// Copyright (c) ONNX Project Contributors /* * SPDX-License-Identifier: Apache-2.0 */ #include "onnx/defs/controlflow/utils.h" #include "onnx/defs/schema.h" namespace ONNX_NAMESPACE { using SupportType = OpSchema::SupportType; static std::vector control_flow_types_ir9() { auto t = OpSchema::all_tensor_types_ir9(); auto s = OpSchema::all_tensor_sequence_types_ir9(); auto o = OpSchema::all_optional_types_ir9(); t.insert(t.end(), s.begin(), s.end()); t.insert(t.end(), o.begin(), o.end()); return t; } static std::vector control_flow_types_ir4() { auto t = OpSchema::all_tensor_types_ir4(); auto s = OpSchema::all_tensor_sequence_types_ir4(); auto o = OpSchema::all_optional_types_ir4(); t.insert(t.end(), s.begin(), s.end()); t.insert(t.end(), o.begin(), o.end()); return t; } ONNX_OPERATOR_SET_SCHEMA( If, 19, OpSchema() .SetDoc("If conditional") .Input(0, "cond", "Condition for the if. The tensor must contain a single element.", "B") .Output( 0, "outputs", "Values that are live-out to the enclosing scope. The return values in " "the `then_branch` and `else_branch` must be of the same data type. " "The `then_branch` and `else_branch` may produce tensors with the same " "element type and different shapes. " "If corresponding outputs from the then-branch and the else-branch have " "static shapes S1 and S2, then the shape of the corresponding output " "variable of the if-node (if present) must be compatible with both S1 " "and S2 as it represents the union of both possible shapes." "For example, if in a model file, the first " "output of `then_branch` is typed float tensor with shape [2] and the " "first output of `else_branch` is another float tensor with shape [3], " "If's first output should have (a) no shape set, or (b) " "a shape of rank 1 with neither `dim_value` nor `dim_param` set, or (c) " "a shape of rank 1 with a unique `dim_param`. " "In contrast, the first output cannot have the shape [2] since [2] and " "[3] are not compatible.", "V", OpSchema::Variadic, false) .Attr( "then_branch", "Graph to run if condition is true. Has N outputs: values you wish to " "be live-out to the enclosing scope. The number of outputs must match" " the number of outputs in the else_branch.", AttributeProto::GRAPH) .Attr( "else_branch", "Graph to run if condition is false. Has N outputs: values you wish to" " be live-out to the enclosing scope. The number of outputs must match" " the number of outputs in the then_branch.", AttributeProto::GRAPH) .TypeConstraint( "V", control_flow_types_ir9(), "All Tensor, Sequence(Tensor), Optional(Tensor), and Optional(Sequence(Tensor)) types up to IRv9.") .TypeConstraint("B", {"tensor(bool)"}, "Only bool") .TypeAndShapeInferenceFunction(IfInferenceFunction)); ONNX_OPERATOR_SET_SCHEMA( If, 16, OpSchema() .SetDoc("If conditional") .Input(0, "cond", "Condition for the if. The tensor must contain a single element.", "B") .Output( 0, "outputs", "Values that are live-out to the enclosing scope. The return values in " "the `then_branch` and `else_branch` must be of the same data type. " "The `then_branch` and `else_branch` may produce tensors with the same " "element type and different shapes. " "If corresponding outputs from the then-branch and the else-branch have " "static shapes S1 and S2, then the shape of the corresponding output " "variable of the if-node (if present) must be compatible with both S1 " "and S2 as it represents the union of both possible shapes." "For example, if in a model file, the first " "output of `then_branch` is typed float tensor with shape [2] and the " "first output of `else_branch` is another float tensor with shape [3], " "If's first output should have (a) no shape set, or (b) " "a shape of rank 1 with neither `dim_value` nor `dim_param` set, or (c) " "a shape of rank 1 with a unique `dim_param`. " "In contrast, the first output cannot have the shape [2] since [2] and " "[3] are not compatible.", "V", OpSchema::Variadic, false) .Attr( "then_branch", "Graph to run if condition is true. Has N outputs: values you wish to " "be live-out to the enclosing scope. The number of outputs must match" " the number of outputs in the else_branch.", AttributeProto::GRAPH) .Attr( "else_branch", "Graph to run if condition is false. Has N outputs: values you wish to" " be live-out to the enclosing scope. The number of outputs must match" " the number of outputs in the then_branch.", AttributeProto::GRAPH) .TypeConstraint( "V", control_flow_types_ir4(), "All Tensor, Sequence(Tensor), Optional(Tensor), and Optional(Sequence(Tensor)) types up to IRv4.") .TypeConstraint("B", {"tensor(bool)"}, "Only bool") .TypeAndShapeInferenceFunction(IfInferenceFunction)); static const char* Loop_ver16_doc = R"DOC( Generic Looping construct. This loop has multiple termination conditions: 1) Trip count. Iteration count specified at runtime. Set by specifying the input M. Optional. Set to empty string to omit. Note that a static trip count (specified at graph construction time) can be specified by passing in a constant node for input M. 2) Loop termination condition. This is an input to the op that determines whether to run the first iteration and also a loop-carried dependency for the body graph. The body graph must yield a value for the condition variable, whether this input is provided or not. This table summarizes the operating modes of this operator with equivalent C-style code: Operator inputs defined as (max_trip_count, condition_var). * input ("", ""): for (int i=0; ; ++i) { cond = ... // Note this value is ignored, but is required in the body } * input ("", cond) // Note this is analogous to a while loop bool cond = ...; for (int i=0; cond; ++i) { cond = ...; } * input ("", 1) // Note this is analogous to a do-while loop bool cond = true for (int i=0; cond; ++i) { cond = ...; } * input (trip_count, "") // Note this is analogous to a for loop int trip_count = ... for (int i=0; i < trip_count; ++i) { cond = ...; // ignored } * input (trip_count, cond) int trip_count = ...; bool cond = ...; for (int i=0; i < trip_count && cond; ++i) { cond = ...; } *Sample usage - cond as well as trip count* graph predict-net { %a = Constant[value = ]() %b = Constant[value = ]() %keepgoing = Constant[value = ]() %max_trip_count = Constant[value = ]() %keepgoing_out, %b_out, %user_defined_vals = Loop[body = ](%max_trip_count, %keepgoing, %b) return } graph body-net ( %i[INT32, scalar] // iteration number %keepgoing_in[BOOL, scalar] // incoming loop-termination-condition; not used %b_in[INT32, scalar] // incoming value of loop-carried-dependency b ) { %my_local = Add(%a, %b_in) %b_out = Sub(%a, %b_in) // outgoing value of loop-carried-dependency b %keepgoing_out = Greater(%my_local, %b_out) // outgoing loop-termination-condition %user_defined_val = Add(%b_in, %b_in) // scan-output value to be accumulated return %keepgoing_out, %b_out, %user_defined_val } *Sample equivalent C code* { /* User-defined code (enclosing scope) */ int a = 3, b = 6; bool keepgoing = true; // Analogous to input cond /* End user-defined code */ /* Implicitly-defined code */ const int max_trip_count = 10; // Analogous to input M int user_defined_vals[]; // Imagine this is resizable /* End implicitly-defined code */ /* initialize loop-carried variables and scan-output variables */ bool keepgoing_out = keepgoing int b_out = b for (int i=0; i < max_trip_count && keepgoing_out; ++i) { /* Implicitly-defined code: bind actual parameter values to formal parameter variables of loop-body */ bool keepgoing_in = keepgoing_out; bool b_in = b_out; /* User-defined code (loop body) */ int my_local = a + b_in; // Reading value "a" from the enclosing scope is fine b_out = a - b_in; keepgoing_out = my_local > b_out; user_defined_val = b_in + b_in; // b_in and b_out are different variables /* End user-defined code */ /* Implicitly defined-code */ user_defined_vals[i] = user_defined_val // accumulate scan-output values } // int t = my_local; // Can't do this. my_local is not accessible here. // The values below are bound to the output variables of the loop and therefore accessible // b_out; user_defined_vals; keepgoing_out; } There are several things of note in this code snippet: 1) Values from the enclosing scope (i.e. variable "a" here) are in scope and can be referenced in the inputs of the loop. 2) Any values computed in the loop body that needs to be used in a subsequent iteration or after the loop are modelled using a pair of variables in the loop-body, consisting of an input variable (eg., b_in) and an output variable (eg., b_out). These are referred to as loop-carried dependences. The loop operation node supplies the input value of the input variable for the first iteration, and returns the output value of the output variable produced by the final iteration. 3) Scan_output variables are used to implicitly concatenate values computed across all the iterations. In the above example, the value of user_defined_val computed over all iterations are concatenated and returned as the value of user_defined_vals after the loop. 4) Values created in the body cannot be accessed in the enclosing scope, except using the mechanism described above. Note that the semantics of this op support "diagonal" or "wavefront" execution. (See Step 3 here for an example: https://devblogs.nvidia.com/optimizing-recurrent-neural-networks-cudnn-5/). Frontends should emit multi-layer RNNs as a series of While operators (with time being the inner looping dimension), with each successive layer consuming the scan_outputs from the previous layer, possibly going through several point-wise operators (e.g. dropout, residual connections, linear layer). The input/output of subgraph (produced by loop node) matching is based on order instead of name. The implementation will figure out the names based on this order. )DOC"; ONNX_OPERATOR_SET_SCHEMA( Loop, 19, OpSchema() .SetDoc(Loop_ver16_doc) .Input( 0, "M", "A maximum trip-count for the loop specified at runtime. Optional." " Pass empty string to skip.", "I", OpSchema::Optional) .Input( 1, "cond", "A boolean termination condition. Optional. Pass empty string to skip.", "B", OpSchema::Optional) .Input( 2, "v_initial", "The initial values of any loop-carried dependencies (values that " "change across loop iterations)", "V", OpSchema::Variadic, false, 0) .Output( 0, "v_final_and_scan_outputs", "Final N loop carried dependency values then K scan_outputs. " "Scan outputs must be Tensors.", "V", OpSchema::Variadic, false) .Attr( "body", "The graph run each iteration. It has 2+N inputs: (iteration_num, " "condition, loop carried dependencies...). It has 1+N+K outputs: " "(condition, loop carried dependencies..., scan_outputs...). Each " "scan_output is created by concatenating the value of the specified " "output value at the end of each iteration of the loop. It is an error" " if the dimensions or data type of these scan_outputs change across loop" " iterations.", AttributeProto::GRAPH) .TypeConstraint( "V", control_flow_types_ir9(), "All Tensor, Sequence(Tensor), Optional(Tensor), and Optional(Sequence(Tensor)) types up to IRv9.") .TypeConstraint("I", {"tensor(int64)"}, "tensor of int64, which should be a scalar.") .TypeConstraint("B", {"tensor(bool)"}, "tensor of bool, which should be a scalar.") .TypeAndShapeInferenceFunction(LoopInferenceFunction)); ONNX_OPERATOR_SET_SCHEMA( Loop, 16, OpSchema() .SetDoc(Loop_ver16_doc) .Input( 0, "M", "A maximum trip-count for the loop specified at runtime. Optional." " Pass empty string to skip.", "I", OpSchema::Optional) .Input( 1, "cond", "A boolean termination condition. Optional. Pass empty string to skip.", "B", OpSchema::Optional) .Input( 2, "v_initial", "The initial values of any loop-carried dependencies (values that " "change across loop iterations)", "V", OpSchema::Variadic, false, 0) .Output( 0, "v_final_and_scan_outputs", "Final N loop carried dependency values then K scan_outputs. " "Scan outputs must be Tensors.", "V", OpSchema::Variadic, false) .Attr( "body", "The graph run each iteration. It has 2+N inputs: (iteration_num, " "condition, loop carried dependencies...). It has 1+N+K outputs: " "(condition, loop carried dependencies..., scan_outputs...). Each " "scan_output is