Python theano.compile.SharedVariable() Examples

def __init__(self, svobj):
        '''Constructor of the MVSharedVariable

        The constructor will create ArrayTableHandler and associate the shared
        variable with it. The initial value of ArrayTableHandler will be same
        as the value of SharedVariable. If different initial value is used in
        different processes, the average of them will be used as the initial
        value
        '''
        assert(isinstance(svobj, SharedVariable))
        self._svobj = svobj
        self._mv_array = mv.ArrayTableHandler(self._svobj.get_value().size,
                                              init_value=self._svobj.get_value().reshape((-1,)))

        mv.barrier()  # add barrier to make sure the initial values have token effect
        # _last_mv_data restore a copy of value. It will be used for calculate
        # the update for multiverso when calling mv_sync
        self._last_mv_data = self._mv_array.get().reshape(self._svobj.get_value().shape)
        self._svobj.set_value(self._last_mv_data, borrow=False) 

def sparse_constructor(value, name=None, strict=False, allow_downcast=None,
                       borrow=False, format=None):
    """
    SharedVariable Constructor for SparseType.

    writeme

    """
    if not isinstance(value, scipy.sparse.spmatrix):
        raise TypeError("Expected a sparse matrix in the sparse shared variable constructor. Received: ",
                        value.__class__)

    if format is None:
        format = value.format
    type = SparseType(format=format, dtype=value.dtype)
    if not borrow:
        value = copy.deepcopy(value)
    return SparseTensorSharedVariable(type=type, value=value, name=name,
                                      strict=strict, allow_downcast=allow_downcast) 

def sparse_constructor(value, name=None, strict=False, allow_downcast=None,
                       borrow=False, format=None):
    """
    SharedVariable Constructor for SparseType.

    writeme

    """
    if not isinstance(value, scipy.sparse.spmatrix):
        raise TypeError("Expected a sparse matrix in the sparse shared variable constructor. Received: ",
                        value.__class__)

    if format is None:
        format = value.format
    type = SparseType(format=format, dtype=value.dtype)
    if not borrow:
        value = copy.deepcopy(value)
    return SparseTensorSharedVariable(type=type, value=value, name=name,
                                      strict=strict, allow_downcast=allow_downcast) 

def tensor_constructor(value, name=None, strict=False, allow_downcast=None,
                       borrow=False, broadcastable=None, target='cpu'):
    """
    SharedVariable Constructor for TensorType.

    Notes
    -----
    Regarding the inference of the broadcastable pattern...
    The default is to assume that the value might be resized in any
    dimension, so the default broadcastable is ``(False,)*len(value.shape)``.
    The optional `broadcastable` argument will override this default.

    """
    if target != 'cpu':
        raise TypeError('not for cpu')

    if not isinstance(value, numpy.ndarray):
        raise TypeError()

    # if no broadcastable is given, then the default is to assume that
    # the value might be resized in any dimension in the future.
    #
    if broadcastable is None:
        broadcastable = (False,) * len(value.shape)
    type = TensorType(value.dtype, broadcastable=broadcastable)
    return TensorSharedVariable(type=type,
                                value=numpy.array(value, copy=(not borrow)),
                                name=name,
                                strict=strict,
                                allow_downcast=allow_downcast)


# TensorSharedVariable brings in the tensor operators, is not ideal, but works
# as long as we dont do purely scalar-scalar operations
# _tensor_py_operators is first to have its version of __{gt,ge,lt,le}__
#
# N.B. THERE IS ANOTHER CLASS CALLED ScalarSharedVariable in the
# theano.scalar.sharedvar file.  It is not registered as a shared_constructor,
# this one is. 

