{ "id": 317836, "title": "EffNetb4 vs EffNetb5", "url": "/competitions/happy-whale-and-dolphin/discussion/317836", "author_name": "", "post_date": "2022-04-09T05:58:38.657845500Z", "votes": 7, "comment_count": 4, "views": 0, "content": "

I trained two models b4 and b5 with the same data and here are my results,

\n\n\n\n \n \n \n \n \n\n\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n\n
B4B5
Batch Size
(Adjusted for 16 gb ram)		3216
Image Size384512
Traning Time3hr 21min8hr 54min
Epocs1010
Validation Acc.
(During Training)		84.987
\n

What can be done to get more accuracy from models like EffNets b6 &b7?

", "messages": [ { "id": "1749890", "postDate": "04/09/2022 05:58:38", "content": "

I trained two models b4 and b5 with the same data and here are my results,

\n\n\n\n \n \n \n \n \n\n\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n\n
B4B5
Batch Size
(Adjusted for 16 gb ram)		3216
Image Size384512
Traning Time3hr 21min8hr 54min
Epocs1010
Validation Acc.
(During Training)		84.987
\n

What can be done to get more accuracy from models like EffNets b6 &b7?

", "rawMarkdown": "I trained two models b4 and b5 with the same data and here are my results,\n\n\n\n\n \n \n \n \n \n\n\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n\n
B4B5
Batch Size
(Adjusted for 16 gb ram)		3216
Image Size384512
Traning Time3hr 21min8hr 54min
Epocs1010
Validation Acc.
(During Training)		84.987
\n\nWhat can be done to get more accuracy from models like EffNets b6 &b7?", "votes": null }, { "id": "1749936", "postDate": "04/09/2022 06:38:21", "content": "

There are a lot of interesting things to try in this competition:

\n", "rawMarkdown": "There are a lot of interesting things to try in this competition:\n- different datasets (full body, just fins or even you own one if you want to create a new one)\n- train more epochs, use larger images\n- different augmentation techniques\n- different head fully connected layers architectures\n- dropout rate \n- stochasting weight average \n- post process techniques: different embeddings blending methods, etc.\n- pre process techniques: remove background from images, etc.", "votes": null }, { "id": "1749945", "postDate": "04/09/2022 06:52:01", "content": "

I tried with removing the background but my model loss was stuck at 13.
\nI tried everything you mentioned but was not able to train my model to get any accuracy (as the loss got stuck around 13-15 everytime).

\n

I still need to try post process techniques.

", "rawMarkdown": "I tried with removing the background but my model loss was stuck at 13.\nI tried everything you mentioned but was not able to train my model to get any accuracy (as the loss got stuck around 13-15 everytime).\n\nI still need to try post process techniques.", "votes": null }, { "id": "1750041", "postDate": "04/09/2022 09:00:13", "content": "

I also tried to firstly train on small image size (224x224 or 256x256). After that, I did additional training of the same model on larger images (size 384x384 or even 512x512). It also helped to increase validation accuracy and LB score.

", "rawMarkdown": "I also tried to firstly train on small image size (224x224 or 256x256). After that, I did additional training of the same model on larger images (size 384x384 or even 512x512). It also helped to increase validation accuracy and LB score.", "votes": null }, { "id": "1750060", "postDate": "04/09/2022 09:23:03", "content": "

I tried that but couldn't better accuracy

", "rawMarkdown": "I tried that but couldn't better accuracy", "votes": null } ], "comments": [ { "id": 1749936, "author_name": "vladvdv", "author_url": "", "post_date": "04/09/2022 06:38:21", "content": "

There are a lot of interesting things to try in this competition:

\n", "votes": null, "replies": [ { "id": 1749945, "author_name": "jainishsavalia", "author_url": "", "post_date": "04/09/2022 06:52:01", "content": "

I tried with removing the background but my model loss was stuck at 13.
\nI tried everything you mentioned but was not able to train my model to get any accuracy (as the loss got stuck around 13-15 everytime).

\n

I still need to try post process techniques.

", "votes": null, "replies": [] }, { "id": 1750041, "author_name": "alexeyolkhovikov", "author_url": "", "post_date": "04/09/2022 09:00:13", "content": "

I also tried to firstly train on small image size (224x224 or 256x256). After that, I did additional training of the same model on larger images (size 384x384 or even 512x512). It also helped to increase validation accuracy and LB score.

", "votes": null, "replies": [] }, { "id": 1750060, "author_name": "jainishsavalia", "author_url": "", "post_date": "04/09/2022 09:23:03", "content": "

I tried that but couldn't better accuracy

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1749890": "I trained two models b4 and b5 with the same data and here are my results,\n\n\n\n\n \n \n \n \n \n\n\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n\n
B4B5
Batch Size
(Adjusted for 16 gb ram)		3216
Image Size384512
Traning Time3hr 21min8hr 54min
Epocs1010
Validation Acc.
(During Training)		84.987
\n\nWhat can be done to get more accuracy from models like EffNets b6 &b7?", "1749936": "There are a lot of interesting things to try in this competition:\n- different datasets (full body, just fins or even you own one if you want to create a new one)\n- train more epochs, use larger images\n- different augmentation techniques\n- different head fully connected layers architectures\n- dropout rate \n- stochasting weight average \n- post process techniques: different embeddings blending methods, etc.\n- pre process techniques: remove background from images, etc.", "1749945": "I tried with removing the background but my model loss was stuck at 13.\nI tried everything you mentioned but was not able to train my model to get any accuracy (as the loss got stuck around 13-15 everytime).\n\nI still need to try post process techniques.", "1750041": "I also tried to firstly train on small image size (224x224 or 256x256). After that, I did additional training of the same model on larger images (size 384x384 or even 512x512). It also helped to increase validation accuracy and LB score.", "1750060": "I tried that but couldn't better accuracy" }, "source": "meta" }