhikun Xu", "author_url": "", "post_date": "2021-11-03T13:20:30.047000", "content": "I have tried detectron2-ResNeSt model….But it performs quite bad after I have trained it on the competition dataset…. (its quite easy to install the detectron2 0.1.1 since there was someone who uploaded the older version of detectron2's GitHub repo on Kaggle's dateset. If you need, I could send you my pip install code for how to install it.)
", "votes": 1, "replies": [ { "id": 1569570, "author_name": "Slawek Biel", "author_url": "", "post_date": "2021-11-03T14:55:47.360000", "content": "Thanks for the kind offer @jerrykun, I got it to work eventually.
\nThe problem turned out to be that I needed to manually install CUDA toolkit on my machine for it to build. The regular detectron2 didn't require that and worked with just the CUDA libraries I had installed from conda.
", "votes": 1, "replies": [] }, { "id": 1569778, "author_name": "dragon zhang", "author_url": "", "post_date": "2021-11-03T17:54:56.187000", "content": "did you get good result using it?
", "votes": 0, "replies": [] }, { "id": 1570385, "author_name": "", "author_url": "", "post_date": "2021-11-04T05:17:48.677000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 1567004, "author_name": "Konstantina", "author_url": "", "post_date": "2021-11-01T12:36:15.087000", "content": "Hello,
\nDo you have any update on this?
\nI have tried the CenterMask2 model (without retraining it) and the loss is very bad.
Hi. I check with the livecell-anchor_based SH-SY5Y model but the loss does not satisfying me(((
", "votes": 1, "replies": [] }, { "id": 1565452, "author_name": "Sanchit Vijay", "author_url": "", "post_date": "2021-10-30T14:12:39.370000", "content": "@slawekbiel detectron2 repo is not updated with many things. I was trying to use new baselines but finding it difficult to implement with LazyConfig. There is no proper documentation.
", "votes": 1, "replies": [ { "id": 1565471, "author_name": "Slawek Biel", "author_url": "", "post_date": "2021-10-30T14:40:51.640000", "content": "Do you mean the original repo or the ResNeSt fork? In detectron2 I see code changes even from last week, while the fork hasn’t been touched since July 2020
", "votes": 0, "replies": [] }, { "id": 1565480, "author_name": "Sanchit Vijay", "author_url": "", "post_date": "2021-10-30T14:49:53.803000", "content": "No I am saying about original repo of detectron2. I haven't tried ResNest. But will surely look into it.
", "votes": 1, "replies": [] }, { "id": 1588880, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2021-11-19T18:06:55.703000", "content": "def main(args):\n cfg = LazyConfig.load(args.config_file)\n cfg = LazyConfig.apply_overrides(cfg, args.opts)\n\n default_setup(cfg, args)\n cfg.dataloader.train.dataset.names = \"sartorius_train\"\n cfg.dataloader.test.dataset.names = \"sartorius_val\"\n\n cfg.train.init_checkpoint = model_zoo.get_checkpoint_url(\"new_baselines/mask_rcnn_R_101_FPN_400ep_LSJ.py\")\n cfg.train.max_iter = 10000\n cfg.train.amp.enabled = False\n cfg.train.checkpointer.period = 500\n@sanchitvj i am using the LazyConfig like this and able to train , I have not yet done inference , but for the same settings , i didnt find any visible difference in result between two checkpoints - current vs newbaseline . Except for that the newbaseline converges probably a bit faster .. I use the LazyTrainer .. but frankly speaking , I find the changes in parameter names between old and new method quite irritating
", "votes": 3, "replies": [] }, { "id": 1590877, "author_name": "Sanchit Vijay", "author_url": "", "post_date": "2021-11-21T18:09:16.070000", "content": "@phoenix9032 you are right, new_baselines are not helping
", "votes": 0, "replies": [] }, { "id": 1617058, "author_name": "Trushant Kalyanpur", "author_url": "", "post_date": "2021-12-13T20:02:02.043000", "content": "@phoenix9032 do you have train script kernel somewhere? I was trying to get this to word with the SimpleTrainer but it had a lot of issues. I am wondering if you used that which is mentioned in the example or DefaultTrainer.
Same as Zhikun, maybe some parameters setting mistake due to me, the training loss is not very good
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1564553": "The [LIVECell](https://github.com/sartorius-research/LIVECell) repository maintained by the organizers has models trained with [detectron2-ResNeSt](https://github.com/chongruo/detectron2-ResNeSt). I tried running it, but I'm getting `Not compiled with GPU support` even though I have pytorch with CUDA and the regular detectron2 works fine.\n\nIt looks like the detectron2-ResNeSt repo is a fork of an old version (0.1.1 vs current 0.5) and hasn't been changed in over a year, so I'm wondering if it's just broken, or am I doing something wrong?", "1569472": "I have tried detectron2-ResNeSt model....But it performs quite bad after I have trained it on the competition dataset.... (its quite easy to install the detectron2 0.1.1 since there was someone who uploaded the older version of detectron2's GitHub repo on Kaggle's dateset. If you need, I could send you my pip install code for how to install it.)", "1567004": "Hello,\nDo you have any update on this?\nI have tried the CenterMask2 model (without retraining it) and the loss is very bad. ", "1566419": "Hi. I check with the livecell-anchor_based SH-SY5Y model but the loss does not satisfying me(((", "1565452": "@slawekbiel detectron2 repo is not updated with many things. I was trying to use [new baselines](https://github.com/facebookresearch/detectron2/tree/main/configs/new_baselines) but finding it difficult to implement with LazyConfig. There is no proper documentation.", "1604024": "Same as Zhikun, maybe some parameters setting mistake due to me, the training loss is not very good" } }