![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F761268%2Fada3062628ae3efd26a92f6be0039526%2FScreenshot%202020-03-19%20at%2012.27.46%20PM.png?generation=1584601086298610&alt=media\")
Training with https://www.kaggle.com/unkownhihi/dfdc-lrcn-training\n@unkownhihi is it normal or is my train/test split issue ? can i trust training loss ?
", "messages": [ { "id": 779278, "postDate": "2020-03-19T06:58:12.713Z", "content": "
Training with https://www.kaggle.com/unkownhihi/dfdc-lrcn-training\n@unkownhihi is it normal or is my train/test split issue ? can i trust training loss ?
", "rawMarkdown": "\n\n\nTraining with https://www.kaggle.com/unkownhihi/dfdc-lrcn-training\n@unkownhihi is it normal or is my train/test split issue ? can i trust training loss ?\n", "votes": 5 }, { "id": 779750, "postDate": "2020-03-19T16:42:01.357Z", "content": "@unkownhihi You motivated me to do best in this competition thanks. 👍
", "rawMarkdown": "@unkownhihi You motivated me to do best in this competition thanks. 👍 ", "votes": 1 }, { "id": 779647, "postDate": "2020-03-19T15:05:40.553Z", "content": "I never meet this before. When it started training, it will bounce around 0.6-0.7, but never went to this low before.
", "rawMarkdown": "I never meet this before. When it started training, it will bounce around 0.6-0.7, but never went to this low before.", "votes": 1, "replies": [ { "id": 779664, "postDate": "2020-03-19T15:23:16.703Z", "content": "Yes after i fix my input data, it reach between 0.4 to 0.5 Thanks @unkownhihi But i got val_loss 0.0063, i need to improve this.
\n\nwhen i try split based on folders, training going closer to zero again :(. is it means training samples are more similar ?
", "rawMarkdown": "Yes after i fix my input data, it reach between 0.4 to 0.5 Thanks @unkownhihi But i got val_loss 0.0063, i need to improve this.\n\nwhen i try split based on folders, training going closer to zero again :(. is it means training samples are more similar ?" }, { "id": 779711, "postDate": "2020-03-19T16:07:11.247Z", "content": "@seshurajup did you use my exact same model, or you used your custom model?
", "rawMarkdown": "@seshurajup did you use my exact same model, or you used your custom model?" }, { "id": 779716, "postDate": "2020-03-19T16:10:52.447Z", "content": "@unkownhihi same model with 0 to 39 as training and 40 to 49 as validation
", "rawMarkdown": "@unkownhihi same model with 0 to 39 as training and 40 to 49 as validation" }, { "id": 779728, "postDate": "2020-03-19T16:17:48.747Z", "content": "well, first of all, I do not recommend copying the exact same model. Because 0.38 is its limit.\nSecond, it shouldn't happen. Check your code, data, etc. There's a slight possibility that your model is actually performing soooooo good, that it can get to ~0 loss.
", "rawMarkdown": "well, first of all, I do not recommend copying the exact same model. Because 0.38 is its limit.\nSecond, it shouldn't happen. Check your code, data, etc. There's a slight possibility that your model is actually performing soooooo good, that it can get to ~0 loss." }, { "id": 779730, "postDate": "2020-03-19T16:19:22.467Z", "content": "Thanks, i will cross check the issue.
", "rawMarkdown": "Thanks, i will cross check the issue.", "votes": 1 } ] }, { "id": 779308, "postDate": "2020-03-19T07:39:14.310Z", "content": "Loss and accuracy are too random at starting when we train the model. don't be panic, it will be stable after 65% of train samples.
\n\nI stopped multiple times to test is my train/split have issue. Thanks, maybe it will help others not to confuse with loss and accuracy
", "rawMarkdown": "Loss and accuracy are too random at starting when we train the model. don't be panic, it will be stable after 65% of train samples.\n\nI stopped multiple times to test is my train/split have issue. Thanks, maybe it will help others not to confuse with loss and accuracy", "votes": 2 }, { "id": 789260, "postDate": "2020-03-28T14:06:05.947Z", "content": "Had very low training loss as well. Found a bug and fixed it! \nNow, I still have low validation loss, will check my code more and maybe add data augmentation. \nThanks for this reassuring thread. Hopefully, I find and correct the bug before the end of the competition. ;)
", "rawMarkdown": "Had very low training loss as well. Found a bug and fixed it! \nNow, I still have low validation loss, will check my code more and maybe add data augmentation. \nThanks for this reassuring thread. Hopefully, I find and correct the bug before the end of the competition. ;)", "replies": [ { "id": 790557, "postDate": "2020-03-29T17:04:00.593Z", "content": "Same issue with u. I'm trying to change my data creation now.
