Merge pull request #5336 from lamblin/fix_scan_bcastable

theano/scan_module/scan_op.py

@@ -773,8 +773,11 @@ class Scan(PureOp):
                         # function exectution. Also, since an update is
                         # defined, a default value must also be (this is
                         # verified by DebugMode). Use an array of size 0 but
-                        # the right ndim and dtype.
-                        default_val = numpy.zeros([0] * inp.ndim,
+                        # the right ndim and dtype (use a shape of 1 on
+                        # broadcastable dimensions, 0 on the others).
+                        default_shape = [1 if _b else 0
+                                         for _b in inp.broadcastable]
+                        default_val = numpy.zeros(default_shape,
                                                   dtype=inp.dtype)
                         wrapped_inp = In(variable=inp, value=default_val,
                                          update=self.outputs[output_idx])

theano/scan_module/tests/test_scan.py

@@ -5466,3 +5466,29 @@ def test_outputs_taps_check():
     outputs_info = {'initial': y, 'taps': [-1, -1]}
     assert_raises(ValueError, theano.scan, f, x, outputs_info)
     print('done')
+
+
+def test_default_value_broadcasted():
+    def floatx(X):
+        return numpy.asarray(X, dtype=theano.config.floatX)
+
+    def init_weights(shape, name):
+        return theano.shared(floatx(numpy.random.randn(*shape) * 0.1), name)
+
+    X = theano.tensor.matrix('X')
+    in_size = 2
+    out_size = 4
+    W_x = init_weights((in_size, out_size), "W_x")
+
+    def _active(x, pre_h):
+        x = theano.tensor.reshape(x, (1, in_size))
+        pre_h = theano.tensor.dot(x, W_x)
+        return pre_h
+
+    value, scan_updates = theano.scan(_active, sequences=X,
+                                      outputs_info=[theano.tensor.alloc(floatx(0.), 1, out_size)])
+    cost = theano.tensor.mean(value)
+    gW_x = theano.tensor.grad(cost, W_x)
+    updates = [(W_x, W_x - 0.1 * gW_x)]
+    f = theano.function([X], outputs=cost, updates=updates)
+    f(numpy.random.rand(10, in_size).astype(X.dtype))