I upgraded to ENVI 5.5.3/Deep Learning 1.1 from ENVI 5.5.2/Deep Learning 1.0 on June 7.
I have been unable to reproduce results that I had achieved with DL 1.0 with DL 1.1. My single class results under DL 1.0 were very good but using the same* parameters as I did with the DL 1.0 model, I am unable to obtain a classification. The entire test area is now classified or unclassified.
After a lot of trial and error, I obtained a single class result, but only by increasing the amount of training data by roughly 20%. However, the quality of classification that was achieved is not nearly as good as I was achieving with my 1.0 models.
I am aware that the Deep Learning framework has changed considerably, but even so, I would expect to be able to reproduce my previous results at least approximately.
Are there any changes in procedure or adjustments that I need to perform in order to reproduce DL 1.0 results in DL 1.1?
*The patch size pulldown no longer has the same options, so my initial model is slightly different from what I was using previously.
Currently Using ENVI 5.5.3 Build 379705/IDL 8.7 with Deep Learning Module 1.1
Upgraded from ENVI 5.5.2/Deep Learning Module 1.0
Windows 10 Home 64-bit OS.
Intel Core i7-8750H CPU, 16 GB RAM
Intel UHD Graphics 630 with NVIDIA GeForce GTX 1060 with Max-Q graphics cards