Wondering if this is true, tested for epochs 100,300 and 400, though not improved much but still 300 to 400 has LB gain
\n(updated 1 400 epoch model score +4 points then 300 epochs on LB…).
\nBut I did not found gain(300 epochs to 400 epochs) if split particpant ids for cv locally.
\nIf not cosidering participant ids then surely more epochs produce better results locally.
\nSo if there are duplicate participant ids between train and test then we need more epochs of training just as the learning eqaulity competition.
Wondering if this is true, tested for epochs 100,300 and 400, though not improved much but still 300 to 400 has LB gain
\n(updated 1 400 epoch model score +4 points then 300 epochs on LB…).
\nBut I did not found gain(300 epochs to 400 epochs) if split particpant ids for cv locally.
\nIf not cosidering participant ids then surely more epochs produce better results locally.
\nSo if there are duplicate participant ids between train and test then we need more epochs of training just as the learning eqaulity competition.
it reminds me of the old ancient adaboost training ….
\nthen people asked
\n\"why does validation accuracy still improveing when you your training set already have 100% accuracy?\"
answer
\n\"becuase margin is still improving\"
on a side note, it is very difficult for me to do comparsion experiments.
\nsome of the gain only happens at the last (long) epoch. I Since have changed multiple hyperparameters in each experiments, now i don't know which exactly has worked
Large improvement still can be tested locally very well, I think LB 1st also has good local result so they can hide their score untill the last days😀. Personally I think online has both old and new participant ids and still I would recommend local test without participant ids overlap though your local score will be lower then online, I found it has better alignment with LB.
", "rawMarkdown": "Large improvement still can be tested locally very well, I think LB 1st also has good local result so they can hide their score untill the last days😀. Personally I think online has both old and new participant ids and still I would recommend local test without participant ids overlap though your local score will be lower then online, I found it has better alignment with LB." } ] }, { "id": 2405501, "postDate": "2023-08-23T23:37:03.760Z", "content": "For me, increasing epochs from 300 to 400 improved local cv (not split by participant id) but decreased lb score
", "rawMarkdown": "For me, increasing epochs from 300 to 400 improved local cv (not split by participant id) but decreased lb score", "replies": [ { "id": 2405542, "postDate": "2023-08-24T01:04:37.400Z", "content": "\n", "rawMarkdown": "> For me, increasing epochs from 300 to 400 improved local cv (not split by participant id) but decreased lb score\nFor my old model 300 to 400 not improve LB but after increasing aug prob it has 1 point LB gain.\n\n", "replies": [ { "id": 2406090, "postDate": "2023-08-24T08:05:56.960Z", "rawMarkdown": "", "isDeleted": true } ] } ] } ], "comments": [ { "id": 2405507, "author_name": "hengck23", "author_url": "", "post_date": "2023-08-24T00:05:00.950000", "content": "For me, increasing epochs from 300 to 400 improved local cv (not split by participant id) but decreased lb score
\n
\n For my old model 300 to 400 not improve LB but after increasing aug prob it has 1 point LB gain.
it reminds me of the old ancient adaboost training ….
\nthen people asked
\n\"why does validation accuracy still improveing when you your training set already have 100% accuracy?\"
answer
\n\"becuase margin is still improving\"
on a side note, it is very difficult for me to do comparsion experiments.
\nsome of the gain only happens at the last (long) epoch. I Since have changed multiple hyperparameters in each experiments, now i don't know which exactly has worked
Large improvement still can be tested locally very well, I think LB 1st also has good local result so they can hide their score untill the last days😀. Personally I think online has both old and new participant ids and still I would recommend local test without participant ids overlap though your local score will be lower then online, I found it has better alignment with LB.
", "votes": 0, "replies": [] } ] }, { "id": 2405501, "author_name": "Yu Wu", "author_url": "", "post_date": "2023-08-23T23:37:03.760000", "content": "For me, increasing epochs from 300 to 400 improved local cv (not split by participant id) but decreased lb score
", "votes": 0, "replies": [ { "id": 2405542, "author_name": "gezi", "author_url": "", "post_date": "2023-08-24T01:04:37.400000", "content": "\n", "votes": 0, "replies": [ { "id": 2406090, "author_name": "", "author_url": "", "post_date": "2023-08-24T08:05:56.960000", "content": "", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2405492": "Wondering if this is true, tested for epochs 100,300 and 400, though not improved much but still 300 to 400 has LB gain\n(updated 1 400 epoch model score +4 points then 300 epochs on LB...).\nBut I did not found gain(300 epochs to 400 epochs) if split particpant ids for cv locally. \nIf not cosidering participant ids then surely more epochs produce better results locally.\nSo if there are duplicate participant ids between train and test then we need more epochs of training just as the learning eqaulity competition.", "2405507": "it reminds me of the old ancient adaboost training .... \nthen people asked\n\"why does validation accuracy still improveing when you your training set already have 100% accuracy?\"\n\nanswer\n\"becuase margin is still improving\"\n\n---\n\non a side note, it is very difficult for me to do comparsion experiments.\nsome of the gain only happens at the last (long) epoch. I Since have changed multiple hyperparameters in each experiments, now i don't know which exactly has worked", "2405501": "For me, increasing epochs from 300 to 400 improved local cv (not split by participant id) but decreased lb score" } }For me, increasing epochs from 300 to 400 improved local cv (not split by participant id) but decreased lb score
\n
\n For my old model 300 to 400 not improve LB but after increasing aug prob it has 1 point LB gain.