Add optimization to lift CPU CTC Op to GPU

theano/gpuarray/opt.py

@@ -34,6 +34,7 @@ from theano.tensor.nnet.abstract_conv import (BaseAbstractConv,
                                               AbstractConv3d_gradWeights,
                                               AbstractConv3d_gradInputs)
 from theano.tensor.nnet.neighbours import Images2Neibs
+from theano.tensor.nnet.ctc import ConnectionistTemporalClassification
 import theano.tensor.nlinalg as nlinalg
 import theano.tensor.signal.pool as pool
 import theano.tensor.slinalg as slinalg
@@ -78,6 +79,7 @@ from .linalg import (GpuCusolverSolve, MATRIX_STRUCTURES_SOLVE, GpuCholesky,
                      cusolver_available, GpuMagmaMatrixInverse, gpu_svd,
                      GpuMagmaCholesky, gpu_qr, GpuMagmaEigh)
 from .neighbours import GpuImages2Neibs
+from .ctc import GpuConnectionistTemporalClassification
 
 _logger = logging.getLogger("theano.gpuarray.opt")
 
@@ -2278,6 +2280,14 @@ def local_gpu_magma_svd(op, context_name, inputs, outputs):
             out = [out.astype('float16')]
     return out
 
+@register_opt('fast_compile')
+@op_lifter([theano.tensor.nnet.ctc.ConnectionistTemporalClassification])
+@register_opt2([theano.tensor.nnet.ctc.ConnectionistTemporalClassification], 'fast_compile')
+def local_gpu_ctc(op, context_name, inputs, outputs):
+    if not config.ctc.enabled:
+        return
+    return [GpuConnectionistTemporalClassification()(*node.inputs)]
+
 # Do not register in fast_run or fast_compile.
 # It will be added to fast_run if the GPU is enabled.
 optdb.register('gpua_scanOp_make_inplace',