Fix shape errors in the gradient (but it's still not the right value).

theano/sandbox/cuda/blocksparse.py

@@ -50,15 +50,14 @@ def gemv(alpha, A, x, beta, y):
 
 
 def ger(alpha, x, y, A):
-    print A.shape, x.shape, y.shape
     assert A.shape[1] == x.shape[0]
     assert A.shape[0] == y.shape[0]
 
     handle = scikits.cuda.misc._global_cublas_handle
 
-    cublas.cublasSger(handle, A.shape[0], A.shape[1], alpha,
-                      x.gpudata, x.strides[0],
+    cublas.cublasSger(handle, A.shape[1], A.shape[0], alpha,
                       y.gpudata, y.strides[0],
+                      x.gpudata, x.strides[0],
                       A.gpudata, A.strides[0])
 
 
@@ -113,9 +112,10 @@ class SparseBlockGemvDS(GpuOp):
         go = grads[0]
 
         # might revise that interface to not have a huge output
-        Wgrad = sparse_block_outer_ss(W.zeros_like(), go, h,
-                                      outputIdx, inputIdx)
-        hgrad = sparse_block_gemv_ds(h.zeros_like(), W.T, go,
+        Wgrad = sparse_block_outer_ss(W.zeros_like().dimshuffle((1, 0, 3, 2)),
+                                      go, h, outputIdx, inputIdx)
+        hgrad = sparse_block_gemv_ds(h.zeros_like(), W.dimshuffle((1, 0, 3, 2)),
+                                     go,
                                      outputIdx, inputIdx)
         return [go, Wgrad, hgrad,
                 grad_undefined(self, 3, inputIdx,
@@ -146,9 +146,9 @@ class SparseBlockOuterSS(GpuOp):
             o = o.copy()
 
         for i in range(x.shape[0]):
-            out_id = xIdx[i]
+            inp_id = xIdx[i]
             for j in range(y.shape[0]):
-                inp_id = yIdx[j]
+                out_id = yIdx[j]
                 ger(numpy.float32(1.0), x[i],
                     y[j], o[inp_id, out_id])
 

theano/sandbox/cuda/tests/test_blocksparse.py

@@ -90,3 +90,21 @@ def test_blocksparse_op_grad():
         return sparse_block_gemv_ds(b.take(oIdx, axis=0), W, h, iIdx, oIdx)
 
     utt.verify_grad(f, [b_val, W_val, h_val])
+
+
+def test_blocksparse_op_grad2():
+    b = tensor.fmatrix()
+    W = tensor.ftensor4()
+    h = tensor.fmatrix()
+    iIdx = tensor.lvector()
+    oIdx = tensor.lvector()
+
+    o = sparse_block_gemv_ds(b.take(oIdx, axis=0), W, h, iIdx, oIdx)
+    go = theano.grad(o.sum(), [b, W, h])
+
+    f = theano.function([W, h, iIdx, b, oIdx], go)
+
+    W_val, h_val, iIdx_val, b_val, oIdx_val = blocksparse_data()
+
+    # just make sure that it runs correcly and all the shapes are ok.
+    f(W_val, h_val, iIdx_val, b_val, oIdx_val)