Better error message when importing cuda_convnet under new GPU backend

lasagne/layers/cuda_convnet.py

@@ -13,6 +13,18 @@
 from .pool import pool_output_length
 from ..utils import as_tuple
 
+# bail out if using Theano's new GPU backend
+try:
+ from theano import gpuarray
+except ImportError:
+ from theano.sandbox import gpuarray
+if gpuarray.pygpu_activated:
+ raise ImportError("cuda_convnet is not supported under Theano's new "
+ "GPU backend. Please either use the ordinary "
+ "convolutional and pooling layers (they are usually "
+ "faster), or use Theano's old GPU backend (available "
+ "up to Theano 0.9).") # pragma: no cover
+
 from theano.sandbox.cuda.basic_ops import gpu_contiguous
 from pylearn2.sandbox.cuda_convnet.filter_acts import FilterActs