![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2241982%2Fbfca8f11daf9d4d8e26d76bb60209720%2Fcenternet.png?generation=1576241857423986&alt=media\")
In the paper of CenterNet, deeper networks seem to perform better for object detection, but in this competition, my experiment was the best at efficientnet-b0, which was better than b3 and b5. My CV was 0.127(use https://www.kaggle.com/its7171/metrics-evaluation-script and 0.2 for val), and lb was 0.56. Do you get better performance using deeper networks ?
In the paper of CenterNet, deeper networks seem to perform better for object detection, but in this competition, my experiment was the best at efficientnet-b0, which was better than b3 and b5. My CV was 0.127(use https://www.kaggle.com/its7171/metrics-evaluation-script and 0.2 for val), and lb was 0.56. Do you get better performance using deeper networks ?
Yes and No. For me, resnet101 > resnet50, but seresnext50 or 101 doesn't work good...
", "rawMarkdown": "Yes and No. For me, resnet101 > resnet50, but seresnext50 or 101 doesn't work good...", "votes": 5, "replies": [ { "id": 697168, "postDate": "2019-12-17T15:03:15.103Z", "content": "I can confirm the same . I am not sure why though :(
", "rawMarkdown": "I can confirm the same . I am not sure why though :( " }, { "id": 697527, "postDate": "2019-12-18T03:29:22.383Z", "content": "Does using pre-trained weights make a difference?
", "rawMarkdown": "Does using pre-trained weights make a difference?", "votes": 1 }, { "id": 698056, "postDate": "2019-12-18T18:38:36.257Z", "content": "Sounds like you are doing it from scratch?
", "rawMarkdown": "Sounds like you are doing it from scratch?" }, { "id": 698171, "postDate": "2019-12-18T22:09:58.500Z", "content": "No, I'm using pretrained weight.
", "rawMarkdown": "No, I'm using pretrained weight.", "votes": 1 }, { "id": 700414, "postDate": "2019-12-22T01:23:59.347Z", "content": "For me, resnext 101 < resnext 50 from scratch
", "rawMarkdown": "For me, resnext 101 < resnext 50 from scratch" } ] }, { "id": 697283, "postDate": "2019-12-17T17:18:52.463Z", "content": "@hesene what image resolution did u use for b3 and b5 ,generally b5 always outperforms b0 ,same was also experience. Remember b5 is fine tuned on the top of b3,b4,b5 to work at higher image resolution base line size is 224 for b0
", "rawMarkdown": "@hesene what image resolution did u use for b3 and b5 ,generally b5 always outperforms b0 ,same was also experience. Remember b5 is fine tuned on the top of b3,b4,b5 to work at higher image resolution base line size is 224 for b0", "votes": 1, "replies": [ { "id": 697490, "postDate": "2019-12-18T01:55:12.997Z", "content": "Thank you for your sharing, my training size is same as kernel（2048*512）
", "rawMarkdown": "Thank you for your sharing, my training size is same as kernel（2048*512）", "votes": 1 }, { "id": 697503, "postDate": "2019-12-18T02:34:24.717Z", "content": "Have you tried image size as per documentation of b5 configs ..,progressive resizing ...which we train b,s .\nIf you prefer to team up let me know we can try some hacks together to boost the solution . Or else no issues I can also tell you otherwise as well
", "rawMarkdown": "Have you tried image size as per documentation of b5 configs ..,progressive resizing ...which we train b,s .\nIf you prefer to team up let me know we can try some hacks together to boost the solution . Or else no issues I can also tell you otherwise as well " }, { "id": 697672, "postDate": "2019-12-18T08:55:51.870Z", "content": "Thank you for you advice! I'm sorry I don't have a team up plan recently
", "rawMarkdown": "Thank you for you advice! I'm sorry I don't have a team up plan recently " }, { "id": 697846, "postDate": "2019-12-18T13:39:36.450Z", "content": "ok no worries.. all i was wanting to suggest is try progressive resizing training when using b series.
", "rawMarkdown": "ok no worries.. all i was wanting to suggest is try progressive resizing training when using b series." } ] }, { "id": 694365, "postDate": "2019-12-13T13:24:52.723Z", "content": "In my experience, yes. Atleast when I compared Resnet 18 and Resnet 52.\nMy current score is with DLA34, trying DLA102 now.
