aboutsummaryrefslogtreecommitdiff
Commit message (Collapse)AuthorAge
* use consistent update calc; clean up code; no need for `direct_update`Determinant2015-11-21
|
* Merge pull request #12 from cloudygoose/txh18/rnnlmTed Yin2015-11-18
|\ | | | | add atomicAdd for cukernel
| * small coding style changetxh182015-11-18
| |
| * h300 and h400 worked well, log addedtxh182015-11-18
| |
| * switch to kernel updatetxh182015-11-17
| |
| * bug fix for select_linear layer-by-layer updatetxh182015-11-17
| |
| * added atomicAdd for select_linear update, however, the result still seems ↵txh182015-11-17
| | | | | | | | unreproducable, I changed select_linear layer update back to line-by-line
| * using atomicAdd for select_linear updatetxh182015-11-17
| |
| * added small opt: use mmatrix in lm_trainer and readertxh182015-11-17
| |
| * coding style changetxh182015-11-17
| |
| * added LOG-tnn-h400 LOGtxh182015-11-16
| |
* | Merge pull request #10 from cloudygoose/txh18/rnnlmTed Yin2015-11-16
|\| | | | | add optimization for parameter update
| * change updateEI to update_by_err_inputtxh182015-11-16
| |
| * coding style changestxh182015-11-16
| |
| * changed the updates for affine layer, now just use update will be okaytxh182015-11-16
| |\ | | | | | | | | | Merge branch 'txh18/rnnlm'
| | * unified param updates, now direct_update is the same speed with undirect_updatetxh182015-11-16
| | |
| | * ...txh182015-11-16
| | |
| | * ...txh182015-11-16
| | |
| | * used os.clock() for timertxh182015-11-16
| | |
| | * fixed direct update, did not know the resulttxh182015-11-16
| | |
| | * added timertxh182015-11-15
| | |
| | * merge lr schedule changetxh182015-11-15
| | |\ | | | | | | | | | | | | Merge branch 'txh18/rnnlm' of github.com:Nerv-SJTU/nerv into txh18/rnnlm
| | | * got good PPL for H400...cloudygoose2015-11-15
| | | |
| | * | added msr_sc settxh182015-11-15
| | |/
| | * small bug: lr_halftxh182015-11-13
| | |
| | * ...txh182015-11-13
| | |
| | * added random seedtxh182015-11-13
| | |
| | * added loadstringtxh182015-11-13
| | |
| | * saving param file for every itertxh182015-11-13
| | |
| | * change ppl_net to ppl_alltxh182015-11-13
| | |
| | * added wcost for select_linear layertxh182015-11-12
| | |
| | * cleaning files...txh182015-11-12
| | |
| | * get good PPL for h300, see m-tests/LOG-tnn-h300txh182015-11-12
| | |
| | * added a little debug info in readertxh182015-11-11
| | |
| | * got good result when batch_size=1, strange!txh182015-11-11
| | |
| | * bug fix : changed zero-filling across borderstxh182015-11-10
| | |
| | * small bug in lmseqreadertxh182015-11-09
| | |
| | * small bug about zeroing errortxh182015-11-09
| | |
| | * another bit error bug foundtxh182015-11-09
| | |
| | * serious bug about bit operation in tnntxh182015-11-08
| | |
| | * switched to softmax_ce_ttxh182015-11-08
| | |
| | * ih[0] nolonger zerotxh182015-11-08
| | |
| | * trying to test softmax_ce_ttxh182015-11-08
| | |
| | * meet a bug, don not know whytxh182015-11-07
| | |
| | * ready to test on ptbtxh182015-11-07
| | |
| | * writing the schedulertxh182015-11-06
| | |
| | * added cross-border flushtxh182015-11-06
| | |
| | * first small tnn test seems to worktxh182015-11-06
| | |
| | * working on TNN:net_propagatetxh182015-11-03
| | |
| | * still working on TNNtxh182015-11-03
| | |
Powered by cgit, Themed by Ted Yin.