Add tests for grad shape.

theano/sandbox/cuda/blocksparse.py

@@ -69,7 +69,7 @@ def bptr(a):
                                   dtype=cublas.ctypes.c_void_p)
 
 
-class SparseBlockGemvDS(GpuOp):
+class SparseBlockGemvSS(GpuOp):
     def __init__(self, inplace):
         self.inplace = inplace
         if self.inplace:
@@ -112,9 +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().dimshuffle((1, 0, 3, 2)),
-                                      go, h, outputIdx, inputIdx)
-        hgrad = sparse_block_gemv_ds(h.zeros_like(), W.dimshuffle((1, 0, 3, 2)),
+        Wgrad = sparse_block_outer_ss(W.zeros_like(),
+                                      h, go, inputIdx, outputIdx)
+        hgrad = sparse_block_gemv_ss(h.zeros_like(),
+                                     W.dimshuffle((1, 0, 3, 2)),
                                      go,
                                      outputIdx, inputIdx)
         return [go, Wgrad, hgrad,
@@ -124,7 +125,7 @@ class SparseBlockGemvDS(GpuOp):
                                "grad of outputIdx makes no sense")]
 
 
-sparse_block_gemv_ds = SparseBlockGemvDS(False)
+sparse_block_gemv_ss = SparseBlockGemvSS(False)
 
 
 class SparseBlockOuterSS(GpuOp):
@@ -145,10 +146,10 @@ class SparseBlockOuterSS(GpuOp):
         if not self.inplace:
             o = o.copy()
 
-        for i in range(x.shape[0]):
-            inp_id = xIdx[i]
-            for j in range(y.shape[0]):
-                out_id = yIdx[j]
+        for j in range(y.shape[0]):
+            out_id = yIdx[j]
+            for i in range(x.shape[0]):
+                inp_id = xIdx[i]
                 ger(numpy.float32(1.0), x[i],
                     y[j], o[inp_id, out_id])
 
@@ -161,7 +162,7 @@ sparse_block_outer_ss = SparseBlockOuterSS()
 # All code above this line is unused (except for the imports)
 
 
-def sparse_block_dot_DS(W, h, inputIdx, b, outputIdx):
+def sparse_block_dot_SS(W, h, inputIdx, b, outputIdx):
     """
     var: shape, comment
     W: (iBlocks, oBlocks, iSize, oSize), weight matrix
@@ -171,7 +172,7 @@ def sparse_block_dot_DS(W, h, inputIdx, b, outputIdx):
     outputIdx: (oWin,), indexes of the output blocks
 
     returns (oBlocks, oSize), dot(W[i, j], h[i]) + b[j]
-        but b[j] is only added once
+         but b[j] is only added once
     """
     o = b.take(outputIdx, axis=0)
     def outer_fn(out_id, W, h, b, iIdx):

theano/sandbox/cuda/tests/test_blocksparse.py

@@ -5,8 +5,8 @@ import theano.tests.unittest_tools as utt
 import numpy
 from numpy.random import randn
 
-from theano.sandbox.cuda.blocksparse import (sparse_block_dot_DS,
-                                             sparse_block_gemv_ds)
+from theano.sandbox.cuda.blocksparse import (sparse_block_dot_SS,
+                                             sparse_block_gemv_ss)
 
 def blocksparse_data():
     nInputBlock = 128
@@ -45,7 +45,7 @@ def test_blocksparse():
     iIdx = tensor.lvector()
     oIdx = tensor.lvector()
 
-    o = sparse_block_dot_DS(W, h, iIdx, b, oIdx)
+    o = sparse_block_dot_SS(W, h, iIdx, b, oIdx)
 
     f = theano.function([W, h, iIdx, b, oIdx], o)
 
@@ -64,7 +64,7 @@ def test_blocksparse_op():
     iIdx = tensor.lvector()
     oIdx = tensor.lvector()
 
-    o = sparse_block_gemv_ds(b.take(oIdx, axis=0), W, h, iIdx, oIdx)
+    o = sparse_block_gemv_ss(b.take(oIdx, axis=0), W, h, iIdx, oIdx)
 
     f = theano.function([W, h, iIdx, b, oIdx], o)
 
@@ -80,16 +80,16 @@ def test_blocksparse_op_grad():
     h_val = randn(2, 3).astype('float32')
     iIdx_val = numpy.random.permutation(3)[:2]
     oIdx_val = numpy.random.permutation(3)[:2]
-    W_val = randn(3, 3, 3, 3).astype('float32')
-    b_val = randn(3, 3).astype('float32')
+    W_val = randn(3, 3, 4, 3).astype('float32')
+    b_val = randn(3, 4).astype('float32')
 
     iIdx = theano.tensor.constant(iIdx_val)
     oIdx = theano.tensor.constant(oIdx_val)
 
-    def f(b, W, h):
-        return sparse_block_gemv_ds(b.take(oIdx, axis=0), W, h, iIdx, oIdx)
+    def f(b, h, W):
+        return sparse_block_gemv_ss(b.take(oIdx, axis=0), W, h, iIdx, oIdx)
 
-    utt.verify_grad(f, [b_val, W_val, h_val])
+    utt.verify_grad(f, [b_val, h_val, W_val])
 
 
 def test_blocksparse_op_grad2():
@@ -99,7 +99,7 @@ def test_blocksparse_op_grad2():
     iIdx = tensor.lvector()
     oIdx = tensor.lvector()
 
-    o = sparse_block_gemv_ds(b.take(oIdx, axis=0), W, h, iIdx, oIdx)
+    o = sparse_block_gemv_ss(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)
@@ -107,4 +107,8 @@ def test_blocksparse_op_grad2():
     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)
+    b_g, W_g, h_g = f(W_val, h_val, iIdx_val, b_val, oIdx_val)
+
+    assert b_g.shape == b_val.shape
+    assert h_g.shape == h_val.shape
+    assert W_g.shape == W_val.shape