def mv_shared(*args, **kwargs):
    '''mv_shared works same as `theano.shared`

    It calls `theano.shared` to create the SharedVariable and use
    MVSharedVariable to wrap it.
    '''
    var = shared(*args, **kwargs)
    mv_shared.shared_vars.append(MVSharedVariable(var))
    return var 

def __getattribute__(self, attr):
        '''This function make MVSharedVariable act same as SharedVariable'''
        if attr in ['_svobj', '_mv_array', '_last_mv_data']:
            # If get the attribute of self, use parent __getattribute__ to get
            # attribute from the object, otherwise it will fall into infinite
            # loop
            return object.__getattribute__(self, attr)
        elif attr in ['mv_sync', "__getstate__"]:
            # If call method of MVSharedVariable, then call the method directly
            # and bound the method to self object
            return getattr(MVSharedVariable, attr).__get__(self)
        else:
            # Otherwise I will get attribute from the wrapped object
            return getattr(self._svobj, attr) 

def mv_sync(self):
        ''' sync values with multiverso server

        mv_sync will add the delta of SharedVariable, which is usually the
        gradients in typical examples, to parameter server and then get the
        latest value in multiverso.
        '''
        # because multiverso always use add method to sync value, the delta
        # will be the difference of the current value of last synced value
        self._mv_array.add(self._svobj.get_value() - self._last_mv_data)

        self._svobj.set_value(self._mv_array.get().reshape(self._svobj.get_value().shape))
        self._last_mv_data = self._svobj.get_value(borrow=False) 

def shared(value, name=None, strict=False, allow_downcast=None):
    """
    SharedVariable constructor for scalar values. Default: int64 or float64.

    Notes
    -----
    We implement this using 0-d tensors for now.

    """
    if not isinstance(value, (numpy.number, float, int, complex)):
        raise TypeError()
    try:
        dtype = value.dtype
    except AttributeError:
        dtype = numpy.asarray(value).dtype

    dtype = str(dtype)
    value = getattr(numpy, dtype)(value)
    scalar_type = Scalar(dtype=dtype)
    rval = ScalarSharedVariable(
        type=scalar_type,
        value=value,
        name=name,
        strict=strict,
        allow_downcast=allow_downcast)
    return rval 

def gpuarray_shared_constructor(value, name=None, strict=False,
                                allow_downcast=None, borrow=False,
                                broadcastable=None, target=None):
    """
    SharedVariable constructor for GpuArrayType.

    See :func:`theano.shared`.

    """
    if target == 'gpu' or target == 'cpu':
        raise TypeError('not for me')

    if not isinstance(value, (numpy.ndarray, pygpu.gpuarray.GpuArray)):
        raise TypeError('ndarray or GpuArray required')

    try:
        get_context(target)
    except ContextNotDefined:
        # Don't make this a hard error if we attempt to make a shared
        # variable while there is no default context.
        if target is None:
            raise TypeError('No default context and no context specified')
        raise

    if broadcastable is None:
        broadcastable = (False,) * value.ndim
    type = GpuArrayType(value.dtype, broadcastable, context_name=target)
    deviceval = pygpu.gpuarray.array(value, copy=(not borrow),
                                     context=type.context)
    return GpuArraySharedVariable(type=type, value=deviceval, name=name,
                                  strict=strict) 

def cuda_shared_constructor(value, name=None, strict=False,
                            allow_downcast=None, borrow=False,
                            broadcastable=None, target='gpu'):
    """
    SharedVariable Constructor for CudaNdarrayType.

    """
    if target != 'gpu':
        raise TypeError('not for gpu')

    # THIS CONSTRUCTOR TRIES TO CAST VALUE TO A FLOAT32, WHICH THEN GOES ONTO THE CARD
    # SO INT shared vars, float64 shared vars, etc. all end up on the card.
    # THIS IS NOT THE DEFAULT BEHAVIOUR THAT WE WANT.
    # SEE float32_shared_constructor

    # TODO: what should strict mean in this context, since we always have to make a copy?
    if strict:
        _value = value
    else:
        _value = theano._asarray(value, dtype='float32')

    if not isinstance(_value, numpy.ndarray):
        raise TypeError('ndarray required')
    if _value.dtype.num != CudaNdarrayType.typenum:
        raise TypeError('float32 ndarray required')

    if broadcastable is None:
        broadcastable = (False,) * len(value.shape)
    type = CudaNdarrayType(broadcastable=broadcastable)
    print("trying to return?")
    try:
        rval = CudaNdarraySharedVariable(type=type, value=_value, name=name, strict=strict)
    except Exception as e:
        print("ERROR", e)
        raise
    return rval 

def get_value(self, borrow=False, return_internal_type=False):
        """
        Return the value of this SharedVariable's internal array.

        Parameters
        ----------
        borrow
            Permit the return of internal storage, when used in conjunction with
            ``return_internal_type=True``.
        return_internal_type
            True to return the internal ``cuda_ndarray`` instance rather than a
            ``numpy.ndarray`` (Default False).

        By default ``get_value()`` copies from the GPU to a ``numpy.ndarray``
        and returns that host-allocated array.

        ``get_value(False,True)`` will return a GPU-allocated copy of the
        original GPU array.

        ``get_value(True,True)`` will return the original GPU-allocated array
        without any copying.

        """
        if return_internal_type or not self.get_value_return_ndarray:
            # return a cuda_ndarray
            if borrow:
                return self.container.value
            else:
                return copy.deepcopy(self.container.value)
        else:  # return an ndarray
            return numpy.asarray(self.container.value) 

def __init__(self, inputs, outputs, **kwargs):
        if not isinstance(outputs, list):
            raise TypeError('outputs must be list', outputs)
        for i in inputs + outputs:
            if not isinstance(i, gof.Variable):
                raise TypeError(
                    'inputs and outputs must be Variable instances', i)
        if 'updates' in kwargs or 'givens' in kwargs:
            raise TypeError('updates and givens are not allowed in kwargs')

        # To support correctly shared variables the inner fct should
        # not see them. Otherwise their is problem with the gradient.
        self.shared_inputs = [var for var in gof.graph.inputs(outputs)
                              if isinstance(var, SharedVariable)]
        shared_vars = [var.type() for var in self.shared_inputs]
        new = rebuild_collect_shared(outputs, inputs=inputs + shared_vars,
                                     replace=dict(izip(self.shared_inputs,
                                                       shared_vars)),
                                     copy_inputs_over=False)
        (new_inputs, new_outputs,
         [clone_d, update_d, update_expr, shared_inputs]) = new
        assert len(new_inputs) == len(inputs) + len(self.shared_inputs)
        assert len(new_outputs) == len(outputs)
        assert not update_d
        assert not update_expr
        assert not shared_inputs

        self.new_inputs = new_inputs
        self.new_outputs = new_outputs
        self.inputs = inputs
        self.outputs = outputs
        self.kwargs = kwargs
        self.input_types = [input.type for input in inputs]
        self.output_types = [output.type for output in outputs] 

def scalar_constructor(value, name=None, strict=False, allow_downcast=None,
                       borrow=False, target='cpu'):
    """
    SharedVariable constructor for scalar values. Default: int64 or float64.

    Notes
    -----
    We implement this using 0-d tensors for now.

    We ignore the borrow parameter as we convert ``value`` to an
    ndarray (this is a new object). This respects the semantic of
    borrow, as it is a hint to Theano that we can reuse it.

    """
    if target != 'cpu':
        raise TypeError('not for cpu')

    if not isinstance(value, (numpy.number, float, int, complex)):
        raise TypeError()
    try:
        dtype = value.dtype
    except Exception:
        dtype = numpy.asarray(value).dtype

    dtype = str(dtype)
    value = theano._asarray(value, dtype=dtype)
    tensor_type = TensorType(dtype=str(value.dtype), broadcastable=[])

    try:
        # Do not pass the dtype to asarray because we want this to fail if
        # strict is True and the types do not match.
        rval = ScalarSharedVariable(type=tensor_type,
                                    value=numpy.array(value, copy=True),
                                    name=name,
                                    strict=strict,
                                    allow_downcast=allow_downcast)
        return rval
    except Exception:
        traceback.print_exc()
        raise 

def __init__(self, shape, input_var=None, name=None, testval=None, **kwargs):
        _InputLayer.__init__(self, shape, input_var=input_var, name=name, **kwargs)
        if testval is not None:
            self.input_var.tag.test_value = testval
        if (not isinstance(self.input_var, SharedVariable)
           and not hasattr(self.input_var.tag, 'test_value')):
            shape = [s if s is not None else 2 for s in self.shape]
            dtype = self.input_var.dtype
            self.input_var.tag.test_value = np.random.uniform(size=shape).astype(dtype) 

def shared(value, name=None, strict=False, allow_downcast=None):
    """
    SharedVariable constructor for scalar values. Default: int64 or float64.

