#datascience #deeplearning #LSTM
Entire Time Series Course – https://www.youtube.com/playlist?list=PL3N9eeOlCrP5cK0QRQxeJd6GrQvhAtpBK
In this video we will see how we can build a multi variate time series model using Deep learning LSTM sequence model. We will see end to end time series model building process in this video
Video Rating: / 5
Sir, thank you so much for this great tutorial! just wanna ask if my data is in daily manner and sporadic, how would be length of my window? can it be just days eg. 1,2,7 days?
THAT IS ALL WRONG !
It is time serie you must :
.add stateful = true in LSTM layer
.remove the validation from the fitting and the testing must be back test testing (you must test one by one instance of test and adding the instance to training (LOOP with reset state for each epoch)
Thank you for a great video! I got a lot of help from this. Just one question if I ask,
If I want to add to calculate MAPE for training and validation set, how can I add it from there?
Hi can you help out on only time series (hh:MM) code from object to time
If the data have outliers so do I need to remove outliers or the stationary the time series ..
I have followed the path that you have mention in this video but I need some help if I will give the length on hourly moving data as 24*5=120 that is 5 days then while predictions it is not matching with the len of X_test …Means ln of prediction after reshaping (-1,1) getting beyond the limit..after getting the prediction its seems like batch so please give the conclusion on it …why is so happening but any solution over that please let me?
Do one on Bi-direction as well
Very nice sir, Can this code work on production? …I have seen splitting for the train and test is very complicated but in your case seems very easy …thanks and Also make videos on different splitting approaches on LSTM for x and y…One more thing while using scaler …I think there will be data leakage
in LSTM algorithm what should be the layers if problem is supervised classification.Target having labels 0 and 1 to predict .
Hello, I have a ques that how you took LSTM(128), in general case how to take this value and what if we found 'nan' value in train generator?
Thanks for the great video. Sir, I found the following error when I going to run the code, ( AttributeError: 'TimeseriesGenerator' object has no attribute 'shape' ) Could you kindly explain this issue? Thanks
Dear Srivatsan,
Thanks a lot for your video and notebook.
Could you kindly explain, why we cut first 720 (length in generator) elements, but not last 720 elements ?
this one
df_pred=pd.concat([pd.DataFrame(predictions), pd.DataFrame(x_test[:,1:][win_length:])],axis=1)
instead
df_pred=pd.concat([pd.DataFrame(predictions), pd.DataFrame(x_test[:,1:][:-win_length])],axis=1)
Why do we haven't first 720 elemetns in prediction ?
When we created test_generator, there was first set of examples test_generator[0][0] from 32 batches by 720 elements, and I thought, that first prediction (prediction[0]) matches first element in first batch test_generator[0][0][0]. And we do not take last 720 elements after last batch, because we will not have target for last set of elements. Is it not True?
Thanks for the great video. The coding is user friendly. So we could think this as a supervised regression problem. But in terms of forecasting future, what actions should we take? What if we wanted to forecast (predict) the next 10 days?
@AIEngineering Hello and thanks again , if i have to predict a variable other than the first variable i will only have to change
?
target = data_scaled[:,0] — > target = data_scaled[:,3] and
pd.DataFrame(x_test[:,1:][win_length:] —> pd.DataFrame(x_test[:,3:][win_length:] and
df_final["predicted_value"]= rev_trans[:,0] — > df_final["predicted_value"]= rev_trans[:,3]
@AIEngineering excellent explanation. I really appreciate your knowledge. I manage to find your script in github and I success executed all your code without error using my own dataset. Now, Im still not clearly understand about the function of timeseriesgenerator. I want to make a forecast 3 hours ahead with interval is 15 minutes. Could you advise me how possible to do that? should we change the parameter of timeseries generator? thank you.
in LSTM model architecture you have used 1 dense layer to depicts one period forecast , if i want to return 3 periods values then should i have 3 dense layers?
@AIEngineering Hello and thanks for the amazing, I have a question regarding the last part. What changes should i do in order to predict a variable other than the first variable (a variable in middle for instance " visibility" ). appreciating your kind support
Just to confirm, what we are doing in this video is forecasting correct?
I am using tf version 1.10.1 and MeanSquaredError function is not present and throwing an error. Would be helpful to know if there is any work around?