created by concatenating the value of the specified " "output value at the end of each iteration of the loop. It is an error" " if the dimensions or data type of these scan_outputs change across loop" " iterations.", AttributeProto::GRAPH) .TypeConstraint( "V", control_flow_types_ir4(), "All Tensor, Sequence(Tensor), Optional(Tensor), and Optional(Sequence(Tensor)) types up to IRv4.") .TypeConstraint("I", {"tensor(int64)"}, "tensor of int64, which should be a scalar.") .TypeConstraint("B", {"tensor(bool)"}, "tensor of bool, which should be a scalar.") .TypeAndShapeInferenceFunction(LoopInferenceFunction)); static const char* scan_16_doc = R"DOC( Scan can be used to iterate over one or more scan_input tensors, constructing zero or more scan_output tensors. It combines ideas from general recurrences, functional programming constructs such as scan, fold, map, and zip, and is intended to enable generalizations of RNN-like constructs for sequence-to-sequence processing. Other tensors (referred to as state_variables here) can be used to carry a state when iterating from one element to another (similar to hidden-state in RNNs, also referred to as loop-carried dependences in the context of loops). Many common usages involve a single scan_input tensor (where functionality similar to scan, fold and map can be obtained). When more than one scan_input is used, a behavior similar to zip is obtained. The attribute body must be a graph, specifying the computation to be performed in every iteration. It takes as input the current values of the state_variables and the current iterated element of the scan_inputs. It must return the (updated) values of the state_variables and zero or more scan_output_element tensors. The values of the scan_output_element tensors are concatenated over all the iterations to produce the scan_output values of the scan construct (similar to the concatenated intermediate hidden-state values of RNN-like constructs). All the output tensors (state_variables as well as scan_output_element tensors) are required to have the same shape in each iteration of the loop (a restriction imposed to enable efficient memory allocation). Note that the iterated element passed to the body subgraph does not have a sequence axis. It will have a rank one less than the rank of the corresponding scan_input. The scan operation returns the final values of the state_variables as well as the scan_outputs. The optional attribute scan_input_directions specifies the direction (forward or backward) for each scan input. If this attribute is omitted, all sequences are scanned in the forward direction. A bidirectional scan may be performed by specifying the same tensor input twice in the scan_inputs, once with a forward direction, and once with a backward direction. The scan_output of the operation is produced by concatenating the scan_output_element values produced by the body in each iteration. The optional attribute scan_output_directions specifies the direction in which scan_output is constructed (by appending or prepending the scan_output_element to scan_output in each iteration) for each scan_output. If this attribute is omitted, the scan_output_element is appended to the scan_output in each iteration. The optional attribute scan_input_axes specifies the axis to be scanned for each scan_input. If omitted, every scan_input will be scanned in axis 0. For example, if axis 0 is the batch axis and axis 1 is the time axis (to be scanned), specify an axis value of 1. Note that scanning a non-zero axis may be less efficient than scanning axis zero. The optional attribute scan_output_axes specifies the axis along which the scan_outputs are accumulated for each scan_output. For example, if axis 1 is the time axis (to be scanned) for both inputs and outputs, specify a scan_input axis and scan_output axis value of 1. Note that because of the ONNX restriction that only the last parameter of an operator can be variadic, the initial-states and scan-inputs are listed together as one input parameter. Similarly, the final-states and scan-outputs are listed together as one output parameter. The attribute num_scan_inputs indicates the number M of scan-inputs. The behavior of Scan < num_scan_inputs = m, body = loop-body, scan_input_axes = [axis_1, ..., axis_m] > (init_1, ..., init_n, scan_1, ..., scan_m) is equivalent to the following pseudo-code: // scan_i.shape[axis_i] denotes the (max) sequence-length of scan_i // scan_i.shape[axis_i] is required to be equal to scan_j.shape[axis_j] for all i,j. sequence_length = scan_1.shape[axis_1]; // initialize state-variables st_1 = init_1; ... st_n = init_n; // initialize scan-output variables: [] denotes an empty tensor scan_out_1 = []; ...; scan_out_k = []; // identify number of iterations: // execute loop for (int t = 0; t < sequence_length; ++t) { // generate the scan-input elements: the notation T[t] indicates the sub-tensor // of rank one less than T obtained by indexing T at position t along axis k. si_1 = scan_1[t]; ... ; si_m = scan_m[t]; // execute loop-body st_1, ..., st_n, so_1, ..., so_k = loop-body(st_1, ..., st_n, si_1, ..., si_m) // accumulate the scan-output elements scan_out_1 = Concat(scan_out_1, so_1); ... ; scan_out_k = Concat(scan_out_k, so_k); } return st_1, ..., st_n, scan_out_1, ..., scan_out_k; *Sample usage: Encoding RNN using a Scan* The following example shows how a simple RNN over an input tensor %X, with weight tensor %Wi, recurrence weight tensor %Ri, bias tensors %Wbi and %Rbi, and initial hidden-state %H_0 can be encoded as a ScanLoop. Note that the loop-body is a nested graph, and it directly computes %Wi, %Ri, %Wbi, and %Rbi (typically constants or initializers in the body graph). If these values are computed in the outer graph, they need to be passed in as extra state_variables. graph rnn-encoding { %H_0 = ... %X = ... %Y_h, %Y = Scan[body = , num_scan_inputs=1](%H_0, %X) return %Y, %Y_h } graph rnn-cell-1 ( %H_tminus1[FLOAT, tensor] %X_t[FLOAT, tensor] ) { %Wi = ... %Ri = ... %Wbi = ... %Rbi = ... %t1 = X_t * (Wi^T) %t2 = H_tminus1*(Ri^T) %t3 = Add(%t1, %t2) %t4 = Add(%t3, %Wbi) %t5 = Add(%t4, %Rbi) %Ht = Tanh(%t5) %Accumulate = Identity(%Ht) return %Ht, %Accumulate } )DOC"; ONNX_OPERATOR_SET_SCHEMA( Scan, 19, OpSchema() .SetDoc(scan_16_doc) .Input( 0, "initial_state_and_scan_inputs", "Initial values of the loop's N state variables followed by M scan_inputs", "V", OpSchema::Variadic, false) .Output( 0, "final_state_and_scan_outputs", "Final values of the loop's N state variables followed by K scan_outputs", "V", OpSchema::Variadic, false) .Attr( "body", "The graph run each iteration. It has N+M inputs: " "(loop state variables..., scan_input_elts...). It has N+K outputs: " "(loop state variables..., scan_output_elts...). Each " "scan_output is created by concatenating the value of the specified " "scan_output_elt value at the end of each iteration of the loop. It is an error" " if the dimensions of these values change across loop iterations.", AttributeProto::GRAPH, true) .Attr("num_scan_inputs", "An attribute specifying the number of scan_inputs M. ", AttributeProto::INT, true) .Attr( "scan_input_directions", "An optional list of M flags. The i-th element of the list specifies the direction " "to be scanned for the i-th scan_input tensor: 0 indicates forward direction and 1 " "indicates reverse direction. " "If omitted, all scan_input tensors will be scanned in the forward direction.", AttributeProto::INTS, false) .Attr( "scan_output_directions", "An optional list of K flags, one for each scan_output. The i-th element of the list " "specifies whether the i-th scan_output should be constructed by appending or " "prepending a new value in each iteration: 0 indicates appending and 1 " "indicates prepending. " "If omitted, all scan_output tensors will be produced by appending a value " "in each iteration.", AttributeProto::INTS, false) .Attr( "scan_input_axes", "An optional list of M flags. The i-th element of the list specifies the axis " "to be scanned (the sequence axis) for the i-th scan_input. If omitted, 0 will " "be used as the scan axis for every scan_input. Negative value for an axis means " "counting dimensions from the back. Accepted range is [-r, r-1] where r = rank(input).", AttributeProto::INTS, false) .Attr( "scan_output_axes", "An optional list of K flags. The i-th element of the list specifies the axis " "for the i-th scan_output. The scan outputs are accumulated along the specified " "axis. If omitted, 0 will be used as the scan axis for every scan_output. " "Negative value for an axis means counting dimensions from the back. Accepted " "range is [-r, r-1].", AttributeProto::INTS, false) .TypeConstraint("V", OpSchema::all_tensor_types_ir9(), "All Tensor types up to IRv9.") .TypeAndShapeInferenceFunction(ScanInferenceFunction)); // Shares same shape inference as opset 11 ONNX_OPERATOR_SET_SCHEMA( Scan, 16, OpSchema() .SetDoc(scan_16_doc) .Input( 0, "initial_state_and_scan_inputs", "Initial values of the loop's N state variables followed by M scan_inputs", "V", OpSchema::Variadic, false) .Output( 0, "final_state_and_scan_outputs", "Final values of the loop's N state variables followed by K scan_outputs", "V", OpSchema::Variadic, false) .Attr( "body", "The graph run each iteration. It has N+M inputs: " "(loop state variables..., scan_input_elts...). It has N+K outputs: " "(loop state variables..., scan_output_elts...). Each " "scan_output is created by concatenating the value of the specified " "scan_output_elt value at the end of each iteration of the loop. It is an error" " if the dimensions of these values change across loop iterations.", AttributeProto::GRAPH, true) .Attr("num_scan_inputs", "An attribute specifying the number of scan_inputs M. ", AttributeProto::INT, true) .Attr( "scan_input_directions", "An optional list of M flags. The i-th element of the list specifies the direction " "to be scanned for the i-th scan_input tensor: 0 indicates forward direction and 1 " "indicates reverse direction. " "If omitted, all scan_input tensors will be scanned in the forward direction.", AttributeProto::INTS, false) .Attr( "scan_output_directions", "An optional list of K flags, one for each scan_output. The i-th element of the list " "specifies whether the i-th scan_output should be constructed by appending or " "prepending a new value in each iteration: 0 indicates appending and 1 " "indicates prepending. " "If omitted, all scan_output tensors will be produced by appending a value " "in each iteration.", AttributeProto::INTS, false) .Attr( "scan_input_axes", "An optional list of M flags. The i-th element of the list specifies the axis " "to be scanned (the sequence axis) for the i-th scan_input. If omitted, 0 will " "be used as the scan axis for every scan_input. Negative value for an axis means " "counting dimensions from the back. Accepted range is [-r, r-1] where r = rank(input).", AttributeProto::INTS, false) .Attr( "scan_output_axes", "An optional list of K flags. The i-th element of the list specifies the axis " "for the i-th scan_output. The scan outputs are accumulated along the specified " "axis. If omitted, 0 will be used as the scan axis for every scan_output. " "Negative value for an axis means counting dimensions from the back. Accepted " "range is [-r, r-1].", AttributeProto::INTS, false) .TypeConstraint("V", OpSchema::all_tensor_types_ir4(), "All Tensor types up to IRv4.") .TypeAndShapeInferenceFunction(ScanInferenceFunction)); // Shares same shape inference as opset 11 void ScanInferenceFunctionOpset8(InferenceContext& ctx) { // NOTE: // The first input to Scan is sequence_lens. We skip that when processing // inputs in many places below, so the - 1 in multiple places is due to that. auto num_inputs = ctx.getNumInputs(); auto num_scan_inputs = narrow_cast(ctx.getAttribute("num_scan_inputs")->i()); auto num_loop_state_vars = num_inputs - 1 - num_scan_inputs; std::vector temporary_type_protos; temporary_type_protos.reserve(num_inputs); std::vector subgraph_input_types; TensorShapeProto_Dimension batch_size_dim; TensorShapeProto_Dimension sequence_len_dim; for (size_t i = 1; i < num_inputs; ++i) { bool is_loop_state_var = (i - 1) < num_loop_state_vars; bool has_shape = hasInputShape(ctx, i); const auto* input_type = ctx.getInputType(i); // Enforce type constraint for inputs if (!input_type || !input_type->has_tensor_type()) { fail_type_inference("Scan input ", i, " was not a tensor."); } if (is_loop_state_var) { // If it's a loop state variable we can propagate type and shape 1:1 to // the matching Scan output. // We can also pass through the type and shape to the subgraph but need to // remove the batch size dimension from the shape. propagateElemTypeFromInputToOutput(ctx, i, i - 1); if (has_shape) { propagateShapeFromInputToOutput(ctx, i, i - 1); // remove batch size dimension and add to subgraph_input_types temporary_type_protos.push_back(RemoveDimensionsFromShape(*input_type, 1)); subgraph_input_types.push_back(&temporary_type_protos.back()); } else { subgraph_input_types.push_back(input_type); } } else { // For other inputs there is no fixed relationships to the Scan outputs, // so we don't propagate type/shape information. // We can pass through the type and shape to the subgraph inputs but need // to remove the batch size and sequence length dimensions from the shape. if (has_shape) { // remove batch size and sequence length dimensions and add to // subgraph_input_types temporary_type_protos.push_back(RemoveDimensionsFromShape(*input_type, 2)); subgraph_input_types.push_back(&temporary_type_protos.back()); // update batch_size and sequence_len if a value is available const auto& shape = input_type->tensor_type().shape(); if (shape.dim_size() > 2) { const auto& dims = shape.dim(); mergeInDimensionInfo(dims.Get(0), batch_size_dim, 0); mergeInDimensionInfo(dims.Get(1), sequence_len_dim, 1); } } else { subgraph_input_types.push_back(input_type); } } } // Run inferencing on the subgraph std::vector output_types; GraphInferencer* graphInferencer = ctx.getGraphAttributeInferencer("body"); if (graphInferencer) { std::vector input_data; for (size_t i = 1; i < num_inputs; ++i) { input_data.push_back(ctx.getInputData(i)); } output_types = graphInferencer->doInferencing(subgraph_input_types, input_data); } // if empty(), assume inferencing was skipped if (!output_types.empty()) { auto num_outputs = ctx.getNumOutputs(); if (output_types.size() != num_outputs) { fail_type_inference( "Graph attribute inferencing returned type information for ", output_types.size(), " outputs. Expected ", num_outputs); } // propagate type/shape information for loop state variables and outputs for (size_t i = 0; i < num_outputs; ++i) { const bool is_loop_state_var = i < num_loop_state_vars; auto* subgraph_output_type = output_types[i]; auto* scan_output_type = ctx.getOutputType(i); if (!subgraph_output_type->has_tensor_type()) { fail_type_inference("Scan 'body' subgraph outputs should all be tensors but output ", i, " was not"); } // propagate output type. loop state vars were done in the above code. if (!is_loop_state_var) { scan_output_type->mutable_tensor_type()->set_elem_type(subgraph_output_type->tensor_type().elem_type()); } // propagate shape if (subgraph_output_type->tensor_type().has_shape()) { // we need to add in the batch size and sequence length values if // available before merging with any existing info. Create a copy of the // inferred type info from the subgraph to do that. TypeProto inferred_type(*subgraph_output_type); auto* mutable_inferred_tensor_type = inferred_type.mutable_tensor_type(); auto* mutable_inferred_shape = mutable_inferred_tensor_type->mutable_shape(); mutable_inferred_shape->clear_dim(); *mutable_inferred_shape->add_dim() = batch_size_dim; if (!is_loop_state_var) { *mutable_inferred_shape->add_dim() = sequence_len_dim; } for (const auto& dim : subgraph_output_type->tensor_type().shape().dim()) { (*mutable_inferred_shape->add_dim()) = dim; } auto* mutable_scan_output_tensor_type = scan_output_type->mutable_tensor_type(); mergeInShapeInfo(*mutable_inferred_tensor_type, *mutable_scan_output_tensor_type); } } } } int handle_negative_axis_validate_opset9(const std::string& attrib, int axis, int rank) { if (!(-rank <= axis && axis < rank)) { fail_shape_inference(attrib, " axis value ", axis, " is invalid for a tensor of rank ", rank); } return (axis >= 0 ? axis : axis + rank); } void ScanInferenceFunctionOpset9(InferenceContext& ctx) { auto num_inputs = ctx.getNumInputs(); auto num_scan_inputs = narrow_cast(ctx.getAttribute("num_scan_inputs")->i()); auto num_loop_state_vars = num_inputs - num_scan_inputs; auto num_outputs = ctx.getNumOutputs(); auto num_scan_outputs = num_outputs - num_loop_state_vars; std::vector axes, output_axes; if (getRepeatedAttribute(ctx, "scan_input_axes", axes)) { if (axes.size() != num_scan_inputs) { fail_shape_inference( "Number of scan input axes specified (", axes.size(), ") is not equal to number of scan inputs (", num_scan_inputs, ")."); } } else { axes.insert(axes.end(), num_scan_inputs, 0); } if (getRepeatedAttribute(ctx, "scan_output_axes", output_axes)) { if (output_axes.size() != num_scan_outputs) { fail_shape_inference( "Number of scan output axes specified (", output_axes.size(), ") is not equal to number of scan outputs (", num_scan_outputs, ")."); } } else { output_axes.insert(output_axes.end(), num_scan_outputs, 0); } std::vector temporary_type_protos; temporary_type_protos.reserve(num_inputs); std::vector subgraph_input_types; TensorShapeProto_Dimension sequence_len_dim; for (size_t i = 0; i < num_inputs; ++i) { bool is_loop_state_var = i < num_loop_state_vars; bool has_shape = hasInputShape(ctx, i); const auto* input_type = ctx.getInputType(i); // Enforce type constraint for inputs if (!input_type || !input_type->has_tensor_type()) { fail_type_inference("Scan input ", i, " was not a tensor."); } if (is_loop_state_var) { // If it's a loop state variable we can propagate type and shape 1:1 to // the matching Scan output. // We can also pass through the type and shape to the subgraph but need to // remove the batch size dimension from the shape. propagateElemTypeFromInputToOutput(ctx, i, i); if (has_shape) propagateShapeFromInputToOutput(ctx, i, i); subgraph_input_types.push_back(input_type); } else { // For other inputs there is no fixed relationships to the Scan outputs, // so we don't propagate type/shape information. // We can pass through the type and shape to the subgraph inputs but // need to remove the sequence length dimensions from the shape. if (has_shape) { const auto& shape = input_type->tensor_type().shape(); // remove sequence length dimensions and add to subgraph_input_types int axis = static_cast(axes[i - num_loop_state_vars]); axis = handle_negative_axis_validate_opset9("scan_input_axes", axis, shape.dim_size()); // update sequence_len if a value is available const auto& dims = shape.dim(); mergeInDimensionInfo(dims.Get(axis), sequence_len_dim, 1); temporary_type_protos.push_back(RemoveIthDimensionFromShape(*input_type, axis)); subgraph_input_types.push_back(&temporary_type_protos.back()); } else { subgraph_input_types.push_back(input_type); } } } // Run inferencing on the subgraph std::vector output_types; GraphInferencer* graphInferencer = ctx.getGraphAttributeInferencer("body"); if (graphInferencer) { std::vector input_data; for (size_t i = 0; i < num_inputs; ++i) { // ctx.getInputData(i), the input to scan, does not represent the input to // scan body. So, we pass in null, to represent an unknown value. input_data.push_back(nullptr); } output_types = graphInferencer->doInferencing(subgraph_input_types, input_data); } // if empty(), assume inferencing was skipped if (!output_types.empty()) { if (output_types.size() != num_outputs) { fail_type_inference( "Graph attribute inferencing returned type information for ", output_types.size(), " outputs. Expected ", num_outputs); } // propagate type/shape information for loop state variables and outputs for (size_t i = 0; i < num_outputs; ++i) { const bool is_loop_state_var = i < num_loop_state_vars; auto* subgraph_output_type = output_types[i]; auto* scan_output_type = ctx.getOutputType(i); auto* mutable_scan_output_tensor_type = scan_output_type->mutable_tensor_type(); if (!subgraph_output_type->has_tensor_type()) { fail_type_inference("Scan 'body' subgraph outputs should all be tensors but output ", i, " was not"); } auto& subgraph_output_tensor_type = subgraph_output_type->tensor_type(); if (is_loop_state_var) { // merge shape; type already propagated mergeInShapeInfo(subgraph_output_tensor_type, *mutable_scan_output_tensor_type); } else { scan_output_type->mutable_tensor_type()->set_elem_type(subgraph_output_tensor_type.elem_type()); // propagate shape if (subgraph_output_tensor_type.has_shape()) { // infer shape of scan-output from the shape of scan-output-element // by adding sequence-length at the correct axis position const TensorShapeProto& subgraph_output_shape = subgraph_output_tensor_type.shape(); TensorShapeProto inferred_shape; auto subgraph_output_rank = subgraph_output_shape.dim_size(); auto output_rank = subgraph_output_rank + 1; int output_axis = static_cast(output_axes[i - num_loop_state_vars]); output_axis = handle_negative_axis_validate_opset9("scan_output_axes", output_axis, output_rank); for (int j = 0; j < output_axis; ++j) *(inferred_shape.add_dim()) = subgraph_output_shape.dim(j); *(inferred_shape.add_dim()) = sequence_len_dim; for (int j = output_axis; j < subgraph_output_rank; ++j) *(inferred_shape.add_dim()) = subgraph_output_shape.dim(j); // Merge inferred shape with existing shape information mergeInShapeInfo(inferred_shape, *mutable_scan_output_tensor_type); } } } } } static const char* scan_opset8_doc = R"DOC( Scan can be used to iterate over one or more scan_input tensors, constructing zero or more scan_output tensors. It combines ideas from general recurrences, functional programming constructs such as scan, fold, map, and zip, and is intended to enable generalizations of RNN-like constructs for sequence-to-sequence processing. Other tensors (referred to as state_variables here) can be used to carry a state when iterating from one element to another (similar to hidden-state in RNNs, also referred to as loop-carried dependences in the context of loops). All these tensors are required to have the same shape in each iteration of the loop (a restriction imposed to enable efficient memory allocation). Many common usages involve a single scan_input tensor (where functionality similar to scan, fold and map can be obtained). When more than one scan_input is used, a behavior similar to zip is obtained. The attribute body must be a graph, specifying the computation to be performed in every iteration. It takes as input the current values of the state_variables and the current iterated element of the scan_inputs. It must return the (updated) values of the state_variables and zero or more scan_output_element tensors. The values of the scan_output_element tensors are concatenated over all the iterations to produce the scan_output values of the scan construct (similar to the concatenated intermediate hidden-state values of RNN-like constructs). The scan operation returns the final values of the state_variables as well as the scan_outputs. The operation supports batching, and the batch-axis is required to be 0. When multiple scan_input tensors are used, they must all have the same batch-size, and they must all have the same maximum-sequence-length (the dimensionality of the sequence axis or scan axis). The sequence axis or scan axis is required to be 1. The operation has an optional sequence_lens input (of shape [BATCH_SIZE]) to allow variable length sequences of length <= the maximum-sequence-length. If this input is not specified, all sequences are assumed to be of length equal to maximum-sequence-length. For variable length input sequences, the scan_outputs will consist of a sequence of same length as the input, padded to the maximum-sequence-length. The optional attribute directions can be used to scan a sequence in the reverse direction. If this attribute is omitted, all sequences are scanned in the forward direction. A bidirectional scan be performed by specifying the same tensor input twice in the scan_inputs, once with a forward direction, and once with a