    Notes
    -----
    We implement this using 0-d tensors for now.

    """
    if not isinstance(value, (numpy.number, float, integer_types, complex)):
        raise TypeError()
    try:
        dtype = value.dtype
    except AttributeError:
        dtype = numpy.asarray(value).dtype

    dtype = str(dtype)
    value = getattr(numpy, dtype)(value)
    scalar_type = Scalar(dtype=dtype)
    rval = ScalarSharedVariable(
        type=scalar_type,
        value=value,
        name=name,
        strict=strict,
        allow_downcast=allow_downcast)
    return rval 

def gpuarray_shared_constructor(value, name=None, strict=False,
                                allow_downcast=None, borrow=False,
                                broadcastable=None, target=None):
    """
    SharedVariable constructor for GpuArrayType.

    See :func:`theano.shared`.

    """
    if target == 'gpu' or target == 'cpu':
        raise TypeError('not for me')

    if not isinstance(value, (numpy.ndarray, pygpu.gpuarray.GpuArray)):
        raise TypeError('ndarray or GpuArray required')

    try:
        get_context(target)
    except ContextNotDefined:
        # Don't make this a hard error if we attempt to make a shared
        # variable while there is no default context.
        if target is None:
            raise TypeError('No default context and no context specified')
        raise

    if broadcastable is None:
        broadcastable = (False,) * value.ndim
    type = GpuArrayType(value.dtype, broadcastable, context_name=target)
    deviceval = pygpu.gpuarray.array(value, copy=(not borrow),
                                     context=type.context)
    return GpuArraySharedVariable(type=type, value=deviceval, name=name,
                                  strict=strict) 

def cuda_shared_constructor(value, name=None, strict=False,
                            allow_downcast=None, borrow=False,
                            broadcastable=None, target='gpu'):
    """
    SharedVariable Constructor for CudaNdarrayType.

    """
    if target != 'gpu':
        raise TypeError('not for gpu')

    # THIS CONSTRUCTOR TRIES TO CAST VALUE TO A FLOAT32, WHICH THEN GOES ONTO THE CARD
    # SO INT shared vars, float64 shared vars, etc. all end up on the card.
    # THIS IS NOT THE DEFAULT BEHAVIOUR THAT WE WANT.
    # SEE float32_shared_constructor

    # TODO: what should strict mean in this context, since we always have to make a copy?
    if strict:
        _value = value
    else:
        _value = theano._asarray(value, dtype='float32')

    if not isinstance(_value, numpy.ndarray):
        raise TypeError('ndarray required')
    if _value.dtype.num != CudaNdarrayType.typenum:
        raise TypeError('float32 ndarray required')

    if broadcastable is None:
        broadcastable = (False,) * len(value.shape)
    type = CudaNdarrayType(broadcastable=broadcastable)
    print("trying to return?")
    try:
        rval = CudaNdarraySharedVariable(type=type, value=_value, name=name, strict=strict)
    except Exception as e:
        print("ERROR", e)
        raise
    return rval 

def get_value(self, borrow=False, return_internal_type=False):
        """
        Return the value of this SharedVariable's internal array.

        Parameters
        ----------
        borrow
            Permit the return of internal storage, when used in conjunction with
            ``return_internal_type=True``.
        return_internal_type
            True to return the internal ``cuda_ndarray`` instance rather than a
            ``numpy.ndarray`` (Default False).

        By default ``get_value()`` copies from the GPU to a ``numpy.ndarray``
        and returns that host-allocated array.