", "rawMarkdown": "Same issue with u. I'm trying to change my data creation now.", "votes": 1 }, { "id": 790563, "postDate": "2020-03-29T17:11:22.263Z", "content": "Good luck to you! I am trying to resample the data to avoid overfitting. I have few ideas to try but time is running low. :p
", "rawMarkdown": "Good luck to you! I am trying to resample the data to avoid overfitting. I have few ideas to try but time is running low. :p " } ] }, { "id": 789707, "postDate": "2020-03-28T23:08:16.800Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 779750, "author_name": "SeshuRaju 🧘♂️", "author_url": "", "post_date": "2020-03-19T16:42:01.357000", "content": "@unkownhihi You motivated me to do best in this competition thanks. 👍
", "votes": 1, "replies": [] }, { "id": 779647, "author_name": "Shangqiu Li", "author_url": "", "post_date": "2020-03-19T15:05:40.553000", "content": "I never meet this before. When it started training, it will bounce around 0.6-0.7, but never went to this low before.
", "votes": 1, "replies": [ { "id": 779664, "author_name": "SeshuRaju 🧘♂️", "author_url": "", "post_date": "2020-03-19T15:23:16.703000", "content": "Yes after i fix my input data, it reach between 0.4 to 0.5 Thanks @unkownhihi But i got val_loss 0.0063, i need to improve this.
\n\nwhen i try split based on folders, training going closer to zero again :(. is it means training samples are more similar ?
", "votes": 0, "replies": [] }, { "id": 779711, "author_name": "Shangqiu Li", "author_url": "", "post_date": "2020-03-19T16:07:11.247000", "content": "@seshurajup did you use my exact same model, or you used your custom model?
", "votes": 0, "replies": [] }, { "id": 779716, "author_name": "SeshuRaju 🧘♂️", "author_url": "", "post_date": "2020-03-19T16:10:52.447000", "content": "@unkownhihi same model with 0 to 39 as training and 40 to 49 as validation
", "votes": 0, "replies": [] }, { "id": 779728, "author_name": "Shangqiu Li", "author_url": "", "post_date": "2020-03-19T16:17:48.747000", "content": "well, first of all, I do not recommend copying the exact same model. Because 0.38 is its limit.\nSecond, it shouldn't happen. Check your code, data, etc. There's a slight possibility that your model is actually performing soooooo good, that it can get to ~0 loss.
", "votes": 0, "replies": [] }, { "id": 779730, "author_name": "SeshuRaju 🧘♂️", "author_url": "", "post_date": "2020-03-19T16:19:22.467000", "content": "Thanks, i will cross check the issue.
", "votes": 1, "replies": [] } ] }, { "id": 779308, "author_name": "SeshuRaju 🧘♂️", "author_url": "", "post_date": "2020-03-19T07:39:14.310000", "content": "Loss and accuracy are too random at starting when we train the model. don't be panic, it will be stable after 65% of train samples.
\n\nI stopped multiple times to test is my train/split have issue. Thanks, maybe it will help others not to confuse with loss and accuracy
", "votes": 2, "replies": [] }, { "id": 789260, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2020-03-28T14:06:05.947000", "content": "Had very low training loss as well. Found a bug and fixed it! \nNow, I still have low validation loss, will check my code more and maybe add data augmentation. \nThanks for this reassuring thread. Hopefully, I find and correct the bug before the end of the competition. ;)
", "votes": 0, "replies": [ { "id": 790557, "author_name": "Liang Su", "author_url": "", "post_date": "2020-03-29T17:04:00.593000", "content": "Same issue with u. I'm trying to change my data creation now.
", "votes": 1, "replies": [] }, { "id": 790563, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2020-03-29T17:11:22.263000", "content": "Good luck to you! I am trying to resample the data to avoid overfitting. I have few ideas to try but time is running low. :p
", "votes": 0, "replies": [] } ] }, { "id": 789707, "author_name": "", "author_url": "", "post_date": "2020-03-28T23:08:16.800000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "779278": "\n\n\nTraining with https://www.kaggle.com/unkownhihi/dfdc-lrcn-training\n@unkownhihi is it normal or is my train/test split issue ? can i trust training loss ?\n", "779750": "@unkownhihi You motivated me to do best in this competition thanks. 👍 ", "779647": "I never meet this before. When it started training, it will bounce around 0.6-0.7, but never went to this low before.", "779308": "Loss and accuracy are too random at starting when we train the model. don't be panic, it will be stable after 65% of train samples.\n\nI stopped multiple times to test is my train/split have issue. Thanks, maybe it will help others not to confuse with loss and accuracy", "789260": "Had very low training loss as well. Found a bug and fixed it! \nNow, I still have low validation loss, will check my code more and maybe add data augmentation. \nThanks for this reassuring thread. Hopefully, I find and correct the bug before the end of the competition. ;)", "789707": "" } }