Fix bug introduced in #863

resize the result matrix before back-propagation

Standard way to represent inf

For compatibility with MS Visual Studio.

Speed up SoftmaxComponent::Backprop()

yesno recipe data prep: paths to absolute + misc. 

New: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 16, speed was 0.0165568 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 16, speed was 0.00355242 gigaflops.
New: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 32, speed was 0.0678791 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 32, speed was 0.0145515 gigaflops.
New: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 64, speed was 0.24739 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 64, speed was 0.0583246 gigaflops.
New: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 128, speed was 0.898427 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 128, speed was 0.225076 gigaflops.
New: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 256, speed was 2.89009 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 256, speed was 0.834096 gigaflops.
New: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 512, speed was 6.72164 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 512, speed was 1.92722 gigaflops.
New: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 1024, speed was 10.4916 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<float>, for dim = 1024, speed was 2.78281 gigaflops.
New: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 16, speed was 0.0148584 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 16, speed was 0.00260567 gigaflops.
New: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 32, speed was 0.0586865 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 32, speed was 0.0121077 gigaflops.
New: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 64, speed was 0.22893 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 64, speed was 0.0527767 gigaflops.
New: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 128, speed was 0.763462 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 128, speed was 0.175736 gigaflops.
New: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 256, speed was 2.40457 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 256, speed was 0.58351 gigaflops.
New: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 512, speed was 4.55165 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 512, speed was 1.42464 gigaflops.
New: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 1024, speed was 4.36421 gigaflops.
Old: For CuMatrix::DiffSoftmaxPerRow<double>, for dim = 1024, speed was 1.94971 gigaflops.

Modify error message and documentation to suggest nvidia-smi -c 3 (process exclusive mode) as thread exclusive mode is deprecated now

fix bug in determinize-star-inl.h that affected goodness of hashing function (thanks: EnochLiou); fix problem in lattice-oracle error message found by Nicolas Serrano.

Fixed pitch_functions to behave correctly with snip_edges=false

 * Fixed snip_edges=false behaviour to have an expected feature delay of frame_shift/2 instead of being identical to the delay of snip_edges=true. This makes the behaviour of snip_edges=false consistent with other feature extractions tools.
 * Added a unit test that ensure snip_edges=false behaves as expected, by calculating the actual delay between outputs only differing by snip_edge through cross-correlation.
 * Refactored ExtractFrame to treat both edge cases symmetrically
 * Calculating window size / shift by dividing by 1000 instead of multiplying by 0.001 to fix unexpected behaviour with non integer window size / shift.
 * Made sure all changes pass google code validation
 * Changed all feature window functions to use precalculated factor: in my opinion it make the functions easier to compare, and it may be a bit faster.

Note that the feature delay of snip_edges=true (the default) is rather difficult to predict (and almost certainly not what people assume), which almost certainly makes the use of pitch features problematic at the moment.
I would suggest switching to snip_edges=false for new recipes.

Add option to skip nnet3 decode diagnostics

fix the order >=2 process of bigger files

fix per Dan's instructions

sorry forgot to save before commit
fix per Dan's instructions

another typo fixed

fixes #849

Fast group pnorm code path for p=0,1,2,inf

adding compile support for cuda8

Fix duplicate option registers

Add missing dependency to Makefiles

For compatbility with HTK USEPOWER option have added --use-power opti…

Add some diagnostic scripts to analyze initial/final silence lengths and related things; call them from alignment and decoding scripts.

fix bug make_segmentation_graph and make_utterance_graph

thchs30-add-real-spkinfo