This is a great series, Jeff. I enjoy hearing your thought process on building models and optimizing them. I'm looking forward to the videos on time series data. 😊
Question Jeff: if you're using df['xx']=zscore(df['xx']), how would you test on (new) oos_data without persisting the sigma and mean to apply them to the new data? Am I missing something?
Yes, for new data you would need to keep the zscore params from the old. I cover that in the last module when I discuss deploying models. You might have trained your model on 100K data items, but if your are scoring 1 at a time in real-time REST calls, you need to keep your old sdev/mean or just calling zscore alone will not work. Though the flip side to that is if I need to score a large batch of data say a year after I've fit the model, and I am worried about "drift", there can be times that it makes sense to use a current mean/sdev, but I do that very carefully.
Hi Jeff, I’ve pretty much gone through all your videos and material in GitHub to learn about deep neural networks. Thank you so much for sharing everything and you do an excellent job teaching! Question re this video, how do you do bootstrapping/benchmarking on time series data (since you wouldn’t want to shuffle in that case)? Thanks again, Rami
Thank you so much of this.
Can't believe I am watching for free.
Love from Nepal.
This is a great series, Jeff. I enjoy hearing your thought process on building models and optimizing them. I'm looking forward to the videos on time series data. 😊
Great series ever .thank yu jeff
Question Jeff: if you're using df['xx']=zscore(df['xx']), how would you test on (new) oos_data without persisting the sigma and mean to apply them to the new data? Am I missing something?
Yes, for new data you would need to keep the zscore params from the old. I cover that in the last module when I discuss deploying models. You might have trained your model on 100K data items, but if your are scoring 1 at a time in real-time REST calls, you need to keep your old sdev/mean or just calling zscore alone will not work. Though the flip side to that is if I need to score a large batch of data say a year after I've fit the model, and I am worried about "drift", there can be times that it makes sense to use a current mean/sdev, but I do that very carefully.
Hi Jeff,
I’ve pretty much gone through all your videos and material in GitHub to learn about deep neural networks. Thank you so much for sharing everything and you do an excellent job teaching! Question re this video, how do you do bootstrapping/benchmarking on time series data (since you wouldn’t want to shuffle in that case)?
Thanks again,
Rami
Wish i could give more likes
I know you don't know who i am sir but Jah bless my friend. Jah bless.