backward direction. Note that because of the ONNX restriction that only the last parameter of an operator can be variadic, the initial-states and scan-inputs are listed together as one input parameter. Similarly, the final-states and scan-outputs are listed together as one output parameter. The attribute num_scan_inputs indicates the number M of scan-inputs. The behavior of Scan < num_scan_inputs = m, body = loop-body > (sequence_lengths, init_1, ..., init_n, scan_1, ..., scan_m) is equivalent to the following pseudo-code: // T.shape[0] denotes the batch-size of T // The batch-size of scan_1, ..., scan_m are all required to be equal batch_size = scan_1.shape[0]; // scan_i.shape[1] denotes the (max) sequence-length of scan_i // scan_i.shape[1] is required to be equal to scan_j.shape[1] for all i,j. max_sequence_length = scan_1.shape[1]; for (int batch = 0; batch < batch_size; ++batch) { // initialize state-variables st_1 = init_1; ... st_n = init_n; // initialize scan-output variables: [] denotes an empty tensor scan_out_1 = []; ...; scan_out_k = []; // identify number of iterations: N = (sequence_lengths specified) ? sequence_lengths[batch] : max_sequence_length; // execute loop for (int t = 0; t < N; ++t) { // generate the scan-input elements: the notation T[t] indicates the sub-tensor // of rank one less than T obtained by indexing T at position t along axis k. si_1 = (scan_1[batch])[t]; ... ; si_m = (scan_m[batch])[t]; // execute loop-body st_1, ..., st_n, so_1, ..., so_k = loop-body(st_1, ..., st_n, si_1, ..., si_m) // accumulate the scan-output elements scan_out_1 = Concat(scan_out_1, so_1); ... ; scan_out_k = Concat(scan_out_k, so_k); } // accumulate the outputs for this batch: bst_1[batch] = st_1; ..., bst_n[batch] = st_n; // Note scan-outputs will have size max_sequence_length, but only first N values will be meaningful. // The remaining values have an undefined value. b_scan_out_1[batch] = scan_out_1; ...; b_scan_out_k[batch] = scan_out_k; } return bst_1, ..., bst_n, b_scan_out_1, ..., b_scan_out_k; *Sample usage: Encoding RNN using a Scan* The following example shows how a simple RNN over an input tensor %X, with weight tensor %Wi, recurrence weight tensor %Ri, bias tensors %Wbi and %Rbi, and initial hidden-state %H_0 can be encoded as a ScanLoop. Note that the loop-body is a nested graph, and it directly computes %Wi, %Ri, %Wbi, and %Rbi (typically constants or initializers in the body graph). If these values are computed in the outer graph, they need to be passed in as extra state_variables. graph rnn-encoding { %H_0 = ... %X = ... %Y_h, %Y = Scan[body = , num_scan_inputs=1]("", %H_0, %X) return %Y, %Y_h } graph rnn-cell-1 ( %H_tminus1[FLOAT, tensor] %X_t[FLOAT, tensor] ) { %Wi = ... %Ri = ... %Wbi = ... %Rbi = ... %t1 = X_t * (Wi^T) %t2 = H_tminus1*(Ri^T) %t3 = Add(%t1, %t2) %t4 = Add(%t3, %Wbi) %t5 = Add(%t4, %Rbi) %Ht = Tanh(%t5) %Accumulate = Identity(%Ht) return %Ht, %Accumulate } )DOC"; ONNX_OPERATOR_SET_SCHEMA( Scan, 8, OpSchema() .SetDoc(scan_opset8_doc) .Input( 0, "sequence_lens", "Optional tensor specifying lengths of the sequences in a batch. " "If this input is not specified, all sequences are assumed to be of " "the maximum sequence length (the dimension of the sequence axis of " "the scan_input tensors).", "I", OpSchema::Optional) .Input( 1, "initial_state_and_scan_inputs", "Initial values of the loop's N state variables followed by M scan_inputs", "V", OpSchema::Variadic, false) .Output( 0, "final_state_and_scan_outputs", "Final values of the loop's N state variables followed by K scan_outputs", "V", OpSchema::Variadic, false) .Attr( "body", "The graph run each iteration. It has N+M inputs: " "(loop state variables..., scan_input_elts...). It has N+K outputs: " "(loop state variables..., scan_output_elts...). Each " "scan_output is created by concatenating the value of the specified " "scan_output_elt value at the end of each iteration of the loop. It is an error" " if the dimensions of these values change across loop iterations.", AttributeProto::GRAPH, true) .Attr("num_scan_inputs", "An attribute specifying the number of scan_inputs M. ", AttributeProto::INT, true) .Attr( "directions", "An optional list of M flags. The i-th element of the list specifies the direction " "to be scanned for the i-th scan_input tensor: 0 indicates forward direction and 1 " "indicates reverse direction. " "If omitted, all scan_input tensors will be scanned in the forward direction.", AttributeProto::INTS, false) .TypeConstraint("I", {"tensor(int64)"}, "Int64 tensor") .TypeConstraint("V", OpSchema::all_tensor_types(), "All Tensor types") .TypeAndShapeInferenceFunction(ScanInferenceFunctionOpset8)); void LoopInferenceFunctionOpset8(InferenceContext& ctx) { auto num_inputs = ctx.getNumInputs(); auto num_loop_state_vars = num_inputs - 2; // skip 'M' and 'cond' std::vector subgraph_input_types; std::vector temporary_type_protos; temporary_type_protos.reserve(num_inputs - 2); // create TypeProto to validate iteration number type is the same as the // optional 'M' input for max iterations. TypeProto iter_num_type; iter_num_type.mutable_tensor_type()->set_elem_type(TensorProto_DataType_INT64); subgraph_input_types.push_back(&iter_num_type); // 'cond' subgraph_input_types.push_back(ctx.getInputType(1)); // loop state value types get propagated to outputs, but shape may change // across iterations so don't propagate it to the outputs and don't pass it // into the subgraph inferencing for (size_t i = 2; i < num_inputs; ++i) { propagateElemTypeFromInputToOutput(ctx, i, i - 2); // copy so we can remove the shape before passing to the subgraph // inferencing temporary_type_protos.push_back(*ctx.getInputType(i)); auto& input_type = temporary_type_protos.back(); input_type.mutable_tensor_type()->clear_shape(); subgraph_input_types.push_back(&input_type); } // Run inferencing on the subgraph std::vector subgraph_output_types; GraphInferencer* graphInferencer = ctx.getGraphAttributeInferencer("body"); if (graphInferencer) { std::vector input_data; input_data.push_back(nullptr); // iteration number for (size_t i = 1; i < num_inputs; ++i) { input_data.push_back(ctx.getInputData(i)); } subgraph_output_types = graphInferencer->doInferencing(subgraph_input_types, input_data); } // if empty(), assume inferencing was skipped if (!subgraph_output_types.empty()) { auto num_outputs = ctx.getNumOutputs(); // subgraph outputs the condition value first but that is only used // internally and not returned by Loop. if (subgraph_output_types.size() != num_outputs + 1) { fail_type_inference( "Graph attribute inferencing returned type information for ", subgraph_output_types.size(), " outputs. Expected ", num_outputs + 1); } // check loop state values match. we should already have type/shape info for (size_t i = 0; i < num_outputs; ++i) { auto* subgraph_output_type = subgraph_output_types[i + 1]; // skip 'cond' auto* loop_output_type = ctx.getOutputType(i); const bool is_loop_state_var = i < num_loop_state_vars; if (!subgraph_output_type->has_tensor_type()) { fail_type_inference( "Loop 'body' subgraph outputs should all be tensors but output ", i, " was ", subgraph_output_type->value_case()); } // if there's an existing type check it matches. otherwise propagate propagateElemTypeWithValidation(subgraph_output_type, loop_output_type); if (is_loop_state_var) { // shape may change across iterations so ignore. } else { // propagate shape if (subgraph_output_type->tensor_type().has_shape()) { // per iteration output. first dimension will be number of iterations // but we don't know that value yet TypeProto inferred_type(*subgraph_output_type); auto* mutable_inferred_tensor_type = inferred_type.mutable_tensor_type(); auto* mutable_inferred_shape = mutable_inferred_tensor_type->mutable_shape(); mutable_inferred_shape->clear_dim(); // add empty dimension for number of iterations mutable_inferred_shape->add_dim(); // add dimensions from subgraph output shape for (const auto& dim : subgraph_output_type->tensor_type().shape().dim()) { (*mutable_inferred_shape->add_dim()) = dim; } mergeInShapeInfo(*mutable_inferred_tensor_type, *loop_output_type->mutable_tensor_type()); } } } } } static const char* Loop_ver1_doc = R"DOC( Generic Looping construct. This loop has multiple termination conditions: 1) Trip count. Iteration count specified at runtime. Set by specifying the input M. Optional. Set to empty string to omit. Note that a static trip count (specified at graph construction time) can be specified by passing in a constant node for input M. 2) Loop termination condition. This is an input to the op that determines whether to run the first iteration and also a loop-carried dependency for the body graph. The body graph must yield a value for the condition variable, whether this input is provided or not. This table summarizes the operating modes of this operator with equivalent C-style code: Operator inputs defined as (max_trip_count, condition_var). input ("", ""): for (int i=0; ; ++i) { cond = ... // Note this value is ignored, but is required in the body } input ("", cond) // Note this is analogous to a while loop bool cond = ...; for (int i=0; cond; ++i) { cond = ...; } input ("", 1) // Note this is analogous to a do-while loop bool cond = true for (int i=0; cond; ++i) { cond = ...; } input (trip_count, "") // Note this is analogous to a for loop int trip_count = ... for (int i=0; i < trip_count; ++i) { cond = ...; // ignored } input (trip_count, cond) int trip_count = ...; bool cond = ...; for (int i=0; i < trip_count && cond; ++i) { cond = ...; } *Sample usage - cond as well as trip count* graph predict-net { %a = Constant[value = ]() %b = Constant[value = ]() %keepgoing = Constant[value = ]() %max_trip_count = Constant[value = ]() %keepgoing_out, %b_out, %user_defined_vals = Loop[body = ](%max_trip_count, %keepgoing, %b) return } graph body-net ( %i[INT32, scalar] %keepgoing[BOOL, scalar] %b[INT32, scalar] ) { %my_local = Add(%a, %b) %b_out = Sub(%a, %b) %keepgoing_out = Greater(%my_local, %b_out) %user_defined_vals = Add(%b, %b) return %keepgoing_out, %b_out, %user_defined_vals } *Sample equivalent C code* { /* User-defined code (enclosing scope) */ int a = 3, b = 6; bool keepgoing = true; // Analogous to input cond /* End user-defined code */ /* Implicitly-defined code */ const int max_trip_count = 10; // Analogous to input M int user_defined_vals[]; // Imagine this is resizable /* End implicitly-defined code */ for (int i=0; i < max_trip_count && keepgoing; ++i) { /* User-defined code (loop body) */ int my_local = a + b; // Reading values in the enclosing scope is fine b = a - b; // writes fine if we specify b as a loop-carried dependency keepgoing = my_local > b; // keepgoing is a loop-carried dependency user_defined_vals[i] = b + b; /* End user-defined code */ } // my_local = 123; // Can't do this. my_local was defined in the body // These below values are live-out from the loop and therefore accessible b_out; user_defined_vals; keepgoing_out; } There are several things of note in this code snippet: 1) Values from the enclosing scope (i.e. variable a here) are in scope and can be referenced in the inputs of the loop. 2) Any variables which you wish to make available in the enclosing scope (i.e. the variables b and keepgoing) must be declared as either loop-carried dependencies (both at the op inputs and output and at the body net input and output) or scan_outputs. 