![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5130427%2F7fc6fc2e9a0a0ab766c8a3d1c5b8c7ba%2Faccuracy.png?generation=1610819117695551&alt=media\")
What can we say about our model looking at these?
\nIts clear that the accuracy has increased in both the cases for training, but for validation set it has been spiking a lot. Is there a way this can be prevented of this is supposed to be like this?
\nDoes it say that there was overfitting that happened?
\nAlso in the third image(I have marked a point) at around 14th epoch the val_accuracy was highest, should I have stopped the training there? because the trainining accuracy increased significantly afterwards.
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5130427%2Fe1d9e3ee9f6f73e4bf6a2a728dec70b3%2Floss.png?generation=1610819095343184&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5130427%2F63f913c4ae6ca8916b18339530aa00ca%2Faccuracy.png?generation=1610819528847624&alt=media\")
\nI hope all these things get answered so that I can get 90 percent accuracy in my next sub :)
What can we say about our model looking at these?
\nIts clear that the accuracy has increased in both the cases for training, but for validation set it has been spiking a lot. Is there a way this can be prevented of this is supposed to be like this?
\nDoes it say that there was overfitting that happened?
\nAlso in the third image(I have marked a point) at around 14th epoch the val_accuracy was highest, should I have stopped the training there? because the trainining accuracy increased significantly afterwards.
\n
\nI hope all these things get answered so that I can get 90 percent accuracy in my next sub :)
In one short question you have asked for a full semester course to be presented :)
\nYou will not achieve 90% on your next submission.
\nNormally we use the model with the best validation loss. So add early stopping to your code.
\nMy first step if I had these plots would be to change my learning rate. I like my spikes smaller. My next steps would be all the things in the full semester course :)
", "rawMarkdown": "In one short question you have asked for a full semester course to be presented :)\n\nYou will not achieve 90% on your next submission.\n\nNormally we use the model with the best validation loss. So add early stopping to your code.\n\nMy first step if I had these plots would be to change my learning rate. I like my spikes smaller. My next steps would be all the things in the full semester course :)", "votes": null }, { "id": "1156634", "postDate": "01/17/2021 09:45:44", "content": "That was a bit demotivational but alright.
\nI have always believed in learning by doing and I this idea of classroom learning doesnt fit me well! But thanks for the suggestions :)
Sorry - not my intention to demotivate. My intention was to indicate that the answer to your quest (achieve 90%) is detailed and complex. It's consists of many steps that need to be taken and many that can be tried.
\nReplace one semester of classroom training with 10 hours reading a couple of good books - 10 more hours studying the shared kernels. Etc
\nMy intent was not to tell you that you needed a semester course but rather that a single simple answer does not exist.
\nI have worked using data for over half a century. Pretty much was always using almost state of the art tools and software. Today I could solve problems using ML in a single day that took me over three years to solve in 1970. After retirement I spend my free hours on Kaggle in these competitions for the last several years. The last time I looked I was in the top 700 range for competitions and doing much better in the discussion rating (after 74 years I have learned to provide BS in huge volumes). I have 4 PC's running models for this competition almost 24/7 since it started. I don't yet have a 90% validation accuracy model.
\nAgain - not my intention to demotivate - but this is not a short simple path you have chosen to walk on.
", "rawMarkdown": "Sorry - not my intention to demotivate. My intention was to indicate that the answer to your quest (achieve 90%) is detailed and complex. It's consists of many steps that need to be taken and many that can be tried.\n\nReplace one semester of classroom training with 10 hours reading a couple of good books - 10 more hours studying the shared kernels. Etc \n\nMy intent was not to tell you that you needed a semester course but rather that a single simple answer does not exist. \n\nI have worked using data for over half a century. Pretty much was always using almost state of the art tools and software. Today I could solve problems using ML in a single day that took me over three years to solve in 1970. After retirement I spend my free hours on Kaggle in these competitions for the last several years. The last time I looked I was in the top 700 range for competitions and doing much better in the discussion rating (after 74 years I have learned to provide BS in huge volumes). I have 4 PC's running models for this competition almost 24/7 since it started. I don't yet have a 90% validation accuracy model. \n\nAgain - not my intention to demotivate - but this is not a short simple path you have chosen to walk on.", "votes": null }, { "id": "1156729", "postDate": "01/17/2021 11:21:13", "content": "Thanks for the reply.
\nYes I will read more and try to study kernels to have a better understanding.
\nMy question was not if I can achieve 90 percent in next submission but it was about drawing inferences from the diagrams. Because graphs always signify something and we can always learn something from them to improve :)
Its very inspiring that how much your knowledge must be about data as you have been doing it since 50 years. I am just 24 I cant even imagine having that much of experience.
\nAlso its inspiring that at 74 you are so enthusiastic about doing projects on kaggle.
\nWould love to connect with you on mail.(will drop a message from kaggle profile)
\nRespect for you
\nCheers
In one short question you have asked for a full semester course to be presented :)
\nYou will not achieve 90% on your next submission.
\nNormally we use the model with the best validation loss. So add early stopping to your code.
\nMy first step if I had these plots would be to change my learning rate. I like my spikes smaller. My next steps would be all the things in the full semester course :)
", "votes": null, "replies": [ { "id": 1156634, "author_name": "sarangbhatnagar", "author_url": "", "post_date": "01/17/2021 09:45:44", "content": "That was a bit demotivational but alright.
\nI have always believed in learning by doing and I this idea of classroom learning doesnt fit me well! But thanks for the suggestions :)
Sorry - not my intention to demotivate. My intention was to indicate that the answer to your quest (achieve 90%) is detailed and complex. It's consists of many steps that need to be taken and many that can be tried.
\nReplace one semester of classroom training with 10 hours reading a couple of good books - 10 more hours studying the shared kernels. Etc
\nMy intent was not to tell you that you needed a semester course but rather that a single simple answer does not exist.
\nI have worked using data for over half a century. Pretty much was always using almost state of the art tools and software. Today I could solve problems using ML in a single day that took me over three years to solve in 1970. After retirement I spend my free hours on Kaggle in these competitions for the last several years. The last time I looked I was in the top 700 range for competitions and doing much better in the discussion rating (after 74 years I have learned to provide BS in huge volumes). I have 4 PC's running models for this competition almost 24/7 since it started. I don't yet have a 90% validation accuracy model.
\nAgain - not my intention to demotivate - but this is not a short simple path you have chosen to walk on.
", "votes": null, "replies": [] }, { "id": 1156729, "author_name": "sarangbhatnagar", "author_url": "", "post_date": "01/17/2021 11:21:13", "content": "Thanks for the reply.
\nYes I will read more and try to study kernels to have a better understanding.
\nMy question was not if I can achieve 90 percent in next submission but it was about drawing inferences from the diagrams. Because graphs always signify something and we can always learn something from them to improve :)
Its very inspiring that how much your knowledge must be about data as you have been doing it since 50 years. I am just 24 I cant even imagine having that much of experience.
\nAlso its inspiring that at 74 you are so enthusiastic about doing projects on kaggle.
\nWould love to connect with you on mail.(will drop a message from kaggle profile)
\nRespect for you
\nCheers