[build,scripts] Fix to fast-lstm xconfig layer which was not function…

…al; cosmetic change to 'configure'.

egs/wsj/s5/steps/libs/nnet3/xconfig/lstm.py

@@ -637,19 +637,23 @@ def generate_lstm_config(self):
         configs.append("# See cu-math.h:ComputeLstmNonlinearity() for details.")
         configs.append("component name={0}.lstm_nonlin type=LstmNonlinearityComponent cell-dim={1} {2}".format(name, cell_dim, lstm_str))
         configs.append("# Component for backprop truncation, to avoid gradient blowup in long training examples.")
+        # Note from Dan: I don't remember why we are applying the backprop
+        # truncation on both c and m appended together, instead of just on c.
+        # Possibly there was some memory or speed or WER reason for it which I
+        # have forgotten about now.
         configs.append("component name={0}.cm_trunc type=BackpropTruncationComponent dim={1} {2}".format(name, 2 * cell_dim, bptrunc_str))
 
         configs.append("###  Nodes for the components above.")
         configs.append("component-node name={0}.four_parts component={0}.W_all input=Append({1}, "
-                       "IfDefined(Offset({0}.r_trunc, {2})))".format(name, input_descriptor, delay))
+                       "IfDefined(Offset({0}.c_trunc, {2})))".format(name, input_descriptor, delay))
         configs.append("component-node name={0}.lstm_nonlin component={0}.lstm_nonlin "
                        "input=Append({0}.four_parts, IfDefined(Offset({0}.c_trunc, {1})))".format(name, delay))
         # we can print .c later if needed, but it generates a warning since it's not used.  could use c_trunc instead
         #configs.append("dim-range-node name={0}.c input-node={0}.lstm_nonlin dim-offset=0 dim={1}".format(name, cell_dim))
         configs.append("dim-range-node name={0}.m input-node={0}.lstm_nonlin dim-offset={1} dim={1}".format(name, cell_dim))
         configs.append("component-node name={0}.cm_trunc component={0}.cm_trunc input={0}.lstm_nonlin".format(name))
         configs.append("dim-range-node name={0}.c_trunc input-node={0}.cm_trunc dim-offset=0 dim={1}".format(name, cell_dim))
-        configs.append("dim-range-node name={0}.m_trunc input-node={0}.cm_trunc dim-offset={1} dim={1}".format(name, cell_dim))
+        # configs.append("dim-range-node name={0}.m_trunc input-node={0}.cm_trunc dim-offset={1} dim={1}".format(name, cell_dim))
         configs.append("### End LTSM layer '{0}'".format(name))
         return configs
 

src/configure

@@ -63,7 +63,8 @@ Configuration options:
   --shared              Build and link against shared libraries [default=no]
   --use-cuda            Build with CUDA [default=yes]
   --cudatk-dir=DIR      CUDA toolkit directory
-  --double-precision    Build with double precision floats [default=no]
+  --double-precision    Build with BaseFloat set to double if yes [default=no],
+                        mostly useful for testing purposes.
   --static-fst          Build with static OpenFst libraries [default=no]
   --fst-root=DIR        OpenFst root directory [default=../tools/openfst/]
   --fst-version=STR     OpenFst version string