Merge pull request #371 from pascanur/fixed_grad_of_grad_of_scan

theano/scan_module/__init__.py

@@ -30,11 +30,11 @@ functions: ``scan()``, ``map()``, ``reduce()``, ``foldl()``,
 
 
 __docformat__ = 'restructedtext en'
-__authors__ = ( "Razvan Pascanu "
-                "Frederic Bastien "
-                "James Bergstra "
-                "Pascal Lamblin "
-                "Arnaud Bergeron ")
+__authors__ = ("Razvan Pascanu "
+               "Frederic Bastien "
+               "James Bergstra "
+               "Pascal Lamblin "
+               "Arnaud Bergeron ")
 __copyright__ = "(c) 2010, Universite de Montreal"
 __contact__ = "Razvan Pascanu <r.pascanu@gmail>"
 

theano/scan_module/scan_opt.py

@@ -417,7 +417,8 @@ class ScanSaveMem(gof.Optimizer):
         # change the number of steps in that case. To do this we set
         # global_nsteps to None which is seen as a flag that nothing needs
         # to be done
-        if len(node.outputs) <= c_outs:
+        assert len(node.outputs) >= c_outs
+        if len(node.outputs) == c_outs:
             global_nsteps = {'real': -1, 'sym': []}
         else:
             global_nsteps = None
@@ -474,7 +475,7 @@ class ScanSaveMem(gof.Optimizer):
                         break
 
                     # 2.3.2 extract the begin/end of the first dimension
-                    if i > op.n_mit_mot:
+                    if i >= op.n_mit_mot:
                         try:
                             length = shape_of[out][0]
                         except KeyError:
@@ -650,7 +651,8 @@ class ScanSaveMem(gof.Optimizer):
                             tmp = tensor.as_tensor_variable(val)
                             initl = tensor.as_tensor_variable(init_l[i])
                             tmp = tensor.maximum(tmp, initl)
-                            tmp = pre_greedy_local_optimizer(list_opt_slice, tmp)
+                            tmp = pre_greedy_local_optimizer(list_opt_slice,
+                                                             tmp)
                             tmp = pre_constant_merge([tmp])[0]
                             nw_input = nw_inputs[offset + idx][:tmp]
 

theano/scan_module/scan_views.py

@@ -5,10 +5,10 @@ See scan.py for details on scan
 """
 
 __docformat__ = 'restructedtext en'
-__authors__ = ( "Razvan Pascanu "
-                "Frederic Bastien "
-                "James Bergstra "
-                "Pascal Lamblin "  )
+__authors__ = ("Razvan Pascanu "
+               "Frederic Bastien "
+               "James Bergstra "
+               "Pascal Lamblin ")
 __copyright__ = "(c) 2010, Universite de Montreal"
 __contact__ = "Razvan Pascanu <r.pascanu@gmail>"
 
@@ -27,13 +27,13 @@ _logger = logging.getLogger('theano.scan_module.scan_views')
 # The ``map`` view of Scan Op.
 
 
-def map( fn
-        , sequences
-        , non_sequences     = None
-        , truncate_gradient = -1
-        , go_backwards      = False
-        , mode              = None
-        , name              = None  ):
+def map(fn,
+        sequences,
+        non_sequences=None,
+        truncate_gradient=-1,
+        go_backwards=False,
+        mode=None,
+        name=None):
     """
     Similar behaviour as python's map.
 
@@ -58,24 +58,24 @@ def map( fn
 
     :param name: See ``scan``.
     """
-    return scan.scan( fn                 = fn
-                , sequences         = sequences
-                , outputs_info      = []
-                , non_sequences     = non_sequences
-                , truncate_gradient = truncate_gradient
-                , go_backwards      = go_backwards
-                , mode              = mode
-                , name              = name )
+    return scan.scan(fn=fn,
+                     sequences=sequences,
+                     outputs_info=[],
+                     non_sequences=non_sequences,
+                     truncate_gradient=truncate_gradient,
+                     go_backwards=go_backwards,
+                     mode=mode,
+                     name=name)
 
 
 # The ``reduce`` view of Scan Op.
-def reduce( fn
-           , sequences
-           , outputs_info
-           , non_sequences = None
-           , go_backwards  = False
-           , mode          = None
-           , name          = None ):
+def reduce(fn,
+           sequences,
+           outputs_info,
+           non_sequences=None,
+           go_backwards=False,
+           mode=None,
+           name=None):
     """
     Similar behaviour as python's reduce
 
