![\"\"](\"https://pbs.twimg.com/media/EeCC5phUMAEX2u_?format=png&name=small\")
From Mingxing Tan at https://twitter.com/tanmingxing/status/1288176253829144577?s=20:
\n\nEfficientDet-D7x: new SOTA 55.1 COCO AP at 153ms latency (10x faster than other multi-scale models). Also TF2+keras supported.
\n\nArxiv: https://arxiv.org/abs/1911.09070\ngithub: https://bit.ly/307N6GG \npaperwithcode: https://bit.ly/2Esz89U
\n\nFrom Mingxing Tan at https://twitter.com/tanmingxing/status/1288176253829144577?s=20:
\n\nEfficientDet-D7x: new SOTA 55.1 COCO AP at 153ms latency (10x faster than other multi-scale models). Also TF2+keras supported.
\n\nArxiv: https://arxiv.org/abs/1911.09070\ngithub: https://bit.ly/307N6GG \npaperwithcode: https://bit.ly/2Esz89U
\n\nI was really excited to try this a few days ago when I saw it on paperswithcode, but I can't fit it on my GPU, even with a batch size of 1. 😪
", "rawMarkdown": "I was really excited to try this a few days ago when I saw it on paperswithcode, but I can't fit it on my GPU, even with a batch size of 1. 😪", "votes": null }, { "id": "951208", "postDate": "07/30/2020 01:50:12", "content": "@jonykarki I've not tried it yet, and looking at the code and paper, I'm also a bit hesitant to train it due to the large amount of resources that it can possibly require 😂
", "rawMarkdown": "jonykarki I've not tried it yet, and looking at the code and paper, I'm also a bit hesitant to train it due to the large amount of resources that it can possibly require 😂", "votes": null }, { "id": "951252", "postDate": "07/30/2020 02:40:57", "content": "Exactly. Turns out I can't even train a D6 on my GPU even with Mixed Precision/apex. It would be really fun to give it a try though (if it worked)
", "rawMarkdown": "Exactly. Turns out I can't even train a D6 on my GPU even with Mixed Precision/apex. It would be really fun to give it a try though (if it worked)", "votes": null }, { "id": "952527", "postDate": "07/31/2020 02:44:28", "content": "@khoongweihao I don't know what are you talking about bro 🤥 ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F137053%2Fb51ff2d8868f3255f11605397620af55%2FScreen%20Shot%202020-07-31%20at%2005.42.54.png?generation=1596163428346433&alt=media\")
EDIT: wow at the down votes .. point of my comment is to joke around that I'm already training one (hence the pinocchio emoji 😏 humor doesn't fly well on kaggle I guess?
", "rawMarkdown": "khoongweihao I don't know what are you talking about bro 🤥 \n\nEDIT: wow at the down votes .. point of my comment is to joke around that I'm already training one (hence the pinocchio emoji 😏 humor doesn't fly well on kaggle I guess?", "votes": null }, { "id": "952747", "postDate": "07/31/2020 07:26:52", "content": "@ma7moud I don't think that's how you call d7x. I've not looked through all the code changes, but it seems for PyTorch it has already been updated in this commit: https://github.com/rwightman/efficientdet-pytorch/commit/f822b84e77b279cfc85f2151c97399a9f5b11304
", "rawMarkdown": "ma7moud I don't think that's how you call d7x. I've not looked through all the code changes, but it seems for PyTorch it has already been updated in this commit: https://github.com/rwightman/efficientdet-pytorch/commit/f822b84e77b279cfc85f2151c97399a9f5b11304", "votes": null }, { "id": "952849", "postDate": "07/31/2020 09:11:16", "content": "Same) only 8gb on my GPU
", "rawMarkdown": "Same) only 8gb on my GPU", "votes": null }, { "id": "952932", "postDate": "07/31/2020 10:33:03", "content": "Nice
