fixed bench sweep

I've long suspected that there was smth wrong w backprop

Created on August 17|Last edited on August 18

Comment

﻿
does any backprop work?﻿
avg_accuracy
avg_accuracy
Computing group metrics from first 36 groups
500k1M1.5M2Miter00.20.40.60.81
updater: dfa   no_recur_lr3
updater: backprop   no_recur_lr3
updater: dfa   no_recur_lr4
updater: backprop   no_recur_lr4
 
Run set644
no_recur_lr3160
no_recur_lr4160
﻿
lr, plast_clip, updater
﻿
 
Run set644
no_recur_lr3160
 
no_recur_lr4160
﻿
is it rly  just about effective lr?
﻿
Run set644
 
no_recur_lr3160
 
no_recur_lr4160
﻿
updater vs. recurrence. We hope to see solid performance on both dfa and backprop even without recurrence. 
﻿
Run set644
﻿
is clip_weights useful at all, and if so, what value of it? 
﻿
 
Run set132
no_recur320
﻿
is residual_connections a useful parameter?
﻿
 
Run set132
no_recur64
﻿
﻿
﻿

Add a comment