\n\nDid you train your network until your validation stopped improving?
", "rawMarkdown": "In my experience, yes. Atleast when I compared Resnet 18 and Resnet 52.\nMy current score is with DLA34, trying DLA102 now.\n\n\nDid you train your network until your validation stopped improving?", "votes": 1, "replies": [ { "id": 694366, "postDate": "2019-12-13T13:29:02.663Z", "content": "In your experience, DLA34 > Resnet 52 > Resnet 18? is that right?\nYes
", "rawMarkdown": "In your experience, DLA34 > Resnet 52 > Resnet 18? is that right?\nYes" }, { "id": 694390, "postDate": "2019-12-13T14:12:47.927Z", "content": "Yes.
", "rawMarkdown": "Yes." }, { "id": 696597, "postDate": "2019-12-16T20:17:02.373Z", "content": "@chroteus Hi! Are you intrested in teaming up? I have a 2080Ti and completed competitions previously with decent scores. I also have a pretty good LB score now. I think we can build good solutions together. Thanks!
", "rawMarkdown": "@chroteus Hi! Are you intrested in teaming up? I have a 2080Ti and completed competitions previously with decent scores. I also have a pretty good LB score now. I think we can build good solutions together. Thanks!" }, { "id": 698270, "postDate": "2019-12-19T02:08:13.270Z", "content": "@chroteus Hi! May I ask what input and output size did you use for DLA34? Thanks!
", "rawMarkdown": "@chroteus Hi! May I ask what input and output size did you use for DLA34? Thanks!" }, { "id": 698491, "postDate": "2019-12-19T10:09:15.803Z", "content": "Sure.\n512x512 input with DLA's downscale factor of 4. Not sure if its the optimal one, but it seems to work.
\n\nRegarding teaming up, I'm afraid I can not team up, sorry.
", "rawMarkdown": "Sure.\n512x512 input with DLA's downscale factor of 4. Not sure if its the optimal one, but it seems to work.\n\nRegarding teaming up, I'm afraid I can not team up, sorry.", "votes": 2 }, { "id": 698568, "postDate": "2019-12-19T12:07:14.543Z", "content": "@chroteus 512x512 as input and you get 0.066, is that right? really amazing !
", "rawMarkdown": "@chroteus 512x512 as input and you get 0.066, is that right? really amazing !", "votes": 1 }, { "id": 698732, "postDate": "2019-12-19T16:31:01.647Z", "content": "I really have a hard time working out with smaller image size . Probably there could be a bug in my public kernel or @hocop1 s kernel somewhere which makes it give good local validation but worse LB score.
", "rawMarkdown": "I really have a hard time working out with smaller image size . Probably there could be a bug in my public kernel or @hocop1 s kernel somewhere which makes it give good local validation but worse LB score." }, { "id": 698780, "postDate": "2019-12-19T17:25:41.583Z", "content": "@chroteus you mean crop image for 512x512?? or resize to square?
", "rawMarkdown": "@chroteus you mean crop image for 512x512?? or resize to square?" }, { "id": 698782, "postDate": "2019-12-19T17:27:08.367Z", "content": "Resize.
", "rawMarkdown": "Resize." }, { "id": 698787, "postDate": "2019-12-19T17:30:32.923Z", "content": "@chroteus do you predict on fullsize or on resized test ?
", "rawMarkdown": "@chroteus do you predict on fullsize or on resized test ?" }, { "id": 698793, "postDate": "2019-12-19T17:39:21.873Z", "content": "i too get good CV but havent tried for lb. one more change i did is i replaced hand written bce with std bce. \n,secondly i cast a doubt on model scale as i have not clearly understood rationale behind it\n@phoenix9032 do you understand that part,is there a need for it or can one skip that..
", "rawMarkdown": "i too get good CV but havent tried for lb. one more change i did is i replaced hand written bce with std bce. \n,secondly i cast a doubt on model scale as i have not clearly understood rationale behind it\n@phoenix9032 do you understand that part,is there a need for it or can one skip that.." }, { "id": 700061, "postDate": "2019-12-21T11:37:20.943Z", "content": "Did you downsize to 512 from full size or cropped image?