", "rawMarkdown": "Nice", "votes": null }, { "id": "953601", "postDate": "07/31/2020 22:56:00", "content": "@khoongweihao dude that screenshot is from code training on it right now 😃 the whole point of toolboxes is that it enables you to tune the config easily and it would construct the right network and load the proper pretrained weights if you pass the flag ... the commit you mentioned has the same key for the dict I shared in my screenshot 😬
\n\npython\n tf_efficientdet_d7x=dict(\n name='tf_efficientdet_d7x',\n backbone_name='tf_efficientnet_b7',\n image_size=1536,\n fpn_channels=384,\n fpn_cell_repeats=8,\n box_class_repeats=5,\n anchor_scale=4.0,\n max_level=8,\n fpn_name='bifpn_sum', # Use unweighted sum for training stability.\n backbone_args=dict(drop_path_rate=0.2),\n url='https://github.com/rwightman/efficientdet-pytorch/releases/download/v0.1/tf_efficientdet_d7x-f390b87c.pth'\n ),\n
@khoongweihao which takes 25 mins per epoch with amp on pytorch 1.6 on a P100 with batch size of 2 for the GWD dataset 🤕
", "rawMarkdown": "khoongweihao which takes 25 mins per epoch with amp on pytorch 1.6 on a P100 with batch size of 2 for the GWD dataset 🤕", "votes": null }, { "id": "953616", "postDate": "07/31/2020 23:18:05", "content": "you can fit it on colab or kaggle with a P100 using batch size of 2 :)
", "rawMarkdown": "you can fit it on colab or kaggle with a P100 using batch size of 2 :)", "votes": null }, { "id": "953653", "postDate": "08/01/2020 00:41:41", "content": "Don't forget Google Colab. Free Tesla T4 (16gb) or P100 (16gb) to train on, batch size 2 works. You can just use the Kaggle API to clone the dataset.
", "rawMarkdown": "Don't forget Google Colab. Free Tesla T4 (16gb) or P100 (16gb) to train on, batch size 2 works. You can just use the Kaggle API to clone the dataset.", "votes": null } ], "comments": [ { "id": 951199, "author_name": "jonykarki", "author_url": "", "post_date": "07/30/2020 01:31:11", "content": "I was really excited to try this a few days ago when I saw it on paperswithcode, but I can't fit it on my GPU, even with a batch size of 1. 😪
", "votes": null, "replies": [ { "id": 951208, "author_name": "khoongweihao", "author_url": "", "post_date": "07/30/2020 01:50:12", "content": "@jonykarki I've not tried it yet, and looking at the code and paper, I'm also a bit hesitant to train it due to the large amount of resources that it can possibly require 😂
", "votes": null, "replies": [] }, { "id": 951252, "author_name": "jonykarki", "author_url": "", "post_date": "07/30/2020 02:40:57", "content": "Exactly. Turns out I can't even train a D6 on my GPU even with Mixed Precision/apex. It would be really fun to give it a try though (if it worked)
", "votes": null, "replies": [] }, { "id": 952849, "author_name": "kupchanski", "author_url": "", "post_date": "07/31/2020 09:11:16", "content": "Same) only 8gb on my GPU
", "votes": null, "replies": [] }, { "id": 953616, "author_name": "ma7moud", "author_url": "", "post_date": "07/31/2020 23:18:05", "content": "you can fit it on colab or kaggle with a P100 using batch size of 2 :)
", "votes": null, "replies": [] }, { "id": 953653, "author_name": "stanleyjzheng", "author_url": "", "post_date": "08/01/2020 00:41:41", "content": "Don't forget Google Colab. Free Tesla T4 (16gb) or P100 (16gb) to train on, batch size 2 works. You can just use the Kaggle API to clone the dataset.
", "votes": null, "replies": [] } ] }, { "id": 952527, "author_name": "ma7moud", "author_url": "", "post_date": "07/31/2020 02:44:28", "content": "@khoongweihao I don't know what are you talking about bro 🤥
EDIT: wow at the down votes .. point of my comment is to joke around that I'm already training one (hence the pinocchio emoji 😏 humor doesn't fly well on kaggle I guess?
