Hi everyone, I am planning to join this competition soon, but I basically only see PyTorch kernels, I am pretty new to object detection, so is there any problem using Tensorflow here? does it have the SOTA algorithms like EfficientDet, YOLOv4, or V5? I could not find it around, I would be grateful for any directions, thanks!
", "messages": [ { "id": 886307, "postDate": "2020-06-14T22:46:15.387Z", "content": "Hi everyone, I am planning to join this competition soon, but I basically only see PyTorch kernels, I am pretty new to object detection, so is there any problem using Tensorflow here? does it have the SOTA algorithms like EfficientDet, YOLOv4, or V5? I could not find it around, I would be grateful for any directions, thanks!
", "rawMarkdown": "Hi everyone, I am planning to join this competition soon, but I basically only see PyTorch kernels, I am pretty new to object detection, so is there any problem using Tensorflow here? does it have the SOTA algorithms like EfficientDet, YOLOv4, or V5? I could not find it around, I would be grateful for any directions, thanks!", "votes": 9 }, { "id": 909955, "postDate": "2020-07-01T00:26:37.403Z", "content": "This seems to be good YoloV3 and YoloV4 TF/Keras repo \nhttps://github.com/david8862/keras-YOLOv3-model-set\n(Have no time to try yet :D )
", "rawMarkdown": "\nThis seems to be good YoloV3 and YoloV4 TF/Keras repo \nhttps://github.com/david8862/keras-YOLOv3-model-set\n(Have no time to try yet :D ) ", "votes": 3, "replies": [ { "id": 910059, "postDate": "2020-07-01T01:39:49.927Z", "content": "This seems very solid, thanks @ratthachat !
", "rawMarkdown": "This seems very solid, thanks @ratthachat !" } ] }, { "id": 896399, "postDate": "2020-06-22T06:34:56.817Z", "content": "Welcome @dimitreoliveira !
\n\nIf you want TF/Keras EfficientDet, please try this repo : https://github.com/xuannianz/EfficientDet\nI think we can try RetinaNet just in case to study Keras pipeline , but not about performance.\nRetinaNet was good in the past, but now its best performance is very far behind EfficientDet.
\n\nIt would be interesting to translate @shonenkov kernel to TF/Keras and get similar performance.\nI am still studying and deciding whether to use TF or Pytorch in this comp. :D \nLooking forward to see great kernels from you!
", "rawMarkdown": "Welcome @dimitreoliveira !\n\nIf you want TF/Keras EfficientDet, please try this repo : https://github.com/xuannianz/EfficientDet\nI think we can try RetinaNet just in case to study Keras pipeline , but not about performance.\nRetinaNet was good in the past, but now its best performance is very far behind EfficientDet. \n\nIt would be interesting to translate @shonenkov kernel to TF/Keras and get similar performance.\nI am still studying and deciding whether to use TF or Pytorch in this comp. :D \nLooking forward to see great kernels from you!", "votes": 3, "replies": [ { "id": 896765, "postDate": "2020-06-22T12:24:03.630Z", "content": "Hey @ratthachat ,
\n\nThanks this repo seems like a good starting point on EfficientDet, and thanks for the RetinaNet heads-up, I was watching a video from DeepMind, and was mentioned that it was SOTA, maybe was an old video 😄.\nTranslating is a good exercise. I am waiting for the Jigsaw competition to end, I will take a few days break then I will decide if I'm joining this or the melanoma, but already there is a lot to start with!
", "rawMarkdown": "Hey @ratthachat ,\n\nThanks this repo seems like a good starting point on EfficientDet, and thanks for the RetinaNet heads-up, I was watching a video from DeepMind, and was mentioned that it was SOTA, maybe was an old video 😄.\nTranslating is a good exercise. I am waiting for the Jigsaw competition to end, I will take a few days break then I will decide if I'm joining this or the melanoma, but already there is a lot to start with!", "votes": 1 }, { "id": 896777, "postDate": "2020-06-22T12:29:28.920Z", "content": "Okay Dimitre, @dimitreoliveira \nBTW about SOTA, you can easily check this : https://paperswithcode.com/sota/object-detection-on-coco :)
", "rawMarkdown": "Okay Dimitre, @dimitreoliveira \nBTW about SOTA, you can easily check this : https://paperswithcode.com/sota/object-detection-on-coco :) ", "votes": 2 }, { "id": 896796, "postDate": "2020-06-22T12:41:34.223Z", "content": "Thank @ratthachat , that graph is very good, indeed RetinaNet is waaay behind.