", "rawMarkdown": "Did you downsize to 512 from full size or cropped image? " } ] }, { "id": 700526, "postDate": "2019-12-22T06:35:48.317Z", "content": "mostly yes
", "rawMarkdown": "mostly yes\n", "votes": -1 }, { "id": 707189, "postDate": "2019-12-31T11:39:51.270Z", "content": "this post confused me, add one more confusion:\nfor me, resnet18 > resnet101 > dla34
", "rawMarkdown": "this post confused me, add one more confusion:\nfor me, resnet18 > resnet101 > dla34" }, { "id": 696838, "postDate": "2019-12-17T06:05:28.097Z", "content": "Did you use https://www.kaggle.com/hocop1/centernet-baseline or implement real CenterNet by yourself? I used https://www.kaggle.com/hocop1/centernet-baseline with inception V3 and got a bed result (but I didn't turn the hyperparameters, and maybe iterations were not enough). I think small models are better because deeper networks may end up with overfitting on training data since the amount of test data is almost 50% of that of the training data.\nBut this is only my opinion. 👀
", "rawMarkdown": "Did you use https://www.kaggle.com/hocop1/centernet-baseline or implement real CenterNet by yourself? I used https://www.kaggle.com/hocop1/centernet-baseline with inception V3 and got a bed result (but I didn't turn the hyperparameters, and maybe iterations were not enough). I think small models are better because deeper networks may end up with overfitting on training data since the amount of test data is almost 50% of that of the training data.\nBut this is only my opinion. 👀 ", "replies": [ { "id": 696880, "postDate": "2019-12-17T07:21:23.527Z", "content": "I use the kernel's centernet, Thanks for your sharing
", "rawMarkdown": "I use the kernel's centernet, Thanks for your sharing" } ] }, { "id": 696046, "postDate": "2019-12-16T03:44:30.577Z", "content": "I could got 0.14 local cv, but lb is only 0.045...
", "rawMarkdown": "I could got 0.14 local cv, but lb is only 0.045...", "replies": [ { "id": 696051, "postDate": "2019-12-16T03:54:42.630Z", "content": "Same here, it's quite unstable.
", "rawMarkdown": "Same here, it's quite unstable." }, { "id": 696052, "postDate": "2019-12-16T03:55:37.540Z", "content": "@niuddd Are you intrested in teaming up? I have a 2080Ti and a pretty good LB score now. I have completed competitions in the past with decent scores.
", "rawMarkdown": "@niuddd Are you intrested in teaming up? I have a 2080Ti and a pretty good LB score now. I have completed competitions in the past with decent scores." }, { "id": 696496, "postDate": "2019-12-16T17:13:19.153Z", "content": "I'm sorry I don't plan to put much time for this competition now...
", "rawMarkdown": "I'm sorry I don't plan to put much time for this competition now..." }, { "id": 696661, "postDate": "2019-12-16T22:51:36.443Z", "content": "It's fine! I am always available and looking forward to work with you if you would like to focus more on this competition or team up in future! @niuddd
", "rawMarkdown": "It's fine! I am always available and looking forward to work with you if you would like to focus more on this competition or team up in future! @niuddd " } ] }, { "id": 694747, "postDate": "2019-12-14T03:00:30.593Z", "content": "Hi! I am also testing performances of different efficienents. Do you want to team up? I have a 2080Ti and got pretty good scores in previous competitions. I am also planning to investigate more into efficientdet. @hesene Thanks!
", "rawMarkdown": "Hi! I am also testing performances of different efficienents. Do you want to team up? I have a 2080Ti and got pretty good scores in previous competitions. I am also planning to investigate more into efficientdet. @hesene Thanks!", "replies": [ { "id": 694754, "postDate": "2019-12-14T03:18:14.727Z", "content": "I'm sorry I don't have a team up plan recently, you can contact other people.
", "rawMarkdown": "I'm sorry I don't have a team up plan recently, you can contact other people. " }, { "id": 694760, "postDate": "2019-12-14T03:22:58.660Z", "content": "Sure!
", "rawMarkdown": "Sure!" } ] }, { "id": 694380, "postDate": "2019-12-13T13:49:12.837Z", "content": "Did you change the imaze size while trying for deeper model? or you are comparing against same imagesize,bacthsize and other hyperparameters
", "rawMarkdown": "Did you change the imaze size while trying for deeper model? or you are comparing against same imagesize,bacthsize and other hyperparameters ", "replies": [ { "id": 694385, "postDate": "2019-12-13T13:59:25.693Z", "content": "The larger network use the smaller batch size
", "rawMarkdown": "The larger network use the smaller batch size" }, { "id": 694392, "postDate": "2019-12-13T14:15:40.467Z", "content": "I think larger is better for me . Effiecientnetb3 was surely better than b0 for me . But b1 was worse.