        ``get_value(False,True)`` will return a GPU-allocated copy of the
        original GPU array.

        ``get_value(True,True)`` will return the original GPU-allocated array
        without any copying.

        """
        if return_internal_type or not self.get_value_return_ndarray:
            # return a cuda_ndarray
            if borrow:
                return self.container.value
            else:
                return copy.deepcopy(self.container.value)
        else:  # return an ndarray
            return numpy.asarray(self.container.value) 

def __init__(self, inputs, outputs, **kwargs):
        if not isinstance(outputs, list):
            raise TypeError('outputs must be list', outputs)
        for i in inputs + outputs:
            if not isinstance(i, gof.Variable):
                raise TypeError(
                    'inputs and outputs must be Variable instances', i)
        if 'updates' in kwargs or 'givens' in kwargs:
            raise TypeError('updates and givens are not allowed in kwargs')

        # To support correctly shared variables the inner fct should
        # not see them. Otherwise their is problem with the gradient.
        self.shared_inputs = [var for var in gof.graph.inputs(outputs)
                              if isinstance(var, SharedVariable)]
        shared_vars = [var.type() for var in self.shared_inputs]
        new = rebuild_collect_shared(outputs, inputs=inputs + shared_vars,
                                     replace=dict(izip(self.shared_inputs,
                                                       shared_vars)),
                                     copy_inputs_over=False)
        (new_inputs, new_outputs,
         [clone_d, update_d, update_expr, shared_inputs]) = new
        assert len(new_inputs) == len(inputs) + len(self.shared_inputs)
        assert len(new_outputs) == len(outputs)
        assert not update_d
        assert not update_expr
        assert not shared_inputs

        self.new_inputs = new_inputs
        self.new_outputs = new_outputs
        self.inputs = inputs
        self.outputs = outputs
        self.kwargs = kwargs
        self.input_types = [input.type for input in inputs]
        self.output_types = [output.type for output in outputs] 

def scalar_constructor(value, name=None, strict=False, allow_downcast=None,
                       borrow=False, target='cpu'):
    """
    SharedVariable constructor for scalar values. Default: int64 or float64.

    Notes
    -----
    We implement this using 0-d tensors for now.

    We ignore the borrow parameter as we convert ``value`` to an
    ndarray (this is a new object). This respects the semantic of
    borrow, as it is a hint to Theano that we can reuse it.

    """
    if target != 'cpu':
        raise TypeError('not for cpu')

    if not isinstance(value, (numpy.number, float, integer_types, complex)):
        raise TypeError()
    try:
        dtype = value.dtype
    except Exception:
        dtype = numpy.asarray(value).dtype

    dtype = str(dtype)
    value = theano._asarray(value, dtype=dtype)
    tensor_type = TensorType(dtype=str(value.dtype), broadcastable=[])

    try:
        # Do not pass the dtype to asarray because we want this to fail if
        # strict is True and the types do not match.
        rval = ScalarSharedVariable(type=tensor_type,
                                    value=numpy.array(value, copy=True),
                                    name=name,
                                    strict=strict,
                                    allow_downcast=allow_downcast)
        return rval
    except Exception:
        traceback.print_exc()
        raise 

def tensor_constructor(value, name=None, strict=False, allow_downcast=None,
                       borrow=False, broadcastable=None, target='cpu'):
    """
    SharedVariable Constructor for TensorType.

    Notes
    -----
    Regarding the inference of the broadcastable pattern...
    The default is to assume that the value might be resized in any
    dimension, so the default broadcastable is ``(False,)*len(value.shape)``.
    The optional `broadcastable` argument will override this default.

