fix some missing features in pytorch needed for RNNs

Author: adamlerer

torch/autograd/__init__.py

@@ -1,6 +1,6 @@
 import torch
 
 from .variable import Variable
-from .function import Function, NestedInputFunction
+from .function import Function, NestedIOFunction
 
 assert torch._C._autograd_init()

torch/autograd/functions/tensor.py

@@ -28,16 +28,19 @@ def backward(self, grad_output):
  return grad_input
 
 
-class SetValue(InplaceFunction):
+class SetItem(InplaceFunction):
 
- def __init__(self, index, value):
- super(SetValue, self).__init__(True)
+ def __init__(self, index, value=None):
+ super(SetItem, self).__init__(True)
  self.index = index
  self.value = value
 
- def forward(self, i):
+ def forward(self, i, value=None):
  self.mark_dirty(i)
- i[self.index].fill_(self.value)
+ if self.value is None:
+ i[self.index].copy_(value)
+ else:
+ i[self.index].fill_(self.value)
  return i
 
  def backward(self, grad_output):

torch/autograd/variable.py

@@ -56,7 +56,12 @@ def __setitem__(self, key, value):
  if (isinstance(key, Variable) and
  type(key.data).__name__ == 'ByteTensor'):
  return MaskedFill(value, inplace=True)(self, key)
- return SetValue(key, value)(self)
+ if isinstance(value, Variable):
+ return SetItem(key)(self, value)
+ return SetItem(key, value)(self)
+
+ def __iter__(self):
+ return iter(map(lambda i: self[i], range(self.size(0))))
 
  def __deepcopy__(self, memo):
  if self.creator is None:

torch/nn/modules/module.py

@@ -37,16 +37,25 @@ def _apply(self, fn):
  self._buffers[key] = fn(buf)
  return self
 
+ def _deleteGradParameters(self):
+ for param in self._parameters.values():
+ if hasattr(param, '_grad'):
+ param._grad = None
+
  def cuda(self, device_id=None):
+ self._deleteGradParameters()
  return self._apply(lambda t: t.cuda(device_id))
 
+
  def cpu(self, device_id=None):
+ self._deleteGradParameters()
  return self._apply(lambda t: t.cpu())
 
  def float(self):
  return self._apply(lambda t: t.float())
 
  def double(self):
+ self._deleteGradParameters()
  return self._apply(lambda t: t.double())
 
  def register_backward_hook(self, name, hook):