", "rawMarkdown": "Thank @ratthachat , that graph is very good, indeed RetinaNet is waaay behind.", "votes": 1 }, { "id": 896848, "postDate": "2020-06-22T13:27:08.340Z", "content": "I think it's comparing apples and oranges though. The ranking of the performance on COCO will not hold true for the GWHD due to their differences (masks vs boxes, multi-class vs single class, lots of data vs little data).
\n\nAnd even just from personal experience, FRCNN has worked best on this dataset even though it's ranked nearly last.
", "rawMarkdown": "I think it's comparing apples and oranges though. The ranking of the performance on COCO will not hold true for the GWHD due to their differences (masks vs boxes, multi-class vs single class, lots of data vs little data).\n\nAnd even just from personal experience, FRCNN has worked best on this dataset even though it's ranked nearly last.", "votes": 2 }, { "id": 899040, "postDate": "2020-06-23T23:14:59.743Z", "content": "That is a great point @matthewmasters , are you using Tensorflow?
", "rawMarkdown": "That is a great point @matthewmasters , are you using Tensorflow?" }, { "id": 899043, "postDate": "2020-06-23T23:17:56Z", "content": "Nope, PyTorch. I've experimented with a couple other frameworks but don't think I will use them
", "rawMarkdown": "Nope, PyTorch. I've experimented with a couple other frameworks but don't think I will use them", "votes": 1 } ] }, { "id": 892268, "postDate": "2020-06-18T18:57:16.900Z", "content": "@dimitreoliveira hi,\nJust sharing info on this. \nCurrently, there are less and comparatively low scoring TF or Keras based kernels here. Namely,
\n\nThank you @ipythonx , Somehow I missed those kernels, I've heard that EfficientDet and RetinaNet are very good performers, I will give a look around, in any case I would not like to be using unofficial repositories, may find an issue after putting too much time into it.
", "rawMarkdown": "Thank you @ipythonx , Somehow I missed those kernels, I've heard that EfficientDet and RetinaNet are very good performers, I will give a look around, in any case I would not like to be using unofficial repositories, may find an issue after putting too much time into it.", "votes": 1 }, { "id": 893429, "postDate": "2020-06-19T15:44:30.663Z", "content": "@dimitreoliveira \nIf you've missed somehow, check this thread https://www.kaggle.com/c/global-wheat-detection/discussion/150780
\n\nFYI, EfficientDet and RetinaNet are one of the one-stage detectors, which mainly focus on the fast response (FPS) but for a two-stage detector, like faster RCNN is great for performance score, normally. However, in this competition so far the one-stage detectors are reported great results. I would recommend you look at EfficientDet and YoLO models.
", "rawMarkdown": "@dimitreoliveira \nIf you've missed somehow, check this thread https://www.kaggle.com/c/global-wheat-detection/discussion/150780\n\nFYI, EfficientDet and RetinaNet are one of the one-stage detectors, which mainly focus on the fast response (FPS) but for a two-stage detector, like faster RCNN is great for performance score, normally. However, in this competition so far the one-stage detectors are reported great results. I would recommend you look at EfficientDet and YoLO models. ", "votes": 1 }, { "id": 893760, "postDate": "2020-06-19T21:35:29.047Z", "content": "Thank you @ipythonx , I had missed that one too, thanks for the summary!
", "rawMarkdown": "Thank you @ipythonx , I had missed that one too, thanks for the summary!", "votes": 1 } ] }, { "id": 909330, "postDate": "2020-06-30T14:11:17.917Z", "content": "Highly recommend PyTorch. No brainer.
", "rawMarkdown": "Highly recommend PyTorch. No brainer.", "votes": -2 }, { "id": 886826, "postDate": "2020-06-15T09:52:38.860Z", "content": "yes,could anyone share how to use tensorflow's official EfficientDet？Thanks!!!
", "rawMarkdown": "yes,could anyone share how to use tensorflow's official EfficientDet？Thanks!!!", "replies": [ { "id": 892422, "postDate": "2020-06-18T21:29:59.533Z", "content": "Why don't you start to do this by yourself and share your findings? Like this, I think you won't get any helpful feedback.