3) Values created in the body cannot be accessed in the enclosing scope. Note that the semantics of this op support "diagonal" or "wavefront" execution. (See Step 3 here for an example: https://devblogs.nvidia.com/optimizing-recurrent-neural-networks-cudnn-5/). Frontends should emit multi-layer RNNs as a series of While operators (with time being the inner looping dimension), with each successive layer consuming the scan_outputs from the previous layer, possibly going through several point-wise operators (e.g. dropout, residual connections, linear layer). )DOC"; ONNX_OPERATOR_SET_SCHEMA( Loop, 1, OpSchema() .SetDoc(Loop_ver1_doc) .Input( 0, "M", "A maximum trip-count for the loop specified at runtime. Optional." " Pass empty string to skip.", "I", OpSchema::Optional) .Input( 1, "cond", "A boolean termination condition. Optional. Pass empty string to skip.", "B", OpSchema::Optional) .Input( 2, "v_initial", "The initial values of any loop-carried dependencies (values that " "change across loop iterations)", "V", OpSchema::Variadic, false) .Output( 0, "v_final_and_scan_outputs", "Final N loop carried dependency values then K scan_outputs", "V", OpSchema::Variadic, false) .Attr( "body", "The graph run each iteration. It has 2+N inputs: (iteration_num, " "condition, loop carried dependencies...). It has 1+N+K outputs: " "(condition, loop carried dependencies..., scan_outputs...). Each " "scan_output is created by concatenating the value of the specified " "output value at the end of each iteration of the loop. It is an error" " if the dimensions or data type of these scan_outputs change across loop" " iterations.", AttributeProto::GRAPH) .TypeConstraint("V", OpSchema::all_tensor_types(), "All Tensor types") .TypeConstraint("I", {"tensor(int64)"}, "tensor of int64, which should be a scalar.") .TypeConstraint("B", {"tensor(bool)"}, "tensor of bool, which should be a scalar.") .TypeAndShapeInferenceFunction(LoopInferenceFunctionOpset8)); void LoopInferenceFunctionOpset11(InferenceContext& ctx) { auto num_inputs = ctx.getNumInputs(); auto num_loop_state_vars = num_inputs - 2; // skip 'M' and 'cond' std::vector subgraph_input_types; std::vector temporary_type_protos; temporary_type_protos.reserve(num_inputs - 2); // create TypeProto to validate iteration number type is the same as the // optional 'M' input for max iterations. TypeProto iter_num_type; iter_num_type.mutable_tensor_type()->set_elem_type(TensorProto_DataType_INT64); subgraph_input_types.push_back(&iter_num_type); // 'cond' subgraph_input_types.push_back(ctx.getInputType(1)); // loop state value types get propagated to outputs, but shape may change // across iterations so don't propagate it to the outputs and don't pass it // into the subgraph inferencing for (size_t i = 2; i < num_inputs; ++i) { propagateElemTypeFromInputToOutput(ctx, i, i - 2); // copy so we can remove the shape before passing to the subgraph // inferencing temporary_type_protos.push_back(*ctx.getInputType(i)); auto& input_type = temporary_type_protos.back(); input_type.mutable_tensor_type()->clear_shape(); subgraph_input_types.push_back(&input_type); } // Run inferencing on the subgraph std::vector subgraph_output_types; GraphInferencer* graphInferencer = ctx.getGraphAttributeInferencer("body"); if (graphInferencer) { std::vector input_data; input_data.push_back(nullptr); // iteration number for (size_t i = 1; i < num_inputs; ++i) { input_data.push_back(ctx.getInputData(i)); } subgraph_output_types = graphInferencer->doInferencing(subgraph_input_types, input_data); } // if empty(), assume inferencing was skipped if (!subgraph_output_types.empty()) { auto num_outputs = ctx.getNumOutputs(); // subgraph outputs the condition value first but that is only used // internally and not returned by Loop. if (subgraph_output_types.size() != num_outputs + 1) { fail_type_inference( "Graph attribute inferencing returned type information for ", subgraph_output_types.size(), " outputs. Expected ", num_outputs + 1); } // check loop state values match. we should already have type/shape info for (size_t i = 0; i < num_outputs; ++i) { auto* subgraph_output_type = subgraph_output_types[i + 1]; // skip 'cond' auto* loop_output_type = ctx.getOutputType(i); const bool is_loop_state_var = i < num_loop_state_vars; if (!subgraph_output_type->has_tensor_type()) { fail_type_inference( "Loop 'body' subgraph outputs should all be tensors but output ", i, " was ", subgraph_output_type->value_case()); } // if there's an existing type check it matches. otherwise propagate propagateElemTypeWithValidation(subgraph_output_type, loop_output_type); if (is_loop_state_var) { // shape may change across iterations so ignore. } else { // propagate shape if (subgraph_output_type->tensor_type().has_shape()) { // per iteration output. first dimension will be number of iterations // but we don't know that value yet TypeProto inferred_type(*subgraph_output_type); auto* mutable_inferred_tensor_type = inferred_type.mutable_tensor_type(); auto* mutable_inferred_shape = mutable_inferred_tensor_type->mutable_shape(); mutable_inferred_shape->clear_dim(); // add empty dimension for number of iterations mutable_inferred_shape->add_dim(); // add dimensions from subgraph output shape for (const auto& dim : subgraph_output_type->tensor_type().shape().dim()) { (*mutable_inferred_shape->add_dim()) = dim; } mergeInShapeInfo(*mutable_inferred_tensor_type, *loop_output_type->mutable_tensor_type()); } } } } } static const char* Loop_ver11_doc = R"DOC( Generic Looping construct. This loop has multiple termination conditions: 1) Trip count. Iteration count specified at runtime. Set by specifying the input M. Optional. Set to empty string to omit. Note that a static trip count (specified at graph construction time) can be specified by passing in a constant node for input M. 2) Loop termination condition. This is an input to the op that determines whether to run the first iteration and also a loop-carried dependency for the body graph. The body graph must yield a value for the condition variable, whether this input is provided or not. This table summarizes the operating modes of this operator with equivalent C-style code: Operator inputs defined as (max_trip_count, condition_var). input ("", ""): for (int i=0; ; ++i) { cond = ... // Note this value is ignored, but is required in the body } input ("", cond) // Note this is analogous to a while loop bool cond = ...; for (int i=0; cond; ++i) { cond = ...; } input ("", 1) // Note this is analogous to a do-while loop bool cond = true for (int i=0; cond; ++i) { cond = ...; } input (trip_count, "") // Note this is analogous to a for loop int trip_count = ... for (int i=0; i < trip_count; ++i) { cond = ...; // ignored } input (trip_count, cond) int trip_count = ...; bool cond = ...; for (int i=0; i < trip_count && cond; ++i) { cond = ...; } *Sample usage - cond as well as trip count* graph predict-net { %a = Constant[value = ]() %b = Constant[value = ]() %keepgoing = Constant[value = ]() %max_trip_count = Constant[value = ]() %keepgoing_out, %b_out, %user_defined_vals = Loop[body = ](%max_trip_count, %keepgoing, %b) return } graph body-net ( %i[INT32, scalar] // iteration number %keepgoing_in[BOOL, scalar] // incoming loop-termination-condition; not used %b_in[INT32, scalar] // incoming value of loop-carried-dependency b ) { %my_local = Add(%a, %b_in) %b_out = Sub(%a, %b_in) // outgoing value of loop-carried-dependency b %keepgoing_out = Greater(%my_local, %b_out) // outgoing loop-termination-condition %user_defined_val = Add(%b_in, %b_in) // scan-output value to be accumulated return %keepgoing_out, %b_out, %user_defined_val } *Sample equivalent C code* { /* User-defined code (enclosing scope) */ int a = 3, b = 6; bool keepgoing = true; // Analogous to input cond /* End user-defined code */ /* Implicitly-defined code */ const int max_trip_count = 10; // Analogous to input M int user_defined_vals[]; // Imagine this is resizable /* End implicitly-defined code */ /* initialize loop-carried variables and scan-output variables */ bool keepgoing_out = keepgoing int b_out = b for (int i=0; i < max_trip_count && keepgoing_out; ++i) { /* Implicitly-defined code: bind actual parameter values to formal parameter variables of loop-body */ bool keepgoing_in = keepgoing_out; bool b_in = b_out; /* User-defined code (loop body) */ int my_local = a + b_in; // Reading value "a" from the enclosing scope is fine b_out = a - b_in; keepgoing_out = my_local > b_out; user_defined_val = b_in + b_in; // b_in and b_out are different variables /* End user-defined code */ /* Implicitly defined-code */ user_defined_vals[i] = user_defined_val // accumulate scan-output values } // int t = my_local; // Can't do this. my_local is not accessible here. // The values below are bound to the output variables of the loop and therefore accessible // b_out; user_defined_vals; keepgoing_out; } There are several things of note in this code snippet: 1) Values from the enclosing scope (i.e. variable "a" here) are in scope and can be referenced in the inputs of the loop. 2) Any values computed in the loop body that needs to be used in a subsequent iteration or after the loop are modelled using a pair of variables in the loop-body, consisting of an input variable (eg., b_in) and an output variable (eg., b_out). These are referred to as loop-carried dependences. The loop operation node supplies the input value of the input variable for the first iteration, and returns the output value of the output variable produced by the final iteration. 3) Scan_output variables are used to implicitly concatenate values computed across all the iterations. In the above example, the value of user_defined_val computed over all iterations are concatenated and returned as the value of user_defined_vals after the loop. 4) Values created in the body cannot be accessed in the enclosing scope, except using the mechanism described above. Note that the semantics of this op support "diagonal" or "wavefront" execution. (See Step 3 here for an example: https://devblogs.nvidia.com/optimizing-recurrent-neural-networks-cudnn-5/). Frontends should emit multi-layer RNNs as a series of While operators (with time being the inner looping dimension), with each successive layer consuming the scan_outputs from the previous layer, possibly going through several point-wise operators (e.g. dropout, residual connections, linear layer). )DOC"; ONNX_OPERATOR_SET_SCHEMA( Loop, 11, OpSchema() .SetDoc(Loop_ver11_doc) .Input( 0, "M", "A maximum trip-count for the loop specified at runtime. Optional." " Pass empty string to skip.", "I", OpSchema::Optional) .Input( 1, "cond", "A boolean termination condition. Optional. Pass empty string to skip.", "B", OpSchema::Optional) .Input( 2, "v_initial", "The initial values of any loop-carried dependencies (values that " "change across loop iterations)", "V", OpSchema::Variadic, false, 0) .Output( 0, "v_final_and_scan_outputs", "Final N loop carried dependency values then K scan_outputs", "V", OpSchema::Variadic, false) .Attr( "body", "The graph run each iteration. It has 2+N inputs: (iteration_num, " "condition, loop carried dependencies...). It has 1+N+K outputs: " "(condition, loop carried dependencies..., scan_outputs...). Each " "scan_output is created by concatenating the value of the specified " "output value at the end of each iteration of the loop. It is an error" " if the dimensions or data type of these scan_outputs change across loop" " iterations.", AttributeProto::GRAPH) .TypeConstraint("V", OpSchema::all_tensor_types(), "All Tensor types") .TypeConstraint("I", {"tensor(int64)"}, "tensor of int64, which should be a scalar.") .TypeConstraint("B", {"tensor(bool)"}, "tensor of bool, which should be a scalar.") .TypeAndShapeInferenceFunction(LoopInferenceFunctionOpset11)); static const char* scan_9_doc = R"DOC( Scan can be used to iterate over one or more scan_input tensors, constructing zero or more scan_output tensors. It combines ideas from general recurrences, functional programming constructs such as scan, fold, map, and zip, and is intended to enable generalizations of RNN-like constructs for sequence-to-sequence processing. Other tensors (referred to as state_variables here) can be used to carry a state when iterating from one element to another (similar to hidden-state in RNNs, also referred to as loop-carried dependences in the context of loops). Many common usages involve a single scan_input tensor (where functionality similar to scan, fold and map can be obtained). When more than one scan_input is used, a behavior similar to zip is obtained. The attribute body must be a graph, specifying the computation to be performed in every iteration. It takes as input the current values of the state_variables and the current iterated element of the scan_inputs. It must return the (updated) values of the state_variables and zero or more scan_output_element tensors. The values