    """
    if target != 'cpu':
        raise TypeError('not for cpu')

    if not isinstance(value, numpy.ndarray):
        raise TypeError()

    # if no broadcastable is given, then the default is to assume that
    # the value might be resized in any dimension in the future.
    #
    if broadcastable is None:
        broadcastable = (False,) * len(value.shape)
    type = TensorType(value.dtype, broadcastable=broadcastable)
    return TensorSharedVariable(type=type,
                                value=numpy.array(value, copy=(not borrow)),
                                name=name,
                                strict=strict,
                                allow_downcast=allow_downcast)


# TensorSharedVariable brings in the tensor operators, is not ideal, but works
# as long as we dont do purely scalar-scalar operations
# _tensor_py_operators is first to have its version of __{gt,ge,lt,le}__
#
# N.B. THERE IS ANOTHER CLASS CALLED ScalarSharedVariable in the
# theano.scalar.sharedvar file.  It is not registered as a shared_constructor,
# this one is. 

def float32_shared_constructor(value, name=None, strict=False,
                               allow_downcast=None, borrow=False,
                               broadcastable=None, target='gpu'):
    """
    SharedVariable Constructor for CudaNdarrayType from numpy.ndarray or
    CudaNdarray.

    """
    if target != 'gpu':
        raise TypeError('not for gpu')
    if theano.sandbox.cuda.use.device_number is None:
        theano.sandbox.cuda.use("gpu",
                                force=True,
                                default_to_move_computation_to_gpu=False,
                                move_shared_float32_to_gpu=False,
                                enable_cuda=False)

    # if value isn't a float32 ndarray, or a CudaNdarray then raise

    if not isinstance(value, (numpy.ndarray, theano.sandbox.cuda.CudaNdarray)):
        raise TypeError('ndarray or CudaNdarray required')
    if isinstance(value, numpy.ndarray) and value.dtype.num != CudaNdarrayType.typenum:
        raise TypeError('float32 ndarray required')

    if broadcastable is None:
        broadcastable = (False,) * len(value.shape)
    type = CudaNdarrayType(broadcastable=broadcastable)
    get_value_return_ndarray = True
    if isinstance(value, theano.sandbox.cuda.CudaNdarray):
        get_value_return_ndarray = False
        if borrow:
            deviceval = value
        else:
            deviceval = value.copy()
    else:
        # type.broadcastable is guaranteed to be a tuple, which this next
        # function requires
        deviceval = type_support_filter(value, type.broadcastable, False, None)

    try:
        rval = CudaNdarraySharedVariable(type=type, value=deviceval, name=name, strict=strict)
    except Exception as e:
        print("ERROR", e)
        raise

    rval.get_value_return_ndarray = get_value_return_ndarray

    return rval 

def float32_shared_constructor(value, name=None, strict=False,
                               allow_downcast=None, borrow=False,
                               broadcastable=None, target='gpu'):
    """
    SharedVariable Constructor for CudaNdarrayType from numpy.ndarray or
    CudaNdarray.

    """
    if target != 'gpu':
        raise TypeError('not for gpu')
    if theano.sandbox.cuda.use.device_number is None:
        theano.sandbox.cuda.use("gpu",
                                force=True,
                                default_to_move_computation_to_gpu=False,
                                move_shared_float32_to_gpu=False,
                                enable_cuda=False)

    # if value isn't a float32 ndarray, or a CudaNdarray then raise

    if not isinstance(value, (numpy.ndarray, theano.sandbox.cuda.CudaNdarray)):
        raise TypeError('ndarray or CudaNdarray required')
    if isinstance(value, numpy.ndarray) and value.dtype.num != CudaNdarrayType.typenum:
        raise TypeError('float32 ndarray required')

    if broadcastable is None:
        broadcastable = (False,) * len(value.shape)
    type = CudaNdarrayType(broadcastable=broadcastable)
    get_value_return_ndarray = True
    if isinstance(value, theano.sandbox.cuda.CudaNdarray):
        get_value_return_ndarray = False
        if borrow:
            deviceval = value
        else:
            deviceval = value.copy()
    else:
        # type.broadcastable is guaranteed to be a tuple, which this next
        # function requires
        deviceval = type_support_filter(value, type.broadcastable, False, None)

    try:
        rval = CudaNdarraySharedVariable(type=type, value=deviceval, name=name, strict=strict)
    except Exception as e:
        print("ERROR", e)
        raise

    rval.get_value_return_ndarray = get_value_return_ndarray

    return rval