![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5609063%2F12bb6f2bd23d7e15b46b38a3d69f0054%2FScreenshot%202025-03-24%20185044.png?generation=1742838673033988&alt=media\")
Hello!
\nDoes it make any sense here for the validation AUC to be lower than the training AUC during the first few epochs (3-4 first epochs)? Considering a base training with EfficientNetB0.
\nI'm getting this consistently on the first few epochs of all the folds.
\nThanks for the feedback!
", "messages": [ { "id": "3158099", "postDate": "03/24/2025 08:40:03", "content": "Hello!
\nDoes it make any sense here for the validation AUC to be lower than the training AUC during the first few epochs (3-4 first epochs)? Considering a base training with EfficientNetB0.
\nI'm getting this consistently on the first few epochs of all the folds.
\nThanks for the feedback!
", "rawMarkdown": "Hello! \n\nDoes it make any sense here for the validation AUC to be lower than the training AUC during the first few epochs (3-4 first epochs)? Considering a base training with EfficientNetB0. \n\nI'm getting this consistently on the first few epochs of all the folds. \n\nThanks for the feedback!", "votes": null }, { "id": "3158615", "postDate": "03/24/2025 17:46:17", "content": "I have seen this often over time - I assume you talking about some kind of ugly values.
", "rawMarkdown": "I have seen this often over time - I assume you talking about some kind of ugly values.", "votes": null }, { "id": "3158619", "postDate": "03/24/2025 17:51:39", "content": "What do you mean?
\nI meant just the training and evaluation after each epoch. Here for example:
Yes I understand what you meant - those are kind of ugly values, ie, they are not near the final result and would make a model that's of low value.
\nNothing looks abnormal about the start - how was the end?
\nFor example, here is the first epoch of the first B0 shared model that I grabbed to look at.
\nEpoch 1/8, Loss: 3.6371, Loss_val: 0.6873
\nAuc: 0.77% Auc_val: 0.93%
Nice to know its normal. Yes, at later epochs it inverts as i expected, train AUC grows faster and gets bigger than val AUC
", "rawMarkdown": "Nice to know its normal. Yes, at later epochs it inverts as i expected, train AUC grows faster and gets bigger than val AUC", "votes": null }, { "id": "3159794", "postDate": "03/25/2025 23:11:30", "content": "Most pre-trained models are going to give you a decent value with no training. In the first epochs, our training starts moving the weights and often some of that movement over corrects and leads to lower train values vs validation.
", "rawMarkdown": "Most pre-trained models are going to give you a decent value with no training. In the first epochs, our training starts moving the weights and often some of that movement over corrects and leads to lower train values vs validation.", "votes": null } ], "comments": [ { "id": 3158615, "author_name": "pcjimmmy", "author_url": "", "post_date": "03/24/2025 17:46:17", "content": "I have seen this often over time - I assume you talking about some kind of ugly values.
", "votes": null, "replies": [ { "id": 3158619, "author_name": "diegoakel", "author_url": "", "post_date": "03/24/2025 17:51:39", "content": "What do you mean?
\nI meant just the training and evaluation after each epoch. Here for example:
Yes I understand what you meant - those are kind of ugly values, ie, they are not near the final result and would make a model that's of low value.
\nNothing looks abnormal about the start - how was the end?
\nFor example, here is the first epoch of the first B0 shared model that I grabbed to look at.
\nEpoch 1/8, Loss: 3.6371, Loss_val: 0.6873
\nAuc: 0.77% Auc_val: 0.93%
Nice to know its normal. Yes, at later epochs it inverts as i expected, train AUC grows faster and gets bigger than val AUC
", "votes": null, "replies": [ { "id": 3159794, "author_name": "pcjimmmy", "author_url": "", "post_date": "03/25/2025 23:11:30", "content": "Most pre-trained models are going to give you a decent value with no training. In the first epochs, our training starts moving the weights and often some of that movement over corrects and leads to lower train values vs validation.
", "votes": null, "replies": [] } ] } ] } ] } ] } ], "raw_markdown_by_id": { "3158099": "Hello! \n\nDoes it make any sense here for the validation AUC to be lower than the training AUC during the first few epochs (3-4 first epochs)? Considering a base training with EfficientNetB0. \n\nI'm getting this consistently on the first few epochs of all the folds. \n\nThanks for the feedback!", "3158615": "I have seen this often over time - I assume you talking about some kind of ugly values.", "3158619": "What do you mean? \n\nI meant just the training and evaluation after each epoch. Here for example: ", "3158815": "Yes I understand what you meant - those are kind of ugly values, ie, they are not near the final result and would make a model that's of low value.\n\nNothing looks abnormal about the start - how was the end? \n\nFor example, here is the first epoch of the first B0 shared model that I grabbed to look at.\n\nEpoch 1/8, Loss: 3.6371, Loss_val: 0.6873\nAuc: 0.77% Auc_val: 0.93%", "3159323": "Nice to know its normal. Yes, at later epochs it inverts as i expected, train AUC grows faster and gets bigger than val AUC", "3159794": "Most pre-trained models are going to give you a decent value with no training. In the first epochs, our training starts moving the weights and often some of that movement over corrects and leads to lower train values vs validation." }, "source": "meta" }