flexflow/keras/models/tensor.py [18:127]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - class Tensor(object): __slots__ = ['_ffhandle', 'to_layers', 'from_layer', 'dtype', \ 'batch_shape', 'name', 'num_dims'] def __init__(self, ffmodel=None, shape=None, batch_size=None, dtype=None, sparse=False, tensor=None, ragged=False, meta_only=False, ffhandle=None, **kwargs): if sparse != False: assert 0, "sparse is not supported" if tensor != None: assert 0, "tensor is not supported" if ragged != False: assert 0, "ragged is not supported" batch_shape = None if "batch_shape" in kwargs: batch_shape = kwargs["batch_shape"] self._ffhandle = ffhandle self.to_layers = [] self.from_layer = 0 if dtype == None or dtype == "float32" or dtype == ff.DataType.DT_FLOAT: self.dtype = ff.DataType.DT_FLOAT elif dtype == "float64" or dtype == ff.DataType.DT_DOUBLE: self.dtype = ff.DataType.DT_DOUBLE elif dtype == "int32" or dtype == ff.DataType.DT_INT32: self.dtype = ff.DataType.DT_INT32 elif dtype == "int64" or dtype == ff.DataType.DT_INT64: self.dtype = ff.DataType.DT_INT64 else: assert 0, "not supported" self.name = "" if "name" in kwargs: self.name = kwargs["name"] # create a tensor if ffhandle == None: if batch_shape != None: self.batch_shape = batch_shape self.num_dims = len(batch_shape) else: self.num_dims = len(shape) + 1 if batch_size == None: self.batch_shape = (0,) + shape else: self.batch_shape = (batch_size,) + shape if (meta_only == False): self.create_ff_tensor(ffmodel) # init from handle else: self.batch_shape = ffhandle.dims self.num_dims = ffhandle.num_dims self.__verify_ffhandle_shape() self.__verify_ffhandle_dtype() @property def ffhandle(self): return self._ffhandle @ffhandle.setter def ffhandle(self, handle): assert isinstance(handle, ff.Tensor) == True, "[Tensor]: ffhandle is not the correct type" assert self._ffhandle == None, "[Tensor]: check handle, already set" self._ffhandle = handle if (self.batch_shape[0] == 0): self.set_batch_size(handle.dims[0]) self.__verify_ffhandle_shape() self.__verify_ffhandle_dtype() @property def dtype_str(self): if self.dtype == ff.DataType.DT_FLOAT: return "float32" elif self.dtype == ff.DataType.DT_DOUBLE: return "float64" elif self.dtype == ff.DataType.DT_INT32: return "int32" elif self.dtype == ff.DataType.DT_INT64: return "int64" def create_ff_tensor(self, ffmodel): assert self.batch_shape[0] != 0, "[Tensor]: batch size is not set" if (self.num_dims == 2 or self.num_dims == 4): self._ffhandle = ffmodel.create_tensor(self.batch_shape, self.dtype); else: assert 0, "un-supported dims" self.__verify_ffhandle_shape() self.__verify_ffhandle_dtype() def set_from_layer(self, layer): assert self.from_layer == 0, "[Tensor]: from layer has been set" self.from_layer = layer def set_to_layer(self, layer): self.to_layers.append(layer) def set_batch_size(self, size): lst = list(self.batch_shape) lst[0] = size self.batch_shape = tuple(lst) def __verify_ffhandle_shape(self): assert self.num_dims == self._ffhandle.num_dims, "[Tensor]: check tensor shape" for i in range(0, self.num_dims): assert self.batch_shape[i] == self._ffhandle.dims[i], "[Tensor]: please check shape dim %d (%d == %d)" %(i, self.batch_shape[i], self._ffhandle.dims[i]) def __verify_ffhandle_dtype(self): assert self.dtype == self._ffhandle.data_type - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - python/flexflow/keras/models/tensor.py [18:127]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - class Tensor(object): __slots__ = ['_ffhandle', 'to_layers', 'from_layer', 'dtype', \ 'batch_shape', 'name', 'num_dims'] def __init__(self, ffmodel=None, shape=None, batch_size=None, dtype=None, sparse=False, tensor=None, ragged=False, meta_only=False, ffhandle=None, **kwargs): if sparse != False: assert 0, "sparse is not supported" if tensor != None: assert 0, "tensor is not supported" if ragged != False: assert 0, "ragged is not supported" batch_shape = None if "batch_shape" in kwargs: batch_shape = kwargs["batch_shape"] self._ffhandle = ffhandle self.to_layers = [] self.from_layer = 0 if dtype == None or dtype == "float32" or dtype == ff.DataType.DT_FLOAT: self.dtype = ff.DataType.DT_FLOAT elif dtype == "float64" or dtype == ff.DataType.DT_DOUBLE: self.dtype = ff.DataType.DT_DOUBLE elif dtype == "int32" or dtype == ff.DataType.DT_INT32: self.dtype = ff.DataType.DT_INT32 elif dtype == "int64" or dtype == ff.DataType.DT_INT64: self.dtype = ff.DataType.DT_INT64 else: assert 0, "not supported" self.name = "" if "name" in kwargs: self.name = kwargs["name"] # create a tensor if ffhandle == None: if batch_shape != None: self.batch_shape = batch_shape self.num_dims = len(batch_shape) else: self.num_dims = len(shape) + 1 if batch_size == None: self.batch_shape = (0,) + shape else: self.batch_shape = (batch_size,) + shape if (meta_only == False): self.create_ff_tensor(ffmodel) # init from handle else: self.batch_shape = ffhandle.dims self.num_dims = ffhandle.num_dims self.__verify_ffhandle_shape() self.__verify_ffhandle_dtype() @property def ffhandle(self): return self._ffhandle @ffhandle.setter def ffhandle(self, handle): assert isinstance(handle, ff.Tensor) == True, "[Tensor]: ffhandle is not the correct type" assert self._ffhandle == None, "[Tensor]: check handle, already set" self._ffhandle = handle if (self.batch_shape[0] == 0): self.set_batch_size(handle.dims[0]) self.__verify_ffhandle_shape() self.__verify_ffhandle_dtype() @property def dtype_str(self): if self.dtype == ff.DataType.DT_FLOAT: return "float32" elif self.dtype == ff.DataType.DT_DOUBLE: return "float64" elif self.dtype == ff.DataType.DT_INT32: return "int32" elif self.dtype == ff.DataType.DT_INT64: return "int64" def create_ff_tensor(self, ffmodel): assert self.batch_shape[0] != 0, "[Tensor]: batch size is not set" if (self.num_dims == 2 or self.num_dims == 4): self._ffhandle = ffmodel.create_tensor(self.batch_shape, self.dtype); else: assert 0, "un-supported dims" self.__verify_ffhandle_shape() self.__verify_ffhandle_dtype() def set_from_layer(self, layer): assert self.from_layer == 0, "[Tensor]: from layer has been set" self.from_layer = layer def set_to_layer(self, layer): self.to_layers.append(layer) def set_batch_size(self, size): lst = list(self.batch_shape) lst[0] = size self.batch_shape = tuple(lst) def __verify_ffhandle_shape(self): assert self.num_dims == self._ffhandle.num_dims, "[Tensor]: check tensor shape" for i in range(0, self.num_dims): assert self.batch_shape[i] == self._ffhandle.dims[i], "[Tensor]: please check shape dim %d (%d == %d)" %(i, self.batch_shape[i], self._ffhandle.dims[i]) def __verify_ffhandle_dtype(self): assert self.dtype == self._ffhandle.data_type - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -