def init_op(self, inputs):
self.op = 'Transpose'
shape = np.zeros(inputs[0].shape).transpose().shape
self.output = intake.array(name=self.op,
shape=shape,
dtype=inputs[0].dtype,
producer=self)
def init_op(self, inputs):
self.op = 'RandomNormal'
shape = (np.zeros(inputs[0].shape) + np.zeros(inputs[1].shape)).shape
self.output = intake.array(name=self.op,
shape=shape,
dtype=config.DTYPE,
producer=self)
def init_op(self, inputs):
self.op = 'Clip'
assert(inputs[1].shape == inputs[2].shape)
shape = inputs[0].shape
self.output = intake.array(name=self.op, shape=shape,
dtype=upcast(inputs),
producer=self)
def init_op(self, inputs):
self.op = 'ArcTan2'
shape = self.inputs[0].shape
self.output = intake.array(name=self.op,
shape=shape,
dtype=config.DTYPE,
producer=self)
def init_op(self, inputs):
self.op = 'Ravel'
shape = np.zeros(inputs[0].shape).ravel().shape
self.output = intake.array(name=self.op,
shape=shape,
dtype=inputs[0].dtype,
producer=self)
def init_op(self, inputs, shape, dtype=config.DTYPE):
self.op = 'Eye'
self.shape = shape
self.output = intake.array(name='eye_mat',
shape=self.shape,
dtype=dtype,
producer=self)
def init_op(self, inputs, axis):
self.op = 'Concatenate'
self.axis=axis
self.shape = np.concatenate((np.zeros(inputs[0].shape),
np.zeros(inputs[1].shape)), axis=axis).shape
self.output = intake.array(name=self.op, shape=self.shape,
dtype=upcast(inputs),
producer=self)
def init_op(self, inputs, at_idx):
self.op = 'ArrayAccess'
if len(at_idx) == 1:
shape = np.zeros(inputs[0].shape)[at_idx[0]].shape
else:
shape = np.zeros(inputs[0].shape)[at_idx[0], at_idx[1]].shape
self.output = intake.array(name=self.op, shape=shape, producer=self)
self.at_idx = at_idx
def __init__(self, inputs, at_idx=None, slices=None):
self.op = 'Assign'
self.inputs = check_type(*inputs)
self.slices = slices
if slices is not None:
if len(slices) == 1:
slice_shape = np.zeros(inputs[0].shape)[slices[0]].shape
else:
slice_shape = np.zeros(inputs[0].shape)[slices[0],
slices[1]].shape
self.output = intake.array(name=self.op,
dtype=inputs[0].dtype,
shape=inputs[0].shape,
producer=self,
slice_shape=slice_shape)
self.__repr__ = lambda: '{}[{}, {}]'.format(
self.inputs[0].name, slice_to_str(slices[0]),
slice_to_str(slices[1])) + '\\n' + str(self.output.shape)
else:
self.output = intake.array(name=self.op,
dtype=inputs[0].dtype,
shape=inputs[0].shape,
producer=self)
self.at_idx = at_idx
if at_idx:
if len(at_idx) == 1:
self.__repr__ = lambda: '{}[{}]'.format(
self.inputs[0].name, at_idx[0]) + '\\n' + str(self.output.
shape)
else:
self.__repr__ = lambda: '{}[{}, {}]'.format(
self.inputs[0].name, at_idx[0], at_idx[1]) + '\\n' + str(
self.output.shape)
# special thing for assign operator... have to add the others
# as dependency
successors = graph.successors(inputs[0].last_producer)
for s in successors:
if s != self.inputs[0].last_producer:
graph.add_edge(s, self, edge_type='helper')
if config.debug or config.group_class or config.group_func:
self.caller_info = get_caller_info('expander.py', 'intake.py')
else:
self.caller_info = None
self.add_to_graph()
self.inputs[0].assignment.append(self)
def init_op(self, inputs, order):
self.op = 'Norm'
if order != 2:
raise NotImplementedError('Only 2 order norm supported currently.')
if inputs[0].ndim > 1:
if inputs[0].shape[1] > 1:
raise NotImplementedError('Only vectors supported.')
self.output = intake.array(name=self.op, shape=(), dtype=upcast(inputs),
producer=self)
def init_op(self, inputs):
self.op = 'Dot'
# result of mat mul = new matrix
# n x m * m x p -> n * p
shape = np.dot(np.zeros(inputs[0].shape), np.zeros(inputs[1].shape)).shape
self.output = intake.array(name=self.op,
shape=shape,
dtype=upcast(inputs),
producer=self)
def init_op(self, inputs):
self.op = 'Solve'
self.add_to_graph()
if not (inputs[0].shape[0] == inputs[0].shape[1] ==
inputs[1].shape[0]):
raise ValueError(
'LHS shape {} and RHS shape {} not compatible for solve.'
.format(inputs[0].shape, inputs[1].shape))
self.output = intake.array(name=self.op, shape=inputs[1].shape,
dtype=upcast(inputs), producer=self)
def init_op(self, inputs):
# TODO check why this happens and what it does
shape = self.inputs[0].shape
if hasattr(self, 'dtype'):
# dtype for sin etc. has to change
dtype = self.dtype
else:
dtype = self.inputs[0].dtype
self.output = intake.array(name=self.op, shape=shape, dtype=dtype,
producer=self)
def init_op(self, inputs, keys):
self.op = 'View'
# find shape by slicing zeros vector
self.slices = keys
if len(self.slices) == 1:
new_shape = np.zeros(inputs[0].shape)[self.slices[0]].shape
else:
new_shape = np.zeros(inputs[0].shape)[self.slices[0], self.slices[1]].shape
self.output = intake.array(name=self.op, shape=new_shape, producer=self)
def init_op(self, inputs):
self.op = 'Min'
self.output = intake.array(name=self.op,
shape=(),
dtype=inputs[0].dtype,
producer=self)
def init_op(self, inputs):
self.output = intake.array(name=self.op,
shape=self.shape_op(inputs),
dtype=upcast(inputs),
producer=self)
def init_op(self, inputs, shape):
self.op = 'Reshape'
self.output = intake.array(name=self.op,
shape=shape,
dtype=inputs[0].dtype,
producer=self)
def init_op(self, inputs):
self.op = 'Any'
self.output = intake.array(name=self.op,
shape=inputs[0].shape,
dtype=bool,
producer=self)
