I have a time-series dataset with 1500 records. The dim of each record is 5000*3*1 and Target is a single value like True/False. The point is that each 3*1 dim input is a location (x,y,z) and 5000 is the number of elements in an area. So it's important to find relationships between these elements after a period of time. In order to solve this time-series problem, I am going to train a CNN/LSTM model but the computation cost of this model is high and I am looking for a solution to train this model efficiently. The output must be calculated based on the relationship between the position of these elements thus I can not reduce the number of elements.
Hi Omid,
I think your time series array of 5000*3*1 is able directly to be train with 2D-CNN without reshaping, however if you want to also explore the temporal features of your data then you can use 3D-CNN. For this I think you can reshape your data into something like (500*3*5000*3*1) where 3 is the data window, so now your data become 500 records of 3*5000*3*1.
Other option is you can combine 2D-CNN with LSTM in 1 networks so you also still can reserve the temporal features of your data.
I hope that's help!
In addition tothe answer of Tommy Sugiarto , maybe this can help you out:
Article Improvement in Land Cover and Crop Classification based on T...
hi, use for training GPU GeForce, it can very increase training efficiency. If you don't have it , you can use kaggle.com powers or take it on cloud.google.com, which give you 300$ per year for different services among them there are GPUs
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training