I tried efficientnet, resnet. They are giving 0.8x accuracy in valid set on my local machine but when i submit the same model it show half the accuracy. I tried augmentation too.
\nWhat can i do ?
I tried efficientnet, resnet. They are giving 0.8x accuracy in valid set on my local machine but when i submit the same model it show half the accuracy. I tried augmentation too.
\nWhat can i do ?
I have done both cross validation and stratified validation.
", "rawMarkdown": "I have done both cross validation and stratified validation.", "votes": null }, { "id": "1099584", "postDate": "12/02/2020 13:06:09", "content": "If you are really getting half the accuracy - 0.40 or 0.61, you have a coding problem. Also, the original TFRecords had errors and were re-released. Make sure you have the current ones.
\nDo you have a \"shuffle\" when retrieving your test data that you shouldn't?
", "rawMarkdown": "If you are really getting half the accuracy - 0.40 or 0.61, you have a coding problem. Also, the original TFRecords had errors and were re-released. Make sure you have the current ones.\n\nDo you have a \"shuffle\" when retrieving your test data that you shouldn't?", "votes": null }, { "id": "1099632", "postDate": "12/02/2020 13:48:07", "content": "You are HERO. It is my first running competition and i spend all of my mind power on this problem. Really it was so depressing haven't slept properly this week.
\nAnd stupid problem was shuffle True for test set.
\nReally thankyou. Great Work.
\nI tested and now my accuracy is 0.812. Feeling motivated
", "rawMarkdown": "You are HERO. It is my first running competition and i spend all of my mind power on this problem. Really it was so depressing haven't slept properly this week. \n\nAnd stupid problem was shuffle True for test set.\n\nReally thankyou. Great Work.\n\nI tested and now my accuracy is 0.812. Feeling motivated", "votes": null }, { "id": "1100347", "postDate": "12/03/2020 03:04:56", "content": "i have 0.89 cv - 0.89 lb
", "rawMarkdown": "i have 0.89 cv - 0.89 lb", "votes": null }, { "id": "1100489", "postDate": "12/03/2020 05:38:32", "content": "Great work. I be Working for same and more socres xD.
", "rawMarkdown": "Great work. I be Working for same and more socres xD.", "votes": null }, { "id": "1104112", "postDate": "12/06/2020 16:20:14", "content": "Hi @rajanlagah , you are not alone. In my last competition I set Shuffle=False for training data. I must also mention that I used external data, so eventually it was (train data+external data) NOT SHUFFLED. It took me many days to realize where my mistake was.
", "rawMarkdown": "Hi @rajanlagah , you are not alone. In my last competition I set Shuffle=False for **training data**. I must also mention that I used external data, so eventually it was (train data+external data) NOT SHUFFLED. It took me many days to realize where my mistake was.", "votes": null }, { "id": "1105459", "postDate": "12/07/2020 22:34:55", "content": "Did you used fast.ai library?
", "rawMarkdown": "Did you used fast.ai library?", "votes": null }, { "id": "1105855", "postDate": "12/08/2020 09:03:47", "content": "No. simple Pytorch
", "rawMarkdown": "No. simple Pytorch", "votes": null }, { "id": "1130956", "postDate": "12/29/2020 12:59:27", "content": "Hahahaha ..
\nKaggle is awesome. This is my first running competition and i have learn a lot of things. Will definitely continue
Right, awesome. And awesome Kaggle shit also happens.
", "rawMarkdown": "Right, awesome. And awesome Kaggle shit also happens.", "votes": null } ], "comments": [ { "id": 1099495, "author_name": "rajanlagah", "author_url": "", "post_date": "12/02/2020 12:03:30", "content": "I have done both cross validation and stratified validation.
", "votes": null, "replies": [] }, { "id": 1099584, "author_name": "richardepstein", "author_url": "", "post_date": "12/02/2020 13:06:09", "content": "If you are really getting half the accuracy - 0.40 or 0.61, you have a coding problem. Also, the original TFRecords had errors and were re-released. Make sure you have the current ones.
\nDo you have a \"shuffle\" when retrieving your test data that you shouldn't?
", "votes": null, "replies": [ { "id": 1099632, "author_name": "rajanlagah", "author_url": "", "post_date": "12/02/2020 13:48:07", "content": "You are HERO. It is my first running competition and i spend all of my mind power on this problem. Really it was so depressing haven't slept properly this week.
\nAnd stupid problem was shuffle True for test set.
\nReally thankyou. Great Work.
\nI tested and now my accuracy is 0.812. Feeling motivated
", "votes": null, "replies": [] }, { "id": 1104112, "author_name": "dunklerwald", "author_url": "", "post_date": "12/06/2020 16:20:14", "content": "Hi @rajanlagah , you are not alone. In my last competition I set Shuffle=False for training data. I must also mention that I used external data, so eventually it was (train data+external data) NOT SHUFFLED. It took me many days to realize where my mistake was.
", "votes": null, "replies": [] }, { "id": 1130956, "author_name": "rajanlagah", "author_url": "", "post_date": "12/29/2020 12:59:27", "content": "Hahahaha ..
\nKaggle is awesome. This is my first running competition and i have learn a lot of things. Will definitely continue
Right, awesome. And awesome Kaggle shit also happens.
", "votes": null, "replies": [] } ] }, { "id": 1100347, "author_name": "kupchanski", "author_url": "", "post_date": "12/03/2020 03:04:56", "content": "i have 0.89 cv - 0.89 lb
", "votes": null, "replies": [ { "id": 1100489, "author_name": "rajanlagah", "author_url": "", "post_date": "12/03/2020 05:38:32", "content": "Great work. I be Working for same and more socres xD.
", "votes": null, "replies": [] } ] }, { "id": 1105459, "author_name": "volcanoflash", "author_url": "", "post_date": "12/07/2020 22:34:55", "content": "Did you used fast.ai library?
", "votes": null, "replies": [ { "id": 1105855, "author_name": "rajanlagah", "author_url": "", "post_date": "12/08/2020 09:03:47", "content": "No. simple Pytorch
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1099485": "I tried efficientnet, resnet. They are giving 0.8x accuracy in valid set on my local machine but when i submit the same model it show half the accuracy. I tried augmentation too. \nWhat can i do ?", "1099495": "I have done both cross validation and stratified validation.", "1099584": "If you are really getting half the accuracy - 0.40 or 0.61, you have a coding problem. Also, the original TFRecords had errors and were re-released. Make sure you have the current ones.\n\nDo you have a \"shuffle\" when retrieving your test data that you shouldn't?", "1099632": "You are HERO. It is my first running competition and i spend all of my mind power on this problem. Really it was so depressing haven't slept properly this week. \n\nAnd stupid problem was shuffle True for test set.\n\nReally thankyou. Great Work.\n\nI tested and now my accuracy is 0.812. Feeling motivated", "1100347": "i have 0.89 cv - 0.89 lb", "1100489": "Great work. I be Working for same and more socres xD.", "1104112": "Hi @rajanlagah , you are not alone. In my last competition I set Shuffle=False for **training data**. I must also mention that I used external data, so eventually it was (train data+external data) NOT SHUFFLED. It took me many days to realize where my mistake was.", "1105459": "Did you used fast.ai library?", "1105855": "No. simple Pytorch", "1130956": "Hahahaha .. \nKaggle is awesome. This is my first running competition and i have learn a lot of things. Will definitely continue", "1131004": "Right, awesome. And awesome Kaggle shit also happens." }, "source": "meta" }