@@ -101,27 +101,27 @@ def reduce( fn
 
     :param name: See ``scan``.
     """
-    rval = scan.scan(fn             = fn
-                , sequences         = sequences
-                , outputs_info      = outputs_info
-                , non_sequences     = non_sequences
-                , go_backwards      = go_backwards
-                , truncate_gradient = -1
-                , mode              = mode
-                , name              = name )
-    if isinstance(rval[0], (list,tuple)):
-        return [ x[-1] for x in rval[0]], rval[1]
+    rval = scan.scan(fn=fn,
+                     sequences=sequences,
+                     outputs_info=outputs_info,
+                     non_sequences=non_sequences,
+                     go_backwards=go_backwards,
+                     truncate_gradient=-1,
+                     mode=mode,
+                     name=name)
+    if isinstance(rval[0], (list, tuple)):
+        return [x[-1] for x in rval[0]], rval[1]
     else:
         return rval[0][-1], rval[1]
 
 
 # The ``foldl`` view of Scan Op.
-def foldl( fn
-          , sequences
-          , outputs_info
-          , non_sequences = None
-          , mode          = None
-          , name          = None  ):
+def foldl(fn,
+          sequences,
+          outputs_info,
+          non_sequences=None,
+          mode=None,
+          name=None):
     """
     Similar behaviour as haskell's foldl
 
@@ -143,22 +143,22 @@ def foldl( fn
 
     :param name: See ``scan``.
     """
-    return reduce( fn             = fn
-                  , sequences     = sequences
-                  , outputs_info  = outputs_info
-                  , non_sequences = non_sequences
-                  , go_backwards  = False
-                  , mode          = mode
-                  , name          = name )
+    return reduce(fn=fn,
+                  sequences=sequences,
+                  outputs_info=outputs_info,
+                  non_sequences=non_sequences,
+                  go_backwards=False,
+                  mode=mode,
+                  name=name)
 
 
 # The ``foldl`` view of Scan Op.
-def foldr( fn
-          , sequences
-          , outputs_info
-          , non_sequences = None
-          , mode          = None
-          , name          = None ):
+def foldr(fn,
+          sequences,
+          outputs_info,
+          non_sequences=None,
+          mode=None,
+          name=None):
     """
     Similar behaviour as haskell' foldr
 
@@ -180,10 +180,10 @@ def foldr( fn
 
     :param name: See ``scan``.
     """
-    return reduce( fn             = fn
-                  , sequences     = sequences
-                  , outputs_info  = outputs_info
-                  , non_sequences = non_sequences
-                  , go_backwards  = True
-                  , mode          = mode
-                  , name          = name )
+    return reduce(fn=fn,
+                  sequences=sequences,
+                  outputs_info=outputs_info,
+                  non_sequences=non_sequences,
+                  go_backwards=True,
+                  mode=mode,
+                  name=name)

theano/scan_module/tests/test_scan.py

@@ -2585,6 +2585,26 @@ class T_Scan(unittest.TestCase):
         tf = theano.function([c, x], dP)
         assert tf([1.0, 2.0, -3.0, 4.0], 2.0) == 38
 
+    def test_grad_of_grad_of_state(self):
+        # Example provided Michael Forbes
+        # This tests ensures that we can compute gradients through cost
+        # defines in terms of gradients of scan
+        c = theano.tensor.vector('c')
+        x = theano.tensor.scalar('x')
+        _max_coefficients_supported = 1000
+        full_range = theano.tensor.arange(_max_coefficients_supported)
+        components, updates = theano.scan(
+            fn=lambda coeff, power, free_var: coeff * (free_var ** power),
+            outputs_info=None,
+            sequences=[c, full_range],
+            non_sequences=x)
+        P = components.sum()
+        dP = theano.tensor.grad(P, x).sum()
+        ddP = theano.tensor.grad(dP, x)
+        tf = theano.function([c, x], ddP)
+        assert tf([1.0, 2.0, -3.0, 4.0], 2.0) == 42
+
+
     def test_return_steps(self):
         rng = numpy.random.RandomState(utt.fetch_seed())
         vW_in2 = asarrayX(rng.uniform(size = (2,), low = -5.,high = 5.))
@@ -2705,9 +2725,9 @@ class T_Scan(unittest.TestCase):
         grad_fn = theano.function([xinit, w], [gx,gw],
                                  allow_input_downcast = True)
         rng  = numpy.random.RandomState(utt.fetch_seed())
-        v_x  = numpy.array(rng.uniform(size=(5,2,3), low=-2., high=2.),
+        v_x  = numpy.array(rng.uniform(size=(5,2,3), low=-3., high=3.),
                            dtype=theano.config.floatX)
-        v_w = numpy.array(rng.uniform(size=(2,2)), dtype= theano.config.floatX)
+        v_w = numpy.array(rng.uniform(size=(2,2), low=-3., high=3.), dtype= theano.config.floatX)
         analytic_grad = grad_fn(v_x, v_w)
         num_grad = multiple_outputs_numeric_grad(cost_fn,
                                                  [v_x, v_w])