of the scan_output_element tensors are concatenated over all the iterations to produce the scan_output values of the scan construct (similar to the concatenated intermediate hidden-state values of RNN-like constructs). All the output tensors (state_variables as well as scan_output_element tensors) are required to have the same shape in each iteration of the loop (a restriction imposed to enable efficient memory allocation). Note that the iterated element passed to the body subgraph does not have a sequence axis. It will have a rank one less than the rank of the corresponding scan_input. The scan operation returns the final values of the state_variables as well as the scan_outputs. The optional attribute scan_input_directions specifies the direction (forward or backward) for each scan input. If this attribute is omitted, all sequences are scanned in the forward direction. A bidirectional scan may be performed by specifying the same tensor input twice in the scan_inputs, once with a forward direction, and once with a backward direction. The scan_output of the operation is produced by concatenating the scan_output_element values produced by the body in each iteration. The optional attribute scan_output_directions specifies the direction in which scan_output is constructed (by appending or prepending the scan_output_element to scan_output in each iteration) for each scan_output. If this attribute is omitted, the scan_output_element is appended to the scan_output in each iteration. The optional attribute scan_input_axes specifies the axis to be scanned for each scan_input. If omitted, every scan_input will be scanned in axis 0. For example, if axis 0 is the batch axis and axis 1 is the time axis (to be scanned), specify an axis value of 1. Note that scanning a non-zero axis may be less efficient than scanning axis zero. The optional attribute scan_output_axes specifies the axis along which the scan_outputs are accumulated for each scan_output. For example, if axis 1 is the time axis (to be scanned) for both inputs and outputs, specify a scan_input axis and scan_output axis value of 1. Note that because of the ONNX restriction that only the last parameter of an operator can be variadic, the initial-states and scan-inputs are listed together as one input parameter. Similarly, the final-states and scan-outputs are listed together as one output parameter. The attribute num_scan_inputs indicates the number M of scan-inputs. The behavior of Scan < num_scan_inputs = m, body = loop-body, scan_input_axes = [axis_1, ..., axis_m] > (init_1, ..., init_n, scan_1, ..., scan_m) is equivalent to the following pseudo-code: // scan_i.shape[axis_i] denotes the (max) sequence-length of scan_i // scan_i.shape[axis_i] is required to be equal to scan_j.shape[axis_j] for all i,j. sequence_length = scan_1.shape[axis_1]; // initialize state-variables st_1 = init_1; ... st_n = init_n; // initialize scan-output variables: [] denotes an empty tensor scan_out_1 = []; ...; scan_out_k = []; // identify number of iterations: // execute loop for (int t = 0; t < sequence_length; ++t) { // generate the scan-input elements: the notation T[t] indicates the sub-tensor // of rank one less than T obtained by indexing T at position t along axis k. si_1 = scan_1[t]; ... ; si_m = scan_m[t]; // execute loop-body st_1, ..., st_n, so_1, ..., so_k = loop-body(st_1, ..., st_n, si_1, ..., si_m) // accumulate the scan-output elements scan_out_1 = Concat(scan_out_1, so_1); ... ; scan_out_k = Concat(scan_out_k, so_k); } return st_1, ..., st_n, scan_out_1, ..., scan_out_k; *Sample usage: Encoding RNN using a Scan* The following example shows how a simple RNN over an input tensor %X, with weight tensor %Wi, recurrence weight tensor %Ri, bias tensors %Wbi and %Rbi, and initial hidden-state %H_0 can be encoded as a ScanLoop. Note that the loop-body is a nested graph, and it directly computes %Wi, %Ri, %Wbi, and %Rbi (typically constants or initializers in the body graph). If these values are computed in the outer graph, they need to be passed in as extra state_variables. graph rnn-encoding { %H_0 = ... %X = ... %Y_h, %Y = Scan[body = , num_scan_inputs=1](%H_0, %X) return %Y, %Y_h } graph rnn-cell-1 ( %H_tminus1[FLOAT, tensor] %X_t[FLOAT, tensor] ) { %Wi = ... %Ri = ... %Wbi = ... %Rbi = ... %t1 = X_t * (Wi^T) %t2 = H_tminus1*(Ri^T) %t3 = Add(%t1, %t2) %t4 = Add(%t3, %Wbi) %t5 = Add(%t4, %Rbi) %Ht = Tanh(%t5) %Accumulate = Identity(%Ht) return %Ht, %Accumulate } )DOC"; ONNX_OPERATOR_SET_SCHEMA( Scan, 9, OpSchema() .SetDoc(scan_9_doc) .Input( 0, "initial_state_and_scan_inputs", "Initial values of the loop's N state variables followed by M scan_inputs", "V", OpSchema::Variadic, false) .Output( 0, "final_state_and_scan_outputs", "Final values of the loop's N state variables followed by K scan_outputs", "V", OpSchema::Variadic, false) .Attr( "body", "The graph run each iteration. It has N+M inputs: " "(loop state variables..., scan_input_elts...). It has N+K outputs: " "(loop state variables..., scan_output_elts...). Each " "scan_output is created by concatenating the value of the specified " "scan_output_elt value at the end of each iteration of the loop. It is an error" " if the dimensions of these values change across loop iterations.", AttributeProto::GRAPH, true) .Attr("num_scan_inputs", "An attribute specifying the number of scan_inputs M. ", AttributeProto::INT, true) .Attr( "scan_input_directions", "An optional list of M flags. The i-th element of the list specifies the direction " "to be scanned for the i-th scan_input tensor: 0 indicates forward direction and 1 " "indicates reverse direction. " "If omitted, all scan_input tensors will be scanned in the forward direction.", AttributeProto::INTS, false) .Attr( "scan_output_directions", "An optional list of K flags, one for each scan_output. The i-th element of the list " "specifies whether the i-th scan_output should be constructed by appending or " "prepending a new value in each iteration: 0 indicates appending and 1 " "indicates prepending. " "If omitted, all scan_output tensors will be produced by appending a value " "in each iteration.", AttributeProto::INTS, false) .Attr( "scan_input_axes", "An optional list of M flags. The i-th element of the list specifies the axis " "to be scanned (the sequence axis) for the i-th scan_input. If omitted, 0 will " "be used as the scan axis for every scan_input.", AttributeProto::INTS, false) .Attr( "scan_output_axes", "An optional list of K flags. The i-th element of the list specifies the axis " "for the i-th scan_output. The scan outputs are accumulated along the specified " "axis. If omitted, 0 will be used as the scan axis for every scan_output.", AttributeProto::INTS, false) .TypeConstraint("V", OpSchema::all_tensor_types(), "All Tensor types") .TypeAndShapeInferenceFunction(ScanInferenceFunctionOpset9)); void IfInferenceFunction1(InferenceContext& ctx) { // there are no inputs so we just need to run the subgraph inferencing for // then/else subgraphs and apply those to the outputs. std::vector subgraph_input_types; // none std::vector input_data; // none std::vector then_output_types; std::vector else_output_types; // Run inferencing on the subgraph GraphInferencer* graphInferencer = ctx.getGraphAttributeInferencer("then_branch"); if (graphInferencer) { then_output_types = graphInferencer->doInferencing(subgraph_input_types, input_data); } graphInferencer = ctx.getGraphAttributeInferencer("else_branch"); if (graphInferencer) { else_output_types = graphInferencer->doInferencing(subgraph_input_types, input_data); } auto num_outputs = ctx.getNumOutputs(); auto num_then_outputs = then_output_types.size(); auto num_else_outputs = else_output_types.size(); // the output types for then and else should be the same if (num_then_outputs != num_else_outputs) { fail_type_inference( "then_branch and else_branch produce different number of outputs. ", num_then_outputs, " != ", num_else_outputs); } if (num_then_outputs != num_outputs) { fail_type_inference("If node has ", num_outputs, " but subgraphs produce ", num_then_outputs); } for (size_t i = 0, end = then_output_types.size(); i < end; ++i) { auto then_output = then_output_types[i]; auto else_output = else_output_types[i]; if (then_output->value_case() != else_output->value_case()) { fail_type_inference( "Mismatched type for output ", i, " then=", then_output->value_case(), " else=", else_output->value_case()); } auto* if_output = ctx.getOutputType(i); *if_output = *then_output; if (then_output->has_tensor_type()) { auto then_elem_type = then_output->tensor_type().elem_type(); auto else_elem_type = else_output->tensor_type().elem_type(); if (then_elem_type != else_elem_type) { fail_type_inference( "Mismatched tensor element type for output ", i, " then=", then_elem_type, " else=", else_elem_type); } // merge the 'else' shape information to check it's consistent and // augment the 'if' output if possible mergeInShapeInfo(else_output->tensor_type(), *if_output->mutable_tensor_type()); } } } ONNX_OPERATOR_SET_SCHEMA( If, 1, OpSchema() .SetDoc("If conditional") .Input(0, "cond", "Condition for the if. The tensor must contain a single element.", "B") .Output( 0, "outputs", "Values that are live-out to the enclosing scope. The return values in " "the `then_branch` and `else_branch` must be of the same shape and same " "data type.", "V", OpSchema::Variadic, false) .Attr( "then_branch", "Graph to run if condition is true. Has N outputs: values you wish to " "be live-out to the enclosing scope. The number of outputs must match" " the number of outputs in the else_branch.", AttributeProto::GRAPH) .Attr( "else_branch", "Graph to run if condition is false. Has N outputs: values you wish to" " be live-out to the enclosing scope. The number of outputs must match" " the number of outputs in the then_branch.", AttributeProto::GRAPH) .TypeConstraint("V", OpSchema::all_tensor_types(), "All Tensor types") .TypeConstraint("B", {"tensor(bool)"}, "Only bool") .TypeAndShapeInferenceFunction(IfInferenceFunction1)); void IfInferenceFunction_11(InferenceContext& ctx) { // there are no inputs so we just need to run the subgraph inferencing for // then/else subgraphs and apply those to the outputs. std::vector subgraph_input_types; // none std::vector input_data; // none std::vector then_output_types; std::vector else_output_types; // Run inferencing on the subgraph GraphInferencer* graphInferencer = ctx.getGraphAttributeInferencer("then_branch"); if (graphInferencer) { then_output_types = graphInferencer->doInferencing(subgraph_input_types, input_data); } graphInferencer = ctx.getGraphAttributeInferencer("else_branch"); if (graphInferencer) { else_output_types = graphInferencer->doInferencing(subgraph_input_types, input_data); } auto num_outputs = ctx.getNumOutputs(); auto num_then_outputs = then_output_types.size(); auto num_else_outputs = else_output_types.size(); // the output types for then and else should be the same if (num_then_outputs != num_else_outputs) { fail_type_inference( "then_branch and else_branch produce different number of outputs. ", num_then_outputs, " != ", num_else_outputs); } if (num_then_outputs != num_outputs) { fail_type_inference("If node has ", num_outputs, " but subgraphs produce ", num_then_outputs); } for (size_t i = 0, end = then_output_types.size(); i < end; ++i) { auto then_output = then_output_types[i]; auto else_output = else_output_types[i]; if (then_output->value_case() != else_output->value_case()) { fail_type_inference( "Mismatched type for output ", i, " then=", then_output->value_case(), " else=", else_output->value_case()); } auto* if_output = ctx.getOutputType(i); *if_output = *then_output; if (then_output->has_tensor_type()) { auto then_elem_type = then_output->tensor_type().elem_type(); auto else_elem_type = else_output->tensor_type().elem_type(); if (then_elem_type != else_elem_type) { fail_type_inference( "Mismatched tensor element type for output ", i, " then=", then_elem_type, " else=", else_elem_type); } UnionShapeInfo(else_output->tensor_type().shape(), *if_output->mutable_tensor_type()); } } } ONNX_OPERATOR_SET_SCHEMA( If, 11, OpSchema() .SetDoc("If