Symbolic shape inference support-3: more ops for data propagation #3593
Conversation
Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
jcwchen
changed the title
Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
| @@ -1595,6 +1664,28 @@ ONNX_OPERATOR_SET_SCHEMA( | |||
| : // i - axis < q | |||
| data_shape.dim(i - q + 1); // i < out_rank < q + r - 1 | |||
| } | |||
| }) | |||
| .PartialDataPropagationFunction([](DataPropagationContext& ctx) { | |||
| if (!axisIsZero(ctx, true)) { | |||
There was a problem hiding this comment.
Actually, I wonder if we should restrict ourselves to the case where input(0) is 1-dimensional tensor? If it is multidimensional, then it is more complicated.
There was a problem hiding this comment.
We do have this restriction now for every data propagation, right? (only 0D or 1D tensor can be getInputData)
For other N-D tensors (which N > 1), their getInputData() will be nullptr and it will stop data propagation.
Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
onnx/defs/tensor/defs.cc
Outdated
| @@ -778,6 +803,36 @@ Example 2: | |||
| ] | |||
| )DOC"; | |||
|
|
|||
| inline int64_t clamp(int64_t val, int64_t low, int64_t high) { | |||
There was a problem hiding this comment.
this can be converted in 1 line:
return (val < low) ? low : (val > high) ? high : val;
also I suggest change low/high to min/max for better readability
another suggestion is you can simply use this 1 line instead of creating a inline function
Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
onnx/defs/math/defs.cc
Outdated
| } else if (input_dim_0.has_dim_param() || input_dim_1.has_dim_param()) { | ||
| tsp.mutable_dim()->Add()->set_dim_param("?"); | ||
| } else { | ||
| // Dim is not set by value or param |
There was a problem hiding this comment.
Hi: I suggest adding a dim in all cases, and setting its value only if both have values, like below:
auto* new_dim = tsp.mutable_dim()->Add();
if (input_dim_0.has_dim_value() && input_dim_1.has_dim_value()) {
new_dim->set_dim_value(
MathOpTwoIntegers(op_type, input_dim_0.dim_value(), input_dim_1.dim_value()));
}Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
| (inferredShape.dim(i).has_dim_value() && expectedShape.dim(i).has_dim_value()) || | ||
| (inferredShape.dim(i).has_dim_param() && expectedShape.dim(i).has_dim_param())) | ||
| (inferredShape.dim(i).has_dim_value() == expectedShape.dim(i).has_dim_value()) && | ||
| (inferredShape.dim(i).has_dim_param() == expectedShape.dim(i).has_dim_param())) |
There was a problem hiding this comment.
Update here to test with "?" shape (which means no value and no param because it is not computable) -- if the dim does not have value or parameter, it is still a valid dim.
Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
Description
Support more ops for data propagation:
Also add tests for them.
Motivation and Context
#3551 ONNX has supported data propagation. Implement more data propagation functions for more ops