for (int32_t i = 0; i < dataDims.nbDims; ++i) { if (indicesDims.d[i] != -1 && dataDims.d[i] != -1) { ASSERT(indicesDims.d[i] <= dataDims.d[i] && "Indices dimensions must be less than data dimensions!", ErrorCode::kUNSUPPORTED_NODE); } if (updatesDims.d[i] != -1 && dataDims.d[i] != -1) { ASSERT(updatesDims.d[i] <= dataDims.d[i] && "Updates dimensions must be less than data dimensions!", ErrorCode::kUNSUPPORTED_NODE); } }
In this section, the assertion indicesDims.d[i] <= dataDims.d[i] should be changed into indices.max() <= dataDims.d[i]. the assertion updatesDims.d[i] <= dataDims.d[i] should be changed into updates.max() <= dataDims.d[i].
This is because that scatter cares about the index instead of the values, hence, dataDims can be irrelevant with indicesDims/updatesDims, besides that their ranks should be the same.
Environment
TensorRT Version: 8.6.1.6
ONNX-TensorRT Version / Branch: 1.16
GPU Type: RTX 4070
Nvidia Driver Version: 535.146.02
CUDA Version: 12.2
CUDNN Version: 11.8
Operating System + Version: Ubuntu 18.4
Python Version (if applicable): 3.10
TensorFlow + TF2ONNX Version (if applicable):
PyTorch Version (if applicable): 2.2.1
Baremetal or Container (if container which image + tag):
Relevant Files
No relevant files.
Steps To Reproduce
You could reproduce it by converting the following into an onnx and then into a trt engine:
y = torch.scatter_add(input=y_, dim=1, index=x, src=x_)
where y is of the shape of (1, 20), x is of the shape of (1, 4096) with the maximum value <= 19, x is of the shape of (1, 4096) and filled with 1's.
You will see that it can be converted into onnx but fails at trt engine.
Description
for (int32_t i = 0; i < dataDims.nbDims; ++i) { if (indicesDims.d