conditional") .Input(0, "cond", "Condition for the if. The tensor must contain a single element.", "B") .Output( 0, "outputs", "Values that are live-out to the enclosing scope. The return values in " "the `then_branch` and `else_branch` must be of the same data type. " "The `then_branch` and `else_branch` may produce tensors with the same " "element type and different shapes. " "If corresponding outputs from the then-branch and the else-branch have " "static shapes S1 and S2, then the shape of the corresponding output " "variable of the if-node (if present) must be compatible with both S1 " "and S2 as it represents the union of both possible shapes." "For example, if in a model file, the first " "output of `then_branch` is typed float tensor with shape [2] and the " "first output of `else_branch` is another float tensor with shape [3], " "If's first output should have (a) no shape set, or (b) " "a shape of rank 1 with neither `dim_value` nor `dim_param` set, or (c) " "a shape of rank 1 with a unique `dim_param`. " "In contrast, the first output cannot have the shape [2] since [2] and " "[3] are not compatible.", "V", OpSchema::Variadic, false) .Attr( "then_branch", "Graph to run if condition is true. Has N outputs: values you wish to " "be live-out to the enclosing scope. The number of outputs must match" " the number of outputs in the else_branch.", AttributeProto::GRAPH) .Attr( "else_branch", "Graph to run if condition is false. Has N outputs: values you wish to" " be live-out to the enclosing scope. The number of outputs must match" " the number of outputs in the then_branch.", AttributeProto::GRAPH) .TypeConstraint("V", OpSchema::all_tensor_types(), "All Tensor types") .TypeConstraint("B", {"tensor(bool)"}, "Only bool") .TypeAndShapeInferenceFunction(IfInferenceFunction_11)); void IfInferenceFunction_13(InferenceContext& ctx) { // there are no inputs so we just need to run the subgraph inferencing for // then/else subgraphs and apply those to the outputs. std::vector subgraph_input_types; // none std::vector input_data; // none std::vector then_output_types; std::vector else_output_types; // Run inferencing on the subgraph GraphInferencer* graphInferencer = ctx.getGraphAttributeInferencer("then_branch"); if (graphInferencer) { then_output_types = graphInferencer->doInferencing(subgraph_input_types, input_data); } graphInferencer = ctx.getGraphAttributeInferencer("else_branch"); if (graphInferencer) { else_output_types = graphInferencer->doInferencing(subgraph_input_types, input_data); } auto num_outputs = ctx.getNumOutputs(); auto num_then_outputs = then_output_types.size(); auto num_else_outputs = else_output_types.size(); // the output types for then and else should be the same if (num_then_outputs != num_else_outputs) { fail_type_inference( "then_branch and else_branch produce different number of outputs. ", num_then_outputs, " != ", num_else_outputs); } if (num_then_outputs != num_outputs) { fail_type_inference("If node has ", num_outputs, " but subgraphs produce ", num_then_outputs); } for (size_t i = 0, end = then_output_types.size(); i < end; ++i) { auto then_output = then_output_types[i]; auto else_output = else_output_types[i]; auto* if_output = ctx.getOutputType(i); *if_output = *then_output; UnionTypeInfo(*else_output, *if_output); } } ONNX_OPERATOR_SET_SCHEMA( If, 13, OpSchema() .SetDoc("If conditional") .Input(0, "cond", "Condition for the if. The tensor must contain a single element.", "B") .Output( 0, "outputs", "Values that are live-out to the enclosing scope. The return values in " "the `then_branch` and `else_branch` must be of the same data type. " "The `then_branch` and `else_branch` may produce tensors with the same " "element type and different shapes. " "If corresponding outputs from the then-branch and the else-branch have " "static shapes S1 and S2, then the shape of the corresponding output " "variable of the if-node (if present) must be compatible with both S1 " "and S2 as it represents the union of both possible shapes." "For example, if in a model file, the first " "output of `then_branch` is typed float tensor with shape [2] and the " "first output of `else_branch` is another float tensor with shape [3], " "If's first output should have (a) no shape set, or (b) " "a shape of rank 1 with neither `dim_value` nor `dim_param` set, or (c) " "a shape of rank 1 with a unique `dim_param`. " "In contrast, the first output cannot have the shape [2] since [2] and " "[3] are not compatible.", "V", OpSchema::Variadic, false) .Attr( "then_branch", "Graph to run if condition is true. Has N outputs: values you wish to " "be live-out to the enclosing scope. The number of outputs must match" " the number of outputs in the else_branch.", AttributeProto::GRAPH) .Attr( "else_branch", "Graph to run if condition is false. Has N outputs: values you wish to" " be live-out to the enclosing scope. The number of outputs must match" " the number of outputs in the then_branch.", AttributeProto::GRAPH) .TypeConstraint( "V", []() { auto t = OpSchema::all_tensor_types(); auto s = OpSchema::all_tensor_sequence_types(); t.insert(t.end(), s.begin(), s.end()); return t; }(), "All Tensor and Sequence types") .TypeConstraint("B", {"tensor(bool)"}, "Only bool") .TypeAndShapeInferenceFunction(IfInferenceFunction_13)); void LoopInferenceFunction_13(InferenceContext& ctx) { auto num_inputs = ctx.getNumInputs(); assert(num_inputs >= 2); auto num_loop_state_vars = num_inputs - 2; // skip 'M' and 'cond' std::vector subgraph_input_types; subgraph_input_types.reserve(num_inputs); std::vector temporary_type_protos; temporary_type_protos.reserve(num_inputs - 2); // create TypeProto to validate iteration number type is the same as the // optional 'M' input for max iterations. TypeProto iter_num_type; iter_num_type.mutable_tensor_type()->set_elem_type(TensorProto_DataType_INT64); subgraph_input_types.push_back(&iter_num_type); // 'cond' subgraph_input_types.push_back(ctx.getInputType(1)); // loop state value types get propagated to outputs, but shape may change // across iterations so don't propagate it to the outputs and don't pass it // into the subgraph inferencing for (size_t i = 2; i < num_inputs; ++i) { propagateElemTypeFromInputToOutput(ctx, i, i - 2); // copy so we can remove the shape before passing to the subgraph // inferencing temporary_type_protos.push_back(*ctx.getInputType(i)); auto& input_type = temporary_type_protos.back(); if (input_type.has_tensor_type()) { input_type.mutable_tensor_type()->clear_shape(); } else if (input_type.has_sequence_type()) { auto& seq_type = *input_type.mutable_sequence_type(); if (seq_type.has_elem_type() && seq_type.elem_type().has_tensor_type()) { seq_type.mutable_elem_type()->mutable_tensor_type()->clear_shape(); } } subgraph_input_types.push_back(&input_type); } // Run inferencing on the subgraph std::vector subgraph_output_types; GraphInferencer* graphInferencer = ctx.getGraphAttributeInferencer("body"); if (graphInferencer) { std::vector input_data; input_data.push_back(nullptr); // iteration number for (size_t i = 1; i < num_inputs; ++i) { input_data.push_back(ctx.getInputData(i)); } subgraph_output_types = graphInferencer->doInferencing(subgraph_input_types, input_data); } // if empty(), assume inferencing was skipped if (!subgraph_output_types.empty()) { auto num_outputs = ctx.getNumOutputs(); // subgraph outputs the condition value first but that is only used // internally and not returned by Loop. if (subgraph_output_types.size() != num_outputs + 1) { fail_type_inference( "Graph attribute inferencing returned type information for ", subgraph_output_types.size(), " outputs. Expected ", num_outputs + 1); } // check loop state values match. we should already have type/shape info for (size_t i = 0; i < num_outputs; ++i) { auto* subgraph_output_type = subgraph_output_types[i + 1]; // skip 'cond' auto* loop_output_type = ctx.getOutputType(i); const bool is_loop_state_var = i < num_loop_state_vars; if (!subgraph_output_type->has_tensor_type() && !subgraph_output_type->has_sequence_type()) { fail_type_inference( "Loop 'body' subgraph outputs should all be tensors or sequences but output ", i, " was ", subgraph_output_type->value_case()); } if (!is_loop_state_var && !subgraph_output_type->has_tensor_type()) { fail_type_inference( "Loop 'body' subgraph scan outputs should all be tensors but output ", i, " was ", subgraph_output_type->value_case()); } // if there's an existing type check it matches. otherwise propagate propagateElemTypeWithValidation(subgraph_output_type, loop_output_type); if (is_loop_state_var) { // shape may change across iterations so ignore. } else { // propagate shape if (subgraph_output_type->tensor_type().has_shape()) { // per iteration output. first dimension will be number of iterations // but we don't know that value yet TypeProto inferred_type(*subgraph_output_type); auto* mutable_inferred_tensor_type = inferred_type.mutable_tensor_type(); auto* mutable_inferred_shape = mutable_inferred_tensor_type->mutable_shape(); mutable_inferred_shape->clear_dim(); // add empty dimension for number of iterations mutable_inferred_shape->add_dim(); // add dimensions from subgraph output shape for (const auto& dim : subgraph_output_type->tensor_type().shape().dim()) { (*mutable_inferred_shape->add_dim()) = dim; } mergeInShapeInfo(*mutable_inferred_tensor_type, *loop_output_type->mutable_tensor_type()); } } } } } static const char* Loop_ver13_doc = R"DOC( Generic Looping construct. This loop has multiple termination conditions: 1) Trip count. Iteration count specified at runtime. Set by specifying the input M. Optional. Set to empty string to omit. Note that a static trip count (specified at graph construction time) can be specified by passing in a constant node for input M. 2) Loop termination condition. This is an input to the op that determines whether to run the first iteration and also a loop-carried dependency for the body graph. The body graph must yield a value for the condition variable, whether this input is provided or not. This table summarizes the operating modes of this operator with equivalent C-style code: Operator inputs defined as (max_trip_count, condition_var). input ("", ""): for (int i=0; ; ++i) { cond = ... // Note this value is ignored, but is required in the body } input ("", cond) // Note this is analogous to a while loop bool cond = ...; for (int i=0; cond; ++i) { cond = ...; } input ("", 1) // Note this is analogous to a do-while loop bool cond = true for (int i=0; cond; ++i) { cond = ...; } input (trip_count, "") // Note this is analogous to a for loop int trip_count = ... for (int i=0; i < trip_count; ++i) { cond = ...; // ignored } input (trip_count, cond) int trip_count = ...; bool cond = ...; for (int i=0; i < trip_count && cond; ++i) { cond = ...; } *Sample usage - cond as well as trip count* graph predict-net { %a = Constant[value = ]() %b = Constant[value = ]() %keepgoing = Constant[value = ]() %max_trip_count = Constant[value = ]() %keepgoing_out, %b_out, %user_defined_vals = Loop[body = ](%max_trip_count, %keepgoing, %b) return } graph body-net ( %i[INT32, scalar] // iteration number %keepgoing_in[BOOL, scalar] // incoming loop-termination-condition; not used %b_in[INT32, scalar] // incoming value of loop-carried-dependency b ) { %my_local = Add(%a, %b_in) %b_out = Sub(%a, %b_in) // outgoing value of loop-carried-dependency b %keepgoing_out = Greater(%my_local, %b_out) // outgoing loop-termination-condition %user_defined_val = Add(%b_in, %b_in) // scan-output value to be accumulated return %keepgoing_out, %b_out, %user_defined_val } *Sample equivalent C code* { /* User-defined code (enclosing scope) */ int a = 3, b = 6; bool keepgoing = true; // Analogous to input cond /* End user-defined code */ /* Implicitly-defined code */ const int max_trip_count = 10; // Analogous to input M int user_defined_vals[]; // Imagine this is resizable /* End implicitly-defined code */ /* initialize loop-carried variables and scan-output variables */ bool keepgoing_out = keepgoing int b_out = b for (int i=0; i < max_trip_count && keepgoing_out; ++i) { /* Implicitly-defined code: bind actual parameter values to formal parameter variables of loop-body */ bool keepgoing_in = keepgoing_out; bool b_in = b_out; /* User-defined code (loop body) */ int my_local = a + b_in; // Reading value "a" from the enclosing scope is fine b_out = a - b_in; keepgoing_out = my_local > b_out; user_defined_val = b_in + b_in; // b_in and b_out are different variables /* End user-defined code */ /* Implicitly defined-code */ user_defined_vals[i] = user_defined_val // accumulate scan-output values } // int t = my_local; // Can't do this. my_local is not accessible here. // The values below are bound to the output variables of the loop and therefore accessible // b_out; user_defined_vals; keepgoing_out; } There are several things of note in this code snippet: 1) Values from the enclosing scope (i.e. variable "a" here) are in scope and can be referenced in the inputs of the loop. 2) Any values computed in the loop body that needs to be used in a subsequent iteration or after the loop are modelled using a pair of variables in the loop-body, consisting of an input variable (eg., b_in) and an output variable (eg., b_out). These are referred to as loop-carried dependences. The loop operation node supplies the input value of the input variable for the first iteration, and returns the output value of the output variable produced by the final iteration. 