In running the ILSVRC12 example using Digits / NVCaffe I obtained very different results in a very important way, where in my execution I got at best an 80% accuracy while the example in the above tutorial shows a 98% BIG DIFFERENCE.
I tried to upload pictures but the NVIDIA forum does not have that ability!!!???
Also, I noticed that in the demo, it used 196,349 images in the train set, while in my run, following the tutorial I got 940,144 in my train set.
In my run, the model doesn’t change accuracy from epoch 3 to 30!
I guess some behind the scenes manipulation was done here, but I can’t understand why the difference.
Researching Keras for my project in the meantime