2

I've a lot of values from all days over one entire year. I'm wanna verify if they have a kind of similarity for each month (verify if these days values correspond to the correct month and/or predict for future same months from another future year). From the https://github.com/NVIDIA/DIGITS/tree/master/examples/regression and the alternative method: manually creating LMDB files, what can I do? I've to put all these values from all days over one entire year into the val_db and train_db? After this, generate one test image for each month with all days from these months? Or I've to make a val_db and train_db for each month separately? Thank you.

Eduardo Andrade
  • 111
  • 1
  • 1
  • 13

1 Answers1

0

Honestly DIGITS might be a little overkill for you. You don't need much computation power to train a little logistic regression problem and the results should be pretty easy to interpret.

But, if you want to forge ahead with DIGITS, you'll have to

  1. Format your data as a CxHxW data blob (maybe just 1x1xN where N is the number of features you have)
  2. Format your label as a CxHxW data blob (maybe just 1x1x1)
  3. Create a Caffe model that works for simple logistic regression (maybe just a single inner-product layer?)

To get help with this, please post on our user group instead of here on SO (following the README).

Luke Yeager
  • 1,400
  • 1
  • 17
  • 30
  • I understand but needs to be from DIGITS. But about the questions? My problem is more in understanding the problem to generate the LMDB correct because it takes many hours (millions and millions of values). Let's suppose it's one feature problem. – Eduardo Andrade Oct 27 '16 at 17:01
  • I don't really have an opinion on how you should structure your data - that's more of a question for a data scientist than for me. Again, I'll refer you to our user group. – Luke Yeager Oct 27 '16 at 17:41
  • Ok, I've put at the user group too. – Eduardo Andrade Oct 27 '16 at 17:46