", "rawMarkdown": "I think larger is better for me . Effiecientnetb3 was surely better than b0 for me . But b1 was worse." } ] }, { "id": 710651, "postDate": "2020-01-05T03:23:54.980Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 697160, "author_name": "Camaro", "author_url": "", "post_date": "2019-12-17T14:46:06.480000", "content": "Yes and No. For me, resnet101 > resnet50, but seresnext50 or 101 doesn't work good...
", "votes": 5, "replies": [ { "id": 697168, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-12-17T15:03:15.103000", "content": "I can confirm the same . I am not sure why though :(
", "votes": 0, "replies": [] }, { "id": 697527, "author_name": "Endi Niu", "author_url": "", "post_date": "2019-12-18T03:29:22.383000", "content": "Does using pre-trained weights make a difference?
", "votes": 1, "replies": [] }, { "id": 698056, "author_name": "Bibek", "author_url": "", "post_date": "2019-12-18T18:38:36.257000", "content": "Sounds like you are doing it from scratch?
", "votes": 0, "replies": [] }, { "id": 698171, "author_name": "Camaro", "author_url": "", "post_date": "2019-12-18T22:09:58.500000", "content": "No, I'm using pretrained weight.
", "votes": 1, "replies": [] }, { "id": 700414, "author_name": "ZHANG Zhi", "author_url": "", "post_date": "2019-12-22T01:23:59.347000", "content": "For me, resnext 101 < resnext 50 from scratch
", "votes": 0, "replies": [] } ] }, { "id": 697283, "author_name": "Jaideep", "author_url": "", "post_date": "2019-12-17T17:18:52.463000", "content": "@hesene what image resolution did u use for b3 and b5 ,generally b5 always outperforms b0 ,same was also experience. Remember b5 is fine tuned on the top of b3,b4,b5 to work at higher image resolution base line size is 224 for b0
", "votes": 1, "replies": [ { "id": 697490, "author_name": "He", "author_url": "", "post_date": "2019-12-18T01:55:12.997000", "content": "Thank you for your sharing, my training size is same as kernel（2048*512）
", "votes": 1, "replies": [] }, { "id": 697503, "author_name": "Jaideep", "author_url": "", "post_date": "2019-12-18T02:34:24.717000", "content": "Have you tried image size as per documentation of b5 configs ..,progressive resizing ...which we train b,s .\nIf you prefer to team up let me know we can try some hacks together to boost the solution . Or else no issues I can also tell you otherwise as well
", "votes": 0, "replies": [] }, { "id": 697672, "author_name": "He", "author_url": "", "post_date": "2019-12-18T08:55:51.870000", "content": "Thank you for you advice! I'm sorry I don't have a team up plan recently
", "votes": 0, "replies": [] }, { "id": 697846, "author_name": "Jaideep", "author_url": "", "post_date": "2019-12-18T13:39:36.450000", "content": "ok no worries.. all i was wanting to suggest is try progressive resizing training when using b series.
", "votes": 0, "replies": [] } ] }, { "id": 694365, "author_name": "Askar Bozcan", "author_url": "", "post_date": "2019-12-13T13:24:52.723000", "content": "In my experience, yes. Atleast when I compared Resnet 18 and Resnet 52.\nMy current score is with DLA34, trying DLA102 now.
\n\nDid you train your network until your validation stopped improving?
", "votes": 1, "replies": [ { "id": 694366, "author_name": "He", "author_url": "", "post_date": "2019-12-13T13:29:02.663000", "content": "In your experience, DLA34 > Resnet 52 > Resnet 18? is that right?\nYes
", "votes": 0, "replies": [] }, { "id": 694390, "author_name": "Askar Bozcan", "author_url": "", "post_date": "2019-12-13T14:12:47.927000", "content": "Yes.
