DataFlow API List

DataFlow graphs can be constructed using DataFlow Python APIs for inference. Graph ProcessPoints and UDF ProcessPoints can be defined to describe the data flow relationship between ProcessPoints. IR files of TensorFlow, ONNX, and MindSpore can be imported to define the computation logic of graph ProcessPoints.

DataFlow Graph Construction APIs

**Table 1** DataFlow graph construction APIs
API	Description
dataflow.FlowData	Indicates data nodes in a DataFlow graph. Each FlowData corresponds to an input.
FlowNode	Indicates compute nodes in a DataFlow graph.
add_process_point	Adds a mapping ProcessPoint (PP) to a FlowNode. Currently, only one PP can be added to a FlowNode. After a PP is added, the input and output of the FlowNode are mapped to those of the PP in sequence by default.
map_input	Maps inputs to FlowNode, indicating that the node_input_index input of FlowNode is sent to the pp_input_index input of ProcessPoint, all attributes in attr are set for the input of ProcessPoint, and the mapped FlowNode is returned. This function is optional. If it is not called, the inputs of FlowNode and ProcessPoint are mapped in sequence by default.
map_output	Maps outputs to FlowNode, indicating that the pp_output_index output of pp is sent to the node_output_index output of FlowNode and then the mapped FlowNode is returned.
set_attr	Sets the attributes of FlowNode.
__call__	Calls FlowNode for computation.
set_balance_scatter	Sets the balance scatter attribute. The UDF with the balance scatter attribute can use balance options to set the load balancing output.
set_balance_gather	Sets the balance gather attribute. The UDF with the balance gather attribute can use balance options to set the load balancing and affinity output.
set_alias	Sets the node alias. When option:ge.experiment.data_flow_deploy_info_path is used to specify the node deployment location, the flow_node_list field can be specified using the alias.
dataflow.FlowFlag	Sets flags in the FlowMsg message header.
FlowGraph	Defines a DataFlow graph, which consists of the input node FlowData and compute node FlowNode.
set_contains_n_mapping_node	Sets whether FlowGraph contains an n_mapping node.
set_inputs_align_attrs	Sets the attribute of input alignment in FlowGraph.
set_exception_catch	Enables or disables the user exception catch function.
dataflow.FlowOutput	Describes the output of FlowNode.
dataflow.Framework	Sets the framework of the original network model.
FuncProcessPoint	Defines the FuncProcessPoint constructor, which returns a FuncProcessPoint object.
set_init_param	Sets the initialization parameters of FuncProcessPoint.
add_invoked_closure	Adds GraphProcessPoint or FlowGraphProcessPoint to be called by FuncProcessPoint, and returns FuncProcessPoint.
GraphProcessPoint	Defines the GraphProcessPoint constructor, which returns a GraphProcessPoint object.
fnode	Generates a FlowNode based on the current GraphProcessPoint and returns a FlowNode object.
dataflow.FlowGraphProcessPoint	Defines the GraphProcessPoint constructor, which returns a GraphProcessPoint object.
Tensor	Defines the Tensor constructor.
numpy	Converts a tensor into an ndarray of NumPy.
dataflow.TensorDesc	Describes a tensor.
dataflow.alloc_tensor	Allocates a dataflow tensor based on the shape, data type, and alignment size.
dataflow.utils.generate_deploy_template	Generates a template of files required by the "ge.experiment.data_flow_deploy_info_path" option in a specified deployment location based on FlowGraph.
register	Registers the serialization, deserialization, and size calculation functions corresponding to the user-defined type. This function can be used together with the feed and fetch APIs to feed or fetch any Python type.
registered	Checks whether the message type ID has been registered.
get_msg_type	Obtains the registered message type based on the type definition.
get_serialize_func	Obtains the registered serialization function based on the message type ID.
get_deserialize_func	Obtains the registered deserialization function based on the message type ID.
get_size_func	Obtains the registered function for calculating the size of the serialized memory based on the message type ID.
deserialize_from_file	Deserializes the serialized pickle file to restore Python objects.
pyflow	Enables functions to be run locally or remotely as pipeline tasks.
method	Enables classes to be run locally or remotely as pipeline tasks in complex scenarios.
npu_model	If the UDF is deployed on the host, data needs to be copied from the device to the local host for computation. In PyTorch scenarios, when all computations, inputs, and outputs reside on the device, data still has to be copied from the device to the host during execution, followed by PyTorch transferring the data back to the device, which degrades performance. The npu_model API optimizes this process by eliminating data transfer overhead, allowing computations to be directly offloaded to the device for execution.
dataflow.CountBatch	CountBatch is used to combine multiple data records into batches based on batchSize, by using UDFs as compute ProcessPoints. This function applies to asynchronous DataFlow scenarios.
dataflow.TimeBatch	The TimeBatch function is based on the UDF. In normal scenarios, a model processes a piece of data each time. To process multiple pieces of data at a time, you need to combine the data into a batch. A basic way is to directly combine the data (N pieces), and add N before the shape. In some scenarios, data of one or more time segments needs to be combined into a batch based on specific dimensions. In this case, you can use TimeBatch.