", "rawMarkdown": "Why don't you start to do this by yourself and share your findings? Like this, I think you won't get any helpful feedback. ", "votes": 2 }, { "id": 892787, "postDate": "2020-06-19T06:49:45.310Z", "content": "ok，I will try
", "rawMarkdown": "ok，I will try", "votes": 1 } ] }, { "id": 893413, "postDate": "2020-06-19T15:31:46.627Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 893423, "postDate": "2020-06-19T15:37:58.937Z", "content": "I've encountered this implementation of Eff-Det in TF 2. It's a great effort by the author. But for competition purposes, it may not well-suited. I forget the link where the author himself agrees that this is still WIP.
\n\nAnd about the efficientnet keras import, for some reason, it gave me import error in kaggle kernel (TF 2.1). I didn't bother what's behind the cause, maybe TF 2.2 required.
WOW! It was a GHOST 👀
", "rawMarkdown": "WOW! It was a GHOST 👀 ", "votes": 1 } ] } ], "comments": [ { "id": 909955, "author_name": "Neuron Engineer", "author_url": "", "post_date": "2020-07-01T00:26:37.403000", "content": "This seems to be good YoloV3 and YoloV4 TF/Keras repo \nhttps://github.com/david8862/keras-YOLOv3-model-set\n(Have no time to try yet :D )
", "votes": 3, "replies": [ { "id": 910059, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2020-07-01T01:39:49.927000", "content": "This seems very solid, thanks @ratthachat !
", "votes": 0, "replies": [] } ] }, { "id": 896399, "author_name": "Neuron Engineer", "author_url": "", "post_date": "2020-06-22T06:34:56.817000", "content": "Welcome @dimitreoliveira !
\n\nIf you want TF/Keras EfficientDet, please try this repo : https://github.com/xuannianz/EfficientDet\nI think we can try RetinaNet just in case to study Keras pipeline , but not about performance.\nRetinaNet was good in the past, but now its best performance is very far behind EfficientDet.
\n\nIt would be interesting to translate @shonenkov kernel to TF/Keras and get similar performance.\nI am still studying and deciding whether to use TF or Pytorch in this comp. :D \nLooking forward to see great kernels from you!
", "votes": 3, "replies": [ { "id": 896765, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2020-06-22T12:24:03.630000", "content": "Hey @ratthachat ,
\n\nThanks this repo seems like a good starting point on EfficientDet, and thanks for the RetinaNet heads-up, I was watching a video from DeepMind, and was mentioned that it was SOTA, maybe was an old video 😄.\nTranslating is a good exercise. I am waiting for the Jigsaw competition to end, I will take a few days break then I will decide if I'm joining this or the melanoma, but already there is a lot to start with!
", "votes": 1, "replies": [] }, { "id": 896777, "author_name": "Neuron Engineer", "author_url": "", "post_date": "2020-06-22T12:29:28.920000", "content": "Okay Dimitre, @dimitreoliveira \nBTW about SOTA, you can easily check this : https://paperswithcode.com/sota/object-detection-on-coco :)
", "votes": 2, "replies": [] }, { "id": 896796, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2020-06-22T12:41:34.223000", "content": "Thank @ratthachat , that graph is very good, indeed RetinaNet is waaay behind.
", "votes": 1, "replies": [] }, { "id": 896848, "author_name": "Matt", "author_url": "", "post_date": "2020-06-22T13:27:08.340000", "content": "I think it's comparing apples and oranges though. The ranking of the performance on COCO will not hold true for the GWHD due to their differences (masks vs boxes, multi-class vs single class, lots of data vs little data).
\n\nAnd even just from personal experience, FRCNN has worked best on this dataset even though it's ranked nearly last.
", "votes": 2, "replies": [] }, { "id": 899040, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2020-06-23T23:14:59.743000", "content": "That is a great point @matthewmasters , are you using Tensorflow?
", "votes": 0, "replies": [] }, { "id": 899043, "author_name": "Matt", "author_url": "", "post_date": "2020-06-23T23:17:56", "content": "Nope, PyTorch. I've experimented with a couple other frameworks but don't think I will use them
", "votes": 1, "replies": [] } ] }, { "id": 892268, "author_name": "Innat", "author_url": "", "post_date": "2020-06-18T18:57:16.900000", "content": "@dimitreoliveira hi,\nJust sharing info on this. \nCurrently, there are less and comparatively low scoring TF or Keras based kernels here. Namely,
\n\nThank you @ipythonx , Somehow I missed those kernels, I've heard that EfficientDet and RetinaNet are very good performers, I will give a look around, in any case I would not like to be using unofficial repositories, may find an issue after putting too much time into it.