", "votes": 0, "replies": [] }, { "id": 696597, "author_name": "Gold Retriever", "author_url": "", "post_date": "2019-12-16T20:17:02.373000", "content": "@chroteus Hi! Are you intrested in teaming up? I have a 2080Ti and completed competitions previously with decent scores. I also have a pretty good LB score now. I think we can build good solutions together. Thanks!
", "votes": 0, "replies": [] }, { "id": 698270, "author_name": "Gold Retriever", "author_url": "", "post_date": "2019-12-19T02:08:13.270000", "content": "@chroteus Hi! May I ask what input and output size did you use for DLA34? Thanks!
", "votes": 0, "replies": [] }, { "id": 698491, "author_name": "Askar Bozcan", "author_url": "", "post_date": "2019-12-19T10:09:15.803000", "content": "Sure.\n512x512 input with DLA's downscale factor of 4. Not sure if its the optimal one, but it seems to work.
\n\nRegarding teaming up, I'm afraid I can not team up, sorry.
", "votes": 2, "replies": [] }, { "id": 698568, "author_name": "He", "author_url": "", "post_date": "2019-12-19T12:07:14.543000", "content": "@chroteus 512x512 as input and you get 0.066, is that right? really amazing !
", "votes": 1, "replies": [] }, { "id": 698732, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-12-19T16:31:01.647000", "content": "I really have a hard time working out with smaller image size . Probably there could be a bug in my public kernel or @hocop1 s kernel somewhere which makes it give good local validation but worse LB score.
", "votes": 0, "replies": [] }, { "id": 698780, "author_name": "Camaro", "author_url": "", "post_date": "2019-12-19T17:25:41.583000", "content": "@chroteus you mean crop image for 512x512?? or resize to square?
", "votes": 0, "replies": [] }, { "id": 698782, "author_name": "Askar Bozcan", "author_url": "", "post_date": "2019-12-19T17:27:08.367000", "content": "Resize.
", "votes": 0, "replies": [] }, { "id": 698787, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-12-19T17:30:32.923000", "content": "@chroteus do you predict on fullsize or on resized test ?
", "votes": 0, "replies": [] }, { "id": 698793, "author_name": "Jaideep", "author_url": "", "post_date": "2019-12-19T17:39:21.873000", "content": "i too get good CV but havent tried for lb. one more change i did is i replaced hand written bce with std bce. \n,secondly i cast a doubt on model scale as i have not clearly understood rationale behind it\n@phoenix9032 do you understand that part,is there a need for it or can one skip that..
", "votes": 0, "replies": [] }, { "id": 700061, "author_name": "Endi Niu", "author_url": "", "post_date": "2019-12-21T11:37:20.943000", "content": "Did you downsize to 512 from full size or cropped image?
", "votes": 0, "replies": [] } ] }, { "id": 700526, "author_name": "Bijayan Bhattarai", "author_url": "", "post_date": "2019-12-22T06:35:48.317000", "content": "mostly yes
", "votes": -1, "replies": [] }, { "id": 707189, "author_name": "Endi Niu", "author_url": "", "post_date": "2019-12-31T11:39:51.270000", "content": "this post confused me, add one more confusion:\nfor me, resnet18 > resnet101 > dla34
", "votes": 0, "replies": [] }, { "id": 696838, "author_name": "Math first", "author_url": "", "post_date": "2019-12-17T06:05:28.097000", "content": "Did you use https://www.kaggle.com/hocop1/centernet-baseline or implement real CenterNet by yourself? I used https://www.kaggle.com/hocop1/centernet-baseline with inception V3 and got a bed result (but I didn't turn the hyperparameters, and maybe iterations were not enough). I think small models are better because deeper networks may end up with overfitting on training data since the amount of test data is almost 50% of that of the training data.\nBut this is only my opinion. 👀
", "votes": 0, "replies": [ { "id": 696880, "author_name": "He", "author_url": "", "post_date": "2019-12-17T07:21:23.527000", "content": "I use the kernel's centernet, Thanks for your sharing
", "votes": 0, "replies": [] } ] }, { "id": 696046, "author_name": "Endi Niu", "author_url": "", "post_date": "2019-12-16T03:44:30.577000", "content": "I could got 0.14 local cv, but lb is only 0.045...
", "votes": 0, "replies": [ { "id": 696051, "author_name": "Gold Retriever", "author_url": "", "post_date": "2019-12-16T03:54:42.630000", "content": "Same here, it's quite unstable.