", "votes": null, "replies": [ { "id": 952747, "author_name": "khoongweihao", "author_url": "", "post_date": "07/31/2020 07:26:52", "content": "@ma7moud I don't think that's how you call d7x. I've not looked through all the code changes, but it seems for PyTorch it has already been updated in this commit: https://github.com/rwightman/efficientdet-pytorch/commit/f822b84e77b279cfc85f2151c97399a9f5b11304
", "votes": null, "replies": [] }, { "id": 953601, "author_name": "ma7moud", "author_url": "", "post_date": "07/31/2020 22:56:00", "content": "@khoongweihao dude that screenshot is from code training on it right now 😃 the whole point of toolboxes is that it enables you to tune the config easily and it would construct the right network and load the proper pretrained weights if you pass the flag ... the commit you mentioned has the same key for the dict I shared in my screenshot 😬
\n\npython\n tf_efficientdet_d7x=dict(\n name='tf_efficientdet_d7x',\n backbone_name='tf_efficientnet_b7',\n image_size=1536,\n fpn_channels=384,\n fpn_cell_repeats=8,\n box_class_repeats=5,\n anchor_scale=4.0,\n max_level=8,\n fpn_name='bifpn_sum', # Use unweighted sum for training stability.\n backbone_args=dict(drop_path_rate=0.2),\n url='https://github.com/rwightman/efficientdet-pytorch/releases/download/v0.1/tf_efficientdet_d7x-f390b87c.pth'\n ),\n
@khoongweihao which takes 25 mins per epoch with amp on pytorch 1.6 on a P100 with batch size of 2 for the GWD dataset 🤕
", "votes": null, "replies": [] } ] }, { "id": 952932, "author_name": "adarshsinghdikhit", "author_url": "", "post_date": "07/31/2020 10:33:03", "content": "Nice
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "951190": "From Mingxing Tan at https://twitter.com/tanmingxing/status/1288176253829144577?s=20:\n\nEfficientDet-D7x: new SOTA 55.1 COCO AP at 153ms latency (10x faster than other multi-scale models). Also TF2+keras supported.\n\nArxiv: https://arxiv.org/abs/1911.09070\ngithub: https://bit.ly/307N6GG \npaperwithcode: https://bit.ly/2Esz89U\n\n", "951199": "I was really excited to try this a few days ago when I saw it on paperswithcode, but I can't fit it on my GPU, even with a batch size of 1. 😪", "951208": "jonykarki I've not tried it yet, and looking at the code and paper, I'm also a bit hesitant to train it due to the large amount of resources that it can possibly require 😂", "951252": "Exactly. Turns out I can't even train a D6 on my GPU even with Mixed Precision/apex. It would be really fun to give it a try though (if it worked)", "952527": "khoongweihao I don't know what are you talking about bro 🤥 \n\nEDIT: wow at the down votes .. point of my comment is to joke around that I'm already training one (hence the pinocchio emoji 😏 humor doesn't fly well on kaggle I guess?", "952747": "ma7moud I don't think that's how you call d7x. I've not looked through all the code changes, but it seems for PyTorch it has already been updated in this commit: https://github.com/rwightman/efficientdet-pytorch/commit/f822b84e77b279cfc85f2151c97399a9f5b11304", "952849": "Same) only 8gb on my GPU", "952932": "Nice", "953601": "khoongweihao dude that screenshot is from code training on it right now 😃 the whole point of toolboxes is that it enables you to tune the config easily and it would construct the right network and load the proper pretrained weights if you pass the flag ... the commit you mentioned has the same key for the dict I shared in my screenshot 😬 \n\n```python\n tf_efficientdet_d7x=dict(\n name='tf_efficientdet_d7x',\n backbone_name='tf_efficientnet_b7',\n image_size=1536,\n fpn_channels=384,\n fpn_cell_repeats=8,\n box_class_repeats=5,\n anchor_scale=4.0,\n max_level=8,\n fpn_name='bifpn_sum', # Use unweighted sum for training stability.\n backbone_args=dict(drop_path_rate=0.2),\n url='https://github.com/rwightman/efficientdet-pytorch/releases/download/v0.1/tf_efficientdet_d7x-f390b87c.pth'\n ),\n```", "953602": "khoongweihao which takes 25 mins per epoch with amp on pytorch 1.6 on a P100 with batch size of 2 for the GWD dataset 🤕", "953616": "you can fit it on colab or kaggle with a P100 using batch size of 2 :)", "953653": "Don't forget Google Colab. Free Tesla T4 (16gb) or P100 (16gb) to train on, batch size 2 works. You can just use the Kaggle API to clone the dataset." }, "source": "meta" }