When training model, your model is good if Loss in training/validation is decreasing and Accuracy in training/validation is increasing.
\nI was training my model in twice.
\nThe firstly, the Learning Curse look like good.
\nSo, I found out the epoch with the best Validation value.
\nAnd finally, training again with the best epoch I have found.
\nBut the second Learning Curse is so zigzag.
\nIs is normal?
P/S: my kernel in here
", "messages": [ { "id": 1593839, "postDate": "2021-11-24T10:09:48.933Z", "content": "When training model, your model is good if Loss in training/validation is decreasing and Accuracy in training/validation is increasing.
\nI was training my model in twice.
\nThe firstly, the Learning Curse look like good.
\nSo, I found out the epoch with the best Validation value.
\nAnd finally, training again with the best epoch I have found.
\nBut the second Learning Curse is so zigzag.
\nIs is normal?
P/S: my kernel in here
", "rawMarkdown": "When training model, your model is good if Loss in training/validation is decreasing and Accuracy in training/validation is increasing.\n\nI was training my model in twice. \nThe firstly, the Learning Curse look like good. \nSo, I found out the epoch with the best Validation value. \nAnd finally, training again with the best epoch I have found. \nBut the second Learning Curse is so zigzag. \nIs is normal? \n\nP/S: my kernel in [here](https://www.kaggle.com/phanttan/cnn-keras-99-accuracy/)", "votes": 4 }, { "id": 1611598, "postDate": "2021-12-08T05:05:29.990Z", "content": "I find out my errors, thank you @meowmeowmeowmeowmeow , @haphuongdo
", "rawMarkdown": "I find out my errors, thank you @meowmeowmeowmeowmeow , @haphuongdo ", "votes": 1 }, { "id": 1593847, "postDate": "2021-11-24T10:19:09.657Z", "content": "Did you fix a random seed? Your weights are initialized randomly. That's why training results may differ from run to run. Other think to pay attention - data shuffling. It also depends on random.
\nIn general, it is a red flag, that your model is not stable in traning.
Is \"red flag\" mean the curse will go up and down, up and down?
", "rawMarkdown": "Is \"red flag\" mean the curse will go up and down, up and down?", "votes": 1 }, { "id": 1593922, "postDate": "2021-11-24T11:52:35.507Z", "content": "I mean that your training process is not stable and depends on random. Sometimes it way finishes successfully, sometimes - not. And even extreme case - imagine you have trained model successfully. And next 100 run training may be failed.
", "rawMarkdown": "I mean that your training process is not stable and depends on random. Sometimes it way finishes successfully, sometimes - not. And even extreme case - imagine you have trained model successfully. And next 100 run training may be failed.", "votes": 1 }, { "id": 1596228, "postDate": "2021-11-26T09:34:19.403Z", "content": "Thank for your suggestion @meowmeowmeowmeowmeow I am feel not good in \"reproduction\". I created a small model and ask in here.
\nI appreciate your time to visit in.
I find out my errors, thank you @meowmeowmeowmeowmeow , @haphuongdo
", "votes": 1, "replies": [] }, { "id": 1593847, "author_name": "Mykola", "author_url": "", "post_date": "2021-11-24T10:19:09.657000", "content": "Did you fix a random seed? Your weights are initialized randomly. That's why training results may differ from run to run. Other think to pay attention - data shuffling. It also depends on random.
\nIn general, it is a red flag, that your model is not stable in traning.
Is \"red flag\" mean the curse will go up and down, up and down?
", "votes": 1, "replies": [] }, { "id": 1593922, "author_name": "Mykola", "author_url": "", "post_date": "2021-11-24T11:52:35.507000", "content": "I mean that your training process is not stable and depends on random. Sometimes it way finishes successfully, sometimes - not. And even extreme case - imagine you have trained model successfully. And next 100 run training may be failed.
", "votes": 1, "replies": [] }, { "id": 1596228, "author_name": "Tan Phan", "author_url": "", "post_date": "2021-11-26T09:34:19.403000", "content": "Thank for your suggestion @meowmeowmeowmeowmeow I am feel not good in \"reproduction\". I created a small model and ask in here.
\nI appreciate your time to visit in.