3) Scan_output variables are used to implicitly concatenate values computed across all the iterations. In the above example, the value of user_defined_val computed over all iterations are concatenated and returned as the value of user_defined_vals after the loop. 4) Values created in the body cannot be accessed in the enclosing scope, except using the mechanism described above. Note that the semantics of this op support "diagonal" or "wavefront" execution. (See Step 3 here for an example: https://devblogs.nvidia.com/optimizing-recurrent-neural-networks-cudnn-5/). Frontends should emit multi-layer RNNs as a series of While operators (with time being the inner looping dimension), with each successive layer consuming the scan_outputs from the previous layer, possibly going through several point-wise operators (e.g. dropout, residual connections, linear layer). The input/output of subgraph (produced by loop node) matching is based on order instead of name. The implementation will figure out the names based on this order. )DOC"; ONNX_OPERATOR_SET_SCHEMA( Loop, 13, OpSchema() .SetDoc(Loop_ver13_doc) .Input( 0, "M", "A maximum trip-count for the loop specified at runtime. Optional." " Pass empty string to skip.", "I", OpSchema::Optional) .Input( 1, "cond", "A boolean termination condition. Optional. Pass empty string to skip.", "B", OpSchema::Optional) .Input( 2, "v_initial", "The initial values of any loop-carried dependencies (values that " "change across loop iterations)", "V", OpSchema::Variadic, false, 0) .Output( 0, "v_final_and_scan_outputs", "Final N loop carried dependency values then K scan_outputs. " "Scan outputs must be Tensors.", "V", OpSchema::Variadic, false) .Attr( "body", "The graph run each iteration. It has 2+N inputs: (iteration_num, " "condition, loop carried dependencies...). It has 1+N+K outputs: " "(condition, loop carried dependencies..., scan_outputs...). Each " "scan_output is created by concatenating the value of the specified " "output value at the end of each iteration of the loop. It is an error" " if the dimensions or data type of these scan_outputs change across loop" " iterations.", AttributeProto::GRAPH) .TypeConstraint( "V", []() { auto t = OpSchema::all_tensor_types(); auto s = OpSchema::all_tensor_sequence_types(); t.insert(t.end(), s.begin(), s.end()); return t; }(), "All Tensor and Sequence types") .TypeConstraint("I", {"tensor(int64)"}, "tensor of int64, which should be a scalar.") .TypeConstraint("B", {"tensor(bool)"}, "tensor of bool, which should be a scalar.") .TypeAndShapeInferenceFunction(LoopInferenceFunction_13)); static const char* scan_11_doc = R"DOC( Scan can be used to iterate over one or more scan_input tensors, constructing zero or more scan_output tensors. It combines ideas from general recurrences, functional programming constructs such as scan, fold, map, and zip, and is intended to enable generalizations of RNN-like constructs for sequence-to-sequence processing. Other tensors (referred to as state_variables here) can be used to carry a state when iterating from one element to another (similar to hidden-state in RNNs, also referred to as loop-carried dependences in the context of loops). Many common usages involve a single scan_input tensor (where functionality similar to scan, fold and map can be obtained). When more than one scan_input is used, a behavior similar to zip is obtained. The attribute body must be a graph, specifying the computation to be performed in every iteration. It takes as input the current values of the state_variables and the current iterated element of the scan_inputs. It must return the (updated) values of the state_variables and zero or more scan_output_element tensors. The values of the scan_output_element tensors are concatenated over all the iterations to produce the scan_output values of the scan construct (similar to the concatenated intermediate hidden-state values of RNN-like constructs). All the output tensors (state_variables as well as scan_output_element tensors) are required to have the same shape in each iteration of the loop (a restriction imposed to enable efficient memory allocation). Note that the iterated element passed to the body subgraph does not have a sequence axis. It will have a rank one less than the rank of the corresponding scan_input. The scan operation returns the final values of the state_variables as well as the scan_outputs. The optional attribute scan_input_directions specifies the direction (forward or backward) for each scan input. If this attribute is omitted, all sequences are scanned in the forward direction. A bidirectional scan may be performed by specifying the same tensor input twice in the scan_inputs, once with a forward direction, and once with a backward direction. The scan_output of the operation is produced by concatenating the scan_output_element values produced by the body in each iteration. The optional attribute scan_output_directions specifies the direction in which scan_output is constructed (by appending or prepending the scan_output_element to scan_output in each iteration) for each scan_output. If this attribute is omitted, the scan_output_element is appended to the scan_output in each iteration. The optional attribute scan_input_axes specifies the axis to be scanned for each scan_input. If omitted, every scan_input will be scanned in axis 0. For example, if axis 0 is the batch axis and axis 1 is the time axis (to be scanned), specify an axis value of 1. Note that scanning a non-zero axis may be less efficient than scanning axis zero. The optional attribute scan_output_axes specifies the axis along which the scan_outputs are accumulated for each scan_output. For example, if axis 1 is the time axis (to be scanned) for both inputs and outputs, specify a scan_input axis and scan_output axis value of 1. Note that because of the ONNX restriction that only the last parameter of an operator can be variadic, the initial-states and scan-inputs are listed together as one input parameter. Similarly, the final-states and scan-outputs are listed together as one output parameter. The attribute num_scan_inputs indicates the number M of scan-inputs. The behavior of Scan < num_scan_inputs = m, body = loop-body, scan_input_axes = [axis_1, ..., axis_m] > (init_1, ..., init_n, scan_1, ..., scan_m) is equivalent to the following pseudo-code: // scan_i.shape[axis_i] denotes the (max) sequence-length of scan_i // scan_i.shape[axis_i] is required to be equal to scan_j.shape[axis_j] for all i,j. sequence_length = scan_1.shape[axis_1]; // initialize state-variables st_1 = init_1; ... st_n = init_n; // initialize scan-output variables: [] denotes an empty tensor scan_out_1 = []; ...; scan_out_k = []; // identify number of iterations: // execute loop for (int t = 0; t < sequence_length; ++t) { // generate the scan-input elements: the notation T[t] indicates the sub-tensor // of rank one less than T obtained by indexing T at position t along axis k. si_1 = scan_1[t]; ... ; si_m = scan_m[t]; // execute loop-body st_1, ..., st_n, so_1, ..., so_k = loop-body(st_1, ..., st_n, si_1, ..., si_m) // accumulate the scan-output elements scan_out_1 = Concat(scan_out_1, so_1); ... ; scan_out_k = Concat(scan_out_k, so_k); } return st_1, ..., st_n, scan_out_1, ..., scan_out_k; *Sample usage: Encoding RNN using a Scan* The following example shows how a simple RNN over an input tensor %X, with weight tensor %Wi, recurrence weight tensor %Ri, bias tensors %Wbi and %Rbi, and initial hidden-state %H_0 can be encoded as a ScanLoop. Note that the loop-body is a nested graph, and it directly computes %Wi, %Ri, %Wbi, and %Rbi (typically constants or initializers in the body graph). If these values are computed in the outer graph, they need to be passed in as extra state_variables. graph rnn-encoding { %H_0 = ... %X = ... %Y_h, %Y = Scan[body = , num_scan_inputs=1](%H_0, %X) return %Y, %Y_h } graph rnn-cell-1 ( %H_tminus1[FLOAT, tensor] %X_t[FLOAT, tensor] ) { %Wi = ... %Ri = ... %Wbi = ... %Rbi = ... %t1 = X_t * (Wi^T) %t2 = H_tminus1*(Ri^T) %t3 = Add(%t1, %t2) %t4 = Add(%t3, %Wbi) %t5 = Add(%t4, %Rbi) %Ht = Tanh(%t5) %Accumulate = Identity(%Ht) return %Ht, %Accumulate } )DOC"; extern void ScanInferenceFunction(InferenceContext& ctx); ONNX_OPERATOR_SET_SCHEMA( Scan, 11, OpSchema() .SetDoc(scan_11_doc) .Input( 0, "initial_state_and_scan_inputs", "Initial values of the loop's N state variables followed by M scan_inputs", "V", OpSchema::Variadic, false) .Output( 0, "final_state_and_scan_outputs", "Final values of the loop's N state variables followed by K scan_outputs", "V", OpSchema::Variadic, false) .Attr( "body", "The graph run each iteration. It has N+M inputs: " "(loop state variables..., scan_input_elts...). It has N+K outputs: " "(loop state variables..., scan_output_elts...). Each " "scan_output is created by concatenating the value of the specified " "scan_output_elt value at the end of each iteration of the loop. It is an error" " if the dimensions of these values change across loop iterations.", AttributeProto::GRAPH, true) .Attr("num_scan_inputs", "An attribute specifying the number of scan_inputs M. ", AttributeProto::INT, true) .Attr( "scan_input_directions", "An optional list of M flags. The i-th element of the list specifies the direction " "to be scanned for the i-th scan_input tensor: 0 indicates forward direction and 1 " "indicates reverse direction. " "If omitted, all scan_input tensors will be scanned in the forward direction.", AttributeProto::INTS, false) .Attr( "scan_output_directions", "An optional list of K flags, one for each scan_output. The i-th element of the list " "specifies whether the i-th scan_output should be constructed by appending or " "prepending a new value in each iteration: 0 indicates appending and 1 " "indicates prepending. " "If omitted, all scan_output tensors will be produced by appending a value " "in each iteration.", AttributeProto::INTS, false) .Attr( "scan_input_axes", "An optional list of M flags. The i-th element of the list specifies the axis " "to be scanned (the sequence axis) for the i-th scan_input. If omitted, 0 will " "be used as the scan axis for every scan_input. Negative value for an axis means " "counting dimensions from the back. Accepted range is [-r, r-1] where r = rank(input).", AttributeProto::INTS, false) .Attr( "scan_output_axes", "An optional list of K flags. The i-th element of the list specifies the axis " "for the i-th scan_output. The scan outputs are accumulated along the specified " "axis. If omitted, 0 will be used as the scan axis for every scan_output. " "Negative value for an axis means counting dimensions from the back. Accepted " "range is [-r, r-1].", AttributeProto::INTS, false) .TypeConstraint("V", OpSchema::all_tensor_types(), "All Tensor types") .TypeAndShapeInferenceFunction(ScanInferenceFunction)); } // namespace ONNX_NAMESPACE