nitishsrivastava
PCD issue.
Hello,
-
I am trying to train a RBM with lot of natural images using PCD (CD-k works fine on the very same network but with higher k step value is quite time consuming on millions of images, so i would try PCD)
My RBM net looks like: "input_layer" (3072) -> "hidden1" (4096)
-
Unfortunate i am unable to figure out positive_phase_order and negative_phase_order params properly without having the net to explode like:
$ deep-trainer.py deep_rbm_layer1.pbtxt train-deep-layer1.pbtxt eval.pbtxt Step 50 T_sp: 0.412 T_E: 1050726.2034862 V_E: 12880016.4388040 Step 100 T_sp: 0.496 T_E: 1007395681.2582184 V_E: 8329255899.3403273 Step 150 T_sp: 0.505 T_E: 192380774933.6150818 V_E: 930038333717.6633301 Step 200 T_sp: 0.500 T_E: 63295473747067.3281250 V_E: 448969617844569.6875000
Can someone advise ?