", "votes": 0, "replies": [] }, { "id": 696052, "author_name": "Gold Retriever", "author_url": "", "post_date": "2019-12-16T03:55:37.540000", "content": "@niuddd Are you intrested in teaming up? I have a 2080Ti and a pretty good LB score now. I have completed competitions in the past with decent scores.
", "votes": 0, "replies": [] }, { "id": 696496, "author_name": "Endi Niu", "author_url": "", "post_date": "2019-12-16T17:13:19.153000", "content": "I'm sorry I don't plan to put much time for this competition now...
", "votes": 0, "replies": [] }, { "id": 696661, "author_name": "Gold Retriever", "author_url": "", "post_date": "2019-12-16T22:51:36.443000", "content": "It's fine! I am always available and looking forward to work with you if you would like to focus more on this competition or team up in future! @niuddd
", "votes": 0, "replies": [] } ] }, { "id": 694747, "author_name": "Gold Retriever", "author_url": "", "post_date": "2019-12-14T03:00:30.593000", "content": "Hi! I am also testing performances of different efficienents. Do you want to team up? I have a 2080Ti and got pretty good scores in previous competitions. I am also planning to investigate more into efficientdet. @hesene Thanks!
", "votes": 0, "replies": [ { "id": 694754, "author_name": "He", "author_url": "", "post_date": "2019-12-14T03:18:14.727000", "content": "I'm sorry I don't have a team up plan recently, you can contact other people.
", "votes": 0, "replies": [] }, { "id": 694760, "author_name": "Gold Retriever", "author_url": "", "post_date": "2019-12-14T03:22:58.660000", "content": "Sure!
", "votes": 0, "replies": [] } ] }, { "id": 694380, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-12-13T13:49:12.837000", "content": "Did you change the imaze size while trying for deeper model? or you are comparing against same imagesize,bacthsize and other hyperparameters
", "votes": 0, "replies": [ { "id": 694385, "author_name": "He", "author_url": "", "post_date": "2019-12-13T13:59:25.693000", "content": "The larger network use the smaller batch size
", "votes": 0, "replies": [] }, { "id": 694392, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-12-13T14:15:40.467000", "content": "I think larger is better for me . Effiecientnetb3 was surely better than b0 for me . But b1 was worse.
", "votes": 0, "replies": [] } ] }, { "id": 710651, "author_name": "", "author_url": "", "post_date": "2020-01-05T03:23:54.980000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "694342": "In the paper of CenterNet, deeper networks seem to perform better for object detection, but in this competition, my experiment was the best at efficientnet-b0, which was better than b3 and b5. My CV was 0.127(use https://www.kaggle.com/its7171/metrics-evaluation-script and 0.2 for val), and lb was 0.56. Do you get better performance using deeper networks ? \n", "697160": "Yes and No. For me, resnet101 > resnet50, but seresnext50 or 101 doesn't work good...", "697283": "@hesene what image resolution did u use for b3 and b5 ,generally b5 always outperforms b0 ,same was also experience. Remember b5 is fine tuned on the top of b3,b4,b5 to work at higher image resolution base line size is 224 for b0", "694365": "In my experience, yes. Atleast when I compared Resnet 18 and Resnet 52.\nMy current score is with DLA34, trying DLA102 now.\n\n\nDid you train your network until your validation stopped improving?", "700526": "mostly yes\n", "707189": "this post confused me, add one more confusion:\nfor me, resnet18 > resnet101 > dla34", "696838": "Did you use https://www.kaggle.com/hocop1/centernet-baseline or implement real CenterNet by yourself? I used https://www.kaggle.com/hocop1/centernet-baseline with inception V3 and got a bed result (but I didn't turn the hyperparameters, and maybe iterations were not enough). I think small models are better because deeper networks may end up with overfitting on training data since the amount of test data is almost 50% of that of the training data.\nBut this is only my opinion. 👀 ", "696046": "I could got 0.14 local cv, but lb is only 0.045...", "694747": "Hi! I am also testing performances of different efficienents. Do you want to team up? I have a 2080Ti and got pretty good scores in previous competitions. I am also planning to investigate more into efficientdet. @hesene Thanks!", "694380": "Did you change the imaze size while trying for deeper model? or you are comparing against same imagesize,bacthsize and other hyperparameters ", "710651": "" } }