", "votes": 1, "replies": [] }, { "id": 893429, "author_name": "Innat", "author_url": "", "post_date": "2020-06-19T15:44:30.663000", "content": "@dimitreoliveira \nIf you've missed somehow, check this thread https://www.kaggle.com/c/global-wheat-detection/discussion/150780
\n\nFYI, EfficientDet and RetinaNet are one of the one-stage detectors, which mainly focus on the fast response (FPS) but for a two-stage detector, like faster RCNN is great for performance score, normally. However, in this competition so far the one-stage detectors are reported great results. I would recommend you look at EfficientDet and YoLO models.
", "votes": 1, "replies": [] }, { "id": 893760, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2020-06-19T21:35:29.047000", "content": "Thank you @ipythonx , I had missed that one too, thanks for the summary!
", "votes": 1, "replies": [] } ] }, { "id": 909330, "author_name": "ZaoDianNao", "author_url": "", "post_date": "2020-06-30T14:11:17.917000", "content": "Highly recommend PyTorch. No brainer.
", "votes": -2, "replies": [] }, { "id": 886826, "author_name": "lz1997", "author_url": "", "post_date": "2020-06-15T09:52:38.860000", "content": "yes,could anyone share how to use tensorflow's official EfficientDet？Thanks!!!
", "votes": 0, "replies": [ { "id": 892422, "author_name": "Innat", "author_url": "", "post_date": "2020-06-18T21:29:59.533000", "content": "Why don't you start to do this by yourself and share your findings? Like this, I think you won't get any helpful feedback.
", "votes": 2, "replies": [] }, { "id": 892787, "author_name": "lz1997", "author_url": "", "post_date": "2020-06-19T06:49:45.310000", "content": "ok，I will try
", "votes": 1, "replies": [] } ] }, { "id": 893413, "author_name": "", "author_url": "", "post_date": "2020-06-19T15:31:46.627000", "content": "", "votes": 0, "replies": [ { "id": 893423, "author_name": "Innat", "author_url": "", "post_date": "2020-06-19T15:37:58.937000", "content": "I've encountered this implementation of Eff-Det in TF 2. It's a great effort by the author. But for competition purposes, it may not well-suited. I forget the link where the author himself agrees that this is still WIP.
\n\nAnd about the efficientnet keras import, for some reason, it gave me import error in kaggle kernel (TF 2.1). I didn't bother what's behind the cause, maybe TF 2.2 required.
WOW! It was a GHOST 👀
", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "886307": "Hi everyone, I am planning to join this competition soon, but I basically only see PyTorch kernels, I am pretty new to object detection, so is there any problem using Tensorflow here? does it have the SOTA algorithms like EfficientDet, YOLOv4, or V5? I could not find it around, I would be grateful for any directions, thanks!", "909955": "\nThis seems to be good YoloV3 and YoloV4 TF/Keras repo \nhttps://github.com/david8862/keras-YOLOv3-model-set\n(Have no time to try yet :D ) ", "896399": "Welcome @dimitreoliveira !\n\nIf you want TF/Keras EfficientDet, please try this repo : https://github.com/xuannianz/EfficientDet\nI think we can try RetinaNet just in case to study Keras pipeline , but not about performance.\nRetinaNet was good in the past, but now its best performance is very far behind EfficientDet. \n\nIt would be interesting to translate @shonenkov kernel to TF/Keras and get similar performance.\nI am still studying and deciding whether to use TF or Pytorch in this comp. :D \nLooking forward to see great kernels from you!", "892268": "@dimitreoliveira hi,\nJust sharing info on this. \nCurrently, there are less and comparatively low scoring TF or Keras based kernels here. Namely,\n\n- [GWD with Mask-RCNN and COCO Transfer Learning](https://www.kaggle.com/ipythonx/gwd-with-mask-rcnn-and-coco-transfer-learning)\n- [GWD Sarter EfficientDet Keras (train)](https://www.kaggle.com/ateplyuk/gwd-starter-efficientdet-keras-train)\n- [Global Wheat Detection🌾- Keras RetinaNet❤ + EDA🔥](https://www.kaggle.com/chittalpatel/global-wheat-detection-keras-retinanet-eda/notebook)\n- [keras centernet inference](https://www.kaggle.com/nvnnghia/keras-centernet-inference/notebook)", "909330": "Highly recommend PyTorch. No brainer.", "886826": "yes,could anyone share how to use tensorflow's official EfficientDet？Thanks!!!", "893413": "" } }