Loop#

Versioned name: Loop-5

Category: Infrastructure

Short description: Loop operation performs recurrent execution of the network, which is described in the body, iterating through the data. The operation has similar semantic to the ONNX Loop operation.

Detailed description

The body of the Loop can be executed 0 or more times depending on the values passed to the Loop operation inputs called “trip count”, “execution condition” and input of the Loop body called “current iteration”.

These Loop operation inputs have the following meaning:

  1. Trip count is an integer scalar or 1D tensor with 1 element input specifying maximum number of iterations. To simulate infinite loop Constant -1 can be provided as input.

  2. Loop execution condition input is a boolean scalar or 1D tensor with 1 element input specifying whether to run the first loop iteration or not. Note, that the body of the Loop must yield the condition value for the consecutive iterations.

There are several combinations of these two inputs (trip_count, execution condition) which are described in the following code snippet:

input (-1, true) // infinite loop
    bool cond = true;
    for (int i = 0; cond; ++i)
    {
        cond = true; // sub-graph calculating condition must always return "true"!
    }

input (-1, cond) // while loop
    bool cond = ...;
    for (int i = 0; cond; ++i)
    {
        cond = ...;
    }

input (-1, true) // do-while loop
    bool cond = true;
    for (int i = 0; cond; ++i)
    {
        cond = ...;
    }

input (trip_count, true) // for loop
    int trip_count = ...;
    bool cond = true;
    for (int i = 0; i < trip_count; ++i)
    {
        cond = true; // sub-graph calculating condition must always return "true"!
    }

input (trip_count, cond) // for with condition
    int trip_count = ...;
    bool cond = ...;
    for (int i = 0; i < trip_count && cond; ++i)
    {
        cond = ...;
    }

  1. One of the body graph inputs called “current iteration” is an integer scalar or 1D integer tensor with 1 number specifying current iteration number. The iteration number starts from 0 and incremented by one for each iteration. This input is optional and may not exist if the iteration number value is not used in the body.

  2. One of the body graph outputs is called “condition” is a boolean scalar or 1D tensor with 1 element. This value is used to decide whenever to perform the next iteration or not.

Loop operation description in the IR has regular sections: input and output. They connect Loop body to the outer graph and specify condition(s). Loop operation description in the IR also has several special sections: body, port_map and back_edges similar to the ones from the TensorIterator operation but having some important features described below.

  1. The body operation getting an input from the main graph should have an entry in the port_map section of the Loop operation. These edges connect input ports of the Loop with the body Parameters.

  2. Input tensors to the Loop can be sliced along a specified axis, the Loop can iterates over all sliced parts. The corresponding input entry in the port_map should have axis attribute specifying the axis to slice. Therefore, inputs to the Loop operation corresponding to input entries in the port_map without axis attribute are used “as is” (without slicing).

  3. The body operation producing tensor to be used in the subsequent iterations (like in RNN models) should have a back edge described in the back_edges section of the operation. The back edge connects the respective body Parameter and Result operations. For such a case the Loop operation node provides input for the first iteration, while corresponding Loop operation output produces the tensor computed during the last iteration.

  4. Output tensors produced by a particular body operation across all iterations can be concatenated and returned as a Loop operation output (this is a “scan output” according to the ONNX* Loop operation specification ). The corresponding output entry in the port_map should have axis attribute specifying the axis to concatenate. Therefore, outputs from operations corresponding to output entries in the port_map without axis attribute are returned “as is” (without concatenation).

  5. There is one body Parameter operation not connected through the port_map. This is a “current iteration” input. The Loop operation is responsible for providing the appropriate value for each iteration.

  6. Connection of nodes inside the Loop body with the main graph should be done through Parameter and Result body operations. No other ways to connect graphs are allowed.

Loop attributes:

  • Body:

    body is a network that will be recurrently executed. The network is described operation by operation as a typical IR network.

    • Body attributes:

      No attributes available.

  • Port map:

    port_map is a set of rules to map input or output data tensors of Loop operation onto body data tensors. The port_map entries can be`` input`` and output. Each entry describes a corresponding mapping rule.

    • Port map attributes:

      • external_port_id

        • Description: external_port_id is a port ID of the Loop operation. The value -1 means that the body node is not connected to the Loop operation.

        • Range of values: IDs of the Loop outputs

        • Type: int

        • Default value: None

        • Required: yes

      • internal_layer_id

        • Description: internal_layer_id is a Parameter or Result operation ID inside the body network to map to.

        • Range of values: IDs of the Parameter operations inside in the Loop operation

        • Type: int

        • Default value: None

        • Required: yes

      • axis

        • Description: if axis is specified for output entry, then it is an axis to concatenate the body Result output across all iterations. If axis is specified for input entry, then it is an axis to iterate through, it triggers the slicing of the input tensor.

        • Range of values: an integer. Negative value means counting dimension from the end.

        • Type: int

        • Default value: None

        • Required: no

  • Back edges:

    back_edges is a set of rules to transfer tensor values from body outputs at one iteration to body parameters at the next iteration. Back edge connects some Result operation in the body to Parameter operation in the same body.

    • Back edge attributes:

      • from-layer

        • Description: from-layer is a Result operation ID inside the body network.

        • Range of values: IDs of the Result operations inside the Loop

        • Type: int

        • Default value: None

        • Required: yes

      • to-layer

        • Description: to-layer is a Parameter operation ID inside the body network to end mapping.

        • Range of values: IDs of the Parameter operations inside the Loop

        • Type: int

        • Default value: None

        • Required: yes

Loop Inputs

  • Trip count: A scalar or 1D tensor with 1 element of int64 or int32 type specifying maximum number of iterations. Required.

  • ExecutionCondition: A scalar or 1D tensor with 1 element of boolean type specifying whether to execute the first iteration or not. True value means to execute the 1st iteration. Required.

  • Multiple other inputs: tensors of different types and shapes. Optional.

Loop Outputs

  • Multiple outputs: Results of execution of the body. Tensors of any type and shape.

Body Inputs

  • Multiple inputs: tensors of different types and shapes except the one corresponding to the current iteration number. This input is marked in the port_map with attribute purpose = "current_iteration" and produces a scalar or 1D tensor with 1 element of int64 or int32 type. Optional.

Body Outputs

  • Multiple outputs: Results of execution of the body. Tensors of any type and shape except the one corresponding to the output with execution condition. This output is marked in the port_map with attribute purpose = "execution_condition" and is mandatory and produces a scalar or 1D tensor with 1 element of boolean type. Other outputs are optional.

Examples

Example 1: a typical Loop structure

<layer type="Loop" ... >
    <input> ... </input>
    <output> ... </output>
    <port_map>
        <input external_port_id="0" internal_layer_id="0"/>
        <input external_port_id="1" internal_layer_id="1"/>
        <input external_port_id="-1" internal_layer_id="2" purpose="current_iteration"/>
        ...
        <output external_port_id="3" internal_layer_id="4"/>
        <output external_port_id="4" internal_layer_id="10" axis="1"/>
        <output external_port_id="-1" internal_layer_id="22" purpose="execution_condition"/>
        ...
    </port_map>
    <back_edges>
        <edge from-layer="1" to-layer="5"/>
        ...
    </back_edges>
    <body>
        <layers> ... </layers>
        <edges> ... </edges>
    </body>
</layer>