Fix gradient of scan when some outputs are disconnected

theano/scan_module/scan_op.py

@@ -1720,9 +1720,12 @@ class Scan(PureOp):
 
         offset = self.n_mit_mot
         for idx in xrange(self.n_mit_sot):
+            if isinstance(dC_douts[idx + offset].type, DisconnectedType):
+                outer_inp_mitmot.append(outs[idx + offset].zeros_like())
+            else:
+                outer_inp_mitmot.append(dC_douts[idx + offset][::-1])
             mitmot_inp_taps.append([])
             mitmot_out_taps.append([])
-            outer_inp_mitmot.append(dC_douts[idx + offset][::-1])
             idx_tap = idx + self.n_mit_mot
             inner_inp_mitmot.append(dC_dXts[out_pos])
             out_pos += 1

theano/scan_module/tests/test_scan.py

@@ -1852,6 +1852,80 @@ class T_Scan(unittest.TestCase):
         analytic_grad = reset_rng_grad_fn(v_u, v_x0, vW_in)
         utt.assert_allclose(analytic_grad[0][:2], numpy.zeros((2, 2)))
 
+    def test_grad_multiple_outs_some_disconnected(self):
+        # Created on Tue Oct 07 13:28:51 2014
+        # @author: vaneetke
+        rng = numpy.random.RandomState(utt.fetch_seed())
+        n_hid = 3
+        n_in = 1
+        n_out = 1
+
+        W_hh_v = asarrayX(rng.uniform(size=(n_hid, n_hid), low=-.01, high=.01))
+        h0_v = asarrayX(rng.uniform(size=(2, n_hid), low=-.01, high=.01))
+        b_h_v = asarrayX(rng.uniform(size=(n_hid), low=-.01, high=.01))
+        W_ih_v = asarrayX(rng.uniform(size=(n_in, n_hid), low=-.01, high=.01))
+        W_ho_v = asarrayX(rng.uniform(size=(n_hid, n_out), low=-.01, high=.01))
+        b_o_v = asarrayX(rng.uniform(size=(n_out), low=-.01, high=.01))
+
+        # parameters of the rnn
+        b_h = theano.shared(b_h_v)
+        h0 = theano.shared(h0_v)
+        W_ih = theano.shared(W_ih_v)
+        W_hh = theano.shared(W_hh_v)
+        W_ho = theano.shared(W_ho_v)
+        b_o = theano.shared(b_o_v)
+        params = [W_ih, W_hh, b_h, W_ho, b_o, h0]
+
+        # first dimension is time
+        x = tensor.matrix()
+
+        # sequences: x_t
+        # prior results: h_tm2, h_tm1
+        # non-sequences: W_ih, W_hh, W_ho, b_h
+        def one_step(x_t, h_tm2, h_tm1, W_ih, W_hh, b_h, W_ho, b_o):
+            h_t = tensor.tanh(theano.dot(x_t, W_ih)
+                              + theano.dot(h_tm2, W_hh) + b_h)
+            y_t = theano.dot(h_t, W_ho) + b_o
+            return [h_t, y_t]
+
+        # hidden and outputs of the entire sequence
+        [h, y], _ = theano.scan(
+            fn=one_step,
+            sequences = dict(input=x),
+            # corresponds to the return type of one_step
+            outputs_info = [dict(initial=h0, taps=[-2, -1]), None],
+            non_sequences = [W_ih, W_hh, b_h, W_ho, b_o])
+
+        # target values
+        t = tensor.matrix()
+
+        # learning rate
+        lr = asarrayX(0.1)
+        learning_rate = theano.shared(lr)
+
+        cost = ((0.5 * ((y - t) ** 2.0).mean())
+                + (0.5 * (y.std() - t.std()) ** 2.0))
+
+        gparams = T.grad(cost, params)
+        updates = [(param, param - gparam * learning_rate)
+                   for param, gparam in zip(params, gparams)]
+        learn_rnn_fn = theano.function(inputs=[x, t],
+                                       outputs=cost,
+                                       updates=updates)
+        eval_rnn_fn = theano.function(inputs=[x],
+                                      outputs=y)
+
+        # artificial data
+        x_v = numpy.arange(0., 100., 0.21, dtype=theano.config.floatX)
+        x_v = x_v.reshape(len(x_v), 1)
+        s_v = numpy.sin(x_v)
+        t_v = numpy.roll(s_v, -1)[:-1]
+        s_v = s_v[:-1]
+        for i in xrange(100):
+            cost = learn_rnn_fn(s_v, t_v)
+        pred = eval_rnn_fn(s_v)
+        assert cost < 0.02
+
     def test_draw_as_input_to_scan(self):
         trng = theano.tensor.shared_randomstreams.RandomStreams(123)