Hello kagglers!
\nI'm just wondering that is there anyone who got this two train_loss and val_loss both converge to 0? I used the dataset1 from another competition and used several augmentations but I couldn't get the result that the both val and train's loss converged to 0…. I'm having doubt of this predicting MGMT by MRIs are possible… (( T _ T ))
\nIs there anyone who caught these two val and train loss?? and if so can I ask which method you used roughly?…..
Hello kagglers!
\nI'm just wondering that is there anyone who got this two train_loss and val_loss both converge to 0? I used the dataset1 from another competition and used several augmentations but I couldn't get the result that the both val and train's loss converged to 0…. I'm having doubt of this predicting MGMT by MRIs are possible… (( T _ T ))
\nIs there anyone who caught these two val and train loss?? and if so can I ask which method you used roughly?…..
Depends on the used loss function, but generally speaking (I assume you are using BCE/CE) both losses converging to 0 sounds like something impossible to achieve. You would need to have a perfect model that never makes mistakes, always predicting exactly 0.0 or 1.0 without any numbers after the decimal points. Furthermore, if the validation loss converges to the same value as training loss, it would mean that your model generalizes incredibly well (or your validation data is wrong). Keep in mind that the competition dataset is very small and it is difficult to train a good model with this amount of data.
", "rawMarkdown": "Depends on the used loss function, but generally speaking (I assume you are using BCE/CE) both losses converging to 0 sounds like something impossible to achieve. You would need to have a perfect model that never makes mistakes, always predicting exactly 0.0 or 1.0 without any numbers after the decimal points. Furthermore, if the validation loss converges to the same value as training loss, it would mean that your model generalizes incredibly well (or your validation data is wrong). Keep in mind that the competition dataset is very small and it is difficult to train a good model with this amount of data.", "votes": null }, { "id": "1515311", "postDate": "09/17/2021 01:35:21", "content": "Thank you for your reply! There was a mistake in my comment what I meant was the BCE is not going below 0.68!
", "rawMarkdown": "Thank you for your reply! There was a mistake in my comment what I meant was the BCE is not going below 0.68!", "votes": null }, { "id": "1516974", "postDate": "09/19/2021 03:28:18", "content": "If you can't get below 0.68 on training, try higher epochs to overfit. (just an indication of if model can overfit not relevant to testing). If you cant get test / val below 0.68, thats kinda expected 😅
", "rawMarkdown": "If you can't get below 0.68 on training, try higher epochs to overfit. (just an indication of if model can overfit not relevant to testing). If you cant get test / val below 0.68, thats kinda expected 😅", "votes": null } ], "comments": [ { "id": 1514947, "author_name": "mikecho", "author_url": "", "post_date": "09/16/2021 15:14:52", "content": "Depends on the used loss function, but generally speaking (I assume you are using BCE/CE) both losses converging to 0 sounds like something impossible to achieve. You would need to have a perfect model that never makes mistakes, always predicting exactly 0.0 or 1.0 without any numbers after the decimal points. Furthermore, if the validation loss converges to the same value as training loss, it would mean that your model generalizes incredibly well (or your validation data is wrong). Keep in mind that the competition dataset is very small and it is difficult to train a good model with this amount of data.
", "votes": null, "replies": [ { "id": 1515311, "author_name": "sungjangwon", "author_url": "", "post_date": "09/17/2021 01:35:21", "content": "Thank you for your reply! There was a mistake in my comment what I meant was the BCE is not going below 0.68!
", "votes": null, "replies": [] }, { "id": 1516974, "author_name": "aryamansharma47", "author_url": "", "post_date": "09/19/2021 03:28:18", "content": "If you can't get below 0.68 on training, try higher epochs to overfit. (just an indication of if model can overfit not relevant to testing). If you cant get test / val below 0.68, thats kinda expected 😅
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1514776": "Hello kagglers! \nI'm just wondering that is there anyone who got this two train_loss and val_loss both converge to 0? I used the dataset1 from another competition and used several augmentations but I couldn't get the result that the both val and train's loss converged to 0.... I'm having doubt of this predicting MGMT by MRIs are possible... (( T _ T ))\nIs there anyone who caught these two val and train loss?? and if so can I ask which method you used roughly?.....", "1514947": "Depends on the used loss function, but generally speaking (I assume you are using BCE/CE) both losses converging to 0 sounds like something impossible to achieve. You would need to have a perfect model that never makes mistakes, always predicting exactly 0.0 or 1.0 without any numbers after the decimal points. Furthermore, if the validation loss converges to the same value as training loss, it would mean that your model generalizes incredibly well (or your validation data is wrong). Keep in mind that the competition dataset is very small and it is difficult to train a good model with this amount of data.", "1515311": "Thank you for your reply! There was a mistake in my comment what I meant was the BCE is not going below 0.68!", "1516974": "If you can't get below 0.68 on training, try higher epochs to overfit. (just an indication of if model can overfit not relevant to testing). If you cant get test / val below 0.68, thats kinda expected 😅" }, "source": "meta" }