nerv - Lua-based toolkit for high-performance deep learning

	Commit message (Collapse)	Author	Age
...
*	applying dropout on lstm.h before combinerL seems bad, PPL only 95, trying ↵	txh18	2015-12-03
\| \| \| \|	another way(like in the paper) of dropout for lstm.h after combinerL
*	added two layers to lstmlm_ptb, todo:test it	txh18	2015-12-03
\|
*	added al_sen_start stat for lmseqreader	txh18	2015-12-03
\|
*	moved tnn to main nerv dir and added it to Makefile	txh18	2015-12-03
\|
*	...	txh18	2015-12-03
\|
*	small bug fix in tnn for se_mode, todo:test it	txh18	2015-12-02
\|
*	added se_mode for lmseqreader, todo:check it	txh18	2015-12-02
\|
*	function name change in LMTrainer	txh18	2015-12-02
\|
*	added dropout_t layer	txh18	2015-12-02
\|
*	changed thres_mask function of matrix to a more standard api	txh18	2015-12-02
\|
*	got PPL115 for ptb on h300lr1bat10wc1e-4	txh18	2015-12-01
\|
*	bug fix for lstm_t layer, t not inclueded in propagate!	txh18	2015-11-30
\|
*	small opt for initing tnn:clip_t	txh18	2015-11-30
\|
*	added ooutputGate for lstm_t	txh18	2015-11-30
\|
*	added clip_t for tnn	txh18	2015-11-27
\|
*	lstm_tnn can be run, todo:testing	txh18	2015-11-27
\|
*	still working..	txh18	2015-11-26
\|
*	working on lstm	txh18	2015-11-26
\|
*	changed auto-generating params, won not save in global_conf.param	txh18	2015-11-25
\|
*	let affine supported multiple inputs	txh18	2015-11-24
\|
*	added wcost for biasparam in lm_trainer	txh18	2015-11-24
\|
*	still working on dagL_T	txh18	2015-11-24
\|
*	completed layerdag_t, now testing...	txh18	2015-11-23
\|
*	completed gate_fff layer	txh18	2015-11-23
\|
*	implementing GateFFF layer	txh18	2015-11-23
\|
*	complete auto-generate params	txh18	2015-11-20
\|
*	working on automatic parameter for layers	txh18	2015-11-20
\|
*	changed work_dir setting	txh18	2015-11-18
\|
*	h300 and h400 worked well, log added	txh18	2015-11-18
\|
*	switch to kernel update	txh18	2015-11-17
\|
*	bug fix for select_linear layer-by-layer update	txh18	2015-11-17
\|
*	added atomicAdd for select_linear update, however, the result still seems ↵	txh18	2015-11-17
\| \| \| \|	unreproducable, I changed select_linear layer update back to line-by-line
*	added small opt: use mmatrix in lm_trainer and reader	txh18	2015-11-17
\|
*	coding style change	txh18	2015-11-17
\|
*	added LOG-tnn-h400 LOG	txh18	2015-11-16
\|
*	unified param updates, now direct_update is the same speed with undirect_update	txh18	2015-11-16
\|
*	...	txh18	2015-11-16
\|
*	...	txh18	2015-11-16
\|
*	used os.clock() for timer	txh18	2015-11-16
\|
*	fixed direct update, did not know the result	txh18	2015-11-16
\|
*	added timer	txh18	2015-11-15
\|
*	merge lr schedule change	txh18	2015-11-15
\|\ \| \| \| \| \| \|	Merge branch 'txh18/rnnlm' of github.com:Nerv-SJTU/nerv into txh18/rnnlm
\| *	got good PPL for H400...	cloudygoose	2015-11-15
\| \|
* \|	added msr_sc set	txh18	2015-11-15
\|/
*	small bug: lr_half	txh18	2015-11-13
\|
*	...	txh18	2015-11-13
\|
*	added random seed	txh18	2015-11-13
\|
*	added loadstring	txh18	2015-11-13
\|
*	saving param file for every iter	txh18	2015-11-13
\|
*	change ppl_net to ppl_all	txh18	2015-11-13
\|