DataFlow Execution APIs

**Table 2** DataFlow execution APIs
API	Description
dataflow.init	Initializes options in DataFlow.
FlowInfo	Indicates the flow information of DataFlow.
set_user_data	Sets user information.
get_user_data	Obtains user information.
user_data	Obtains user information.
data_size	Obtains the length of user_data.
start_time	Reads and sets the start time of FlowInfo in attribute mode.
end_time	Reads and sets the end time of FlowInfo in attribute mode.
flow_flags	Reads and sets flow_flags of FlowInfo in attribute mode.
transaction_id	Reads and writes transaction IDs in attribute mode.
feed_data	Inputs data to a graph.
feed	Inputs data to a graph. A serializable input is supported.
fetch_data	Obtains the graph output data.
fetch	Obtains the graph output data. A serializable output is supported.
dataflow.finalize	Frees the DataFlow initialization resources.
dataflow.get_running_device_id	Obtains the ID of the device where the current UDF runs during execution. This information is sourced from the UDF deployment location configuration.
dataflow.get_running_instance_id	Obtains IDs of running instances of the current UDF during execution. This information is sourced from the logic_device_list configuration in the data_flow_deploy_info.json file.
dataflow.get_running_instance_num	Obtains the number of running instances of the current UDF during execution. The information is sourced from the logic_device_list configuration in data_flow_deploy_info.json.

Module

dataflow module: namespace of common APIs

Class

CountBatch: class of the CountBatch attribute
FlowData: input node class
FlowFlag: data mark class
FlowGraph: DataFlow graph class
FlowInfo: information class carried in the input and output data
FlowNode: compute node class
FlowOutput: output class of the compute node
Framework: enumeration class of the IR file framework type
FuncProcessPoint: class of the UDF processing point
GraphProcessPoint: class of the graph processing point
Tensor: tensor data class
TensorDesc: tensor description class
TimeBatch: class of the TimeBatch attribute

Function

init(...): DataFlow resource initialization method
finalize(...): DataFlow resource release method

Other Members

**Table 3** Other members
Name	Description
DT_FLOAT	Object of df.data_type.DType 32-bit single-precision floating point number
DT_FLOAT16	Object of df.data_type.DType 16-bit half-precision floating point number
DT_INT8	Object of df.data_type.DType 8-bit signed integer
DT_INT16	Object of df.data_type.DType 16-bit signed integer
DT_UINT16	Object of df.data_type.DType 16-bit unsigned integer
DT_UINT8	Object of df.data_type.DType 8-bit unsigned integer
DT_INT32	Object of df.data_type.DType 32-bit signed integer
DT_INT64	Object of df.data_type.DType 64-bit signed integer
DT_UINT32	Object of df.data_type.DType 32-bit unsigned integer
DT_UINT64	Object of df.data_type.DType 64-bit unsigned integer
DT_BOOL	Object of df.data_type.DType Boolean
DT_DOUBLE	Object of df.data_type.DType 64-bit double-precision floating point number
DT_STRING	Object of df.data_type.DType String

Parent topic: DataFlow APIs