print not supported in autograd

[tonyyang-svail]

Looks like tangent only supports autograd on operations registered at adjoint. Is it a good idea to simply ignore those non-registered operations when doing autograd?

def f(x):
  a = x * x
  b = x * a
  c = a + b
  print c
  return c

import tangent
df = tangent.grad(f)

Gives

ValueErrorTraceback (most recent call last)
<ipython-input-10-d32ca81b2574> in <module>()
      1 import tangent
----> 2 df = tangent.grad(f)

/usr/local/lib/python2.7/dist-packages/tangent/grad_util.pyc in grad(func, wrt, optimized, preserve_result, check_dims, verbose)
    384       check_dims=check_dims,
    385       input_derivative=INPUT_DERIVATIVE.DefaultOne,
--> 386       verbose=verbose)
    387 
    388 

/usr/local/lib/python2.7/dist-packages/tangent/grad_util.pyc in autodiff(func, wrt, optimized, motion, mode, preserve_result, check_dims, input_derivative, verbose)
    288   # Generate the derivative
    289   node, namespace = autodiff_tree(func, wrt, motion, mode, preserve_result,
--> 290                                   check_dims, verbose)
    291 
    292   if mode == 'reverse' and motion == 'joint':

/usr/local/lib/python2.7/dist-packages/tangent/grad_util.pyc in autodiff_tree(func, wrt, motion, mode, preserve_result, check_dims, verbose)
    142 
    143   node, required = autodiff_ast(func, wrt, motion, mode, preserve_result,
--> 144                                 check_dims, verbose)
    145   final.body.extend(node.body)
    146 

/usr/local/lib/python2.7/dist-packages/tangent/grad_util.pyc in autodiff_ast(func, wrt, motion, mode, preserve_result, check_dims, verbose)
     95   if mode == 'reverse':
     96     node, required, stack = reverse_ad.reverse_ad(node.body[0], wrt,
---> 97                                                   preserve_result, check_dims)
     98     if verbose >= 2:
     99       print('RAW')

/usr/local/lib/python2.7/dist-packages/tangent/reverse_ad.pyc in reverse_ad(node, wrt, preserve_result, check_dims)
    841 
    842   ad = ReverseAD(wrt, preserve_result, check_dims)
--> 843   pri, adj = ad.visit(node)
    844   mod = gast.Module(body=[pri, adj])
    845   mod = annotate.find_stacks(mod)

/usr/local/lib/python2.7/dist-packages/tangent/reverse_ad.pyc in visit(self, node)
    149     if anno.hasanno(node, 'active_in'):
    150       self.active_variables = anno.getanno(node, 'active_in')
--> 151     pri, adj = visitor(node)
    152 
    153     # Annotate primal and adjoint statements

/usr/local/lib/python2.7/dist-packages/tangent/reverse_ad.pyc in visit_FunctionDef(self, node)
    212 
    213     # Perform AD on the function body
--> 214     body, adjoint_body = self.visit_statements(node.body[:-1])
    215 
    216     # Annotate the first statement of the primal and adjoint as such

/usr/local/lib/python2.7/dist-packages/tangent/reverse_ad.pyc in visit_statements(self, nodes)
    285     primals, adjoints = [], collections.deque()
    286     for node in nodes:
--> 287       primal, adjoint = self.visit(node)
    288       if not isinstance(primal, list):
    289         primal = [primal]

/usr/local/lib/python2.7/dist-packages/tangent/reverse_ad.pyc in visit(self, node)
    142     method = 'visit_' + node.__class__.__name__
    143     if not hasattr(self, method):
--> 144       raise ValueError('Unknown node type: %s' % node.__class__.__name__)
    145     visitor = getattr(self, method)
    146 

ValueError: Unknown node type: Print

[mdanatg]

Blanked whitelisting is risky because those functions might have side effects and in general may need to be differentiated. However, whitelisting known functions like print, assert that are guaranteed not to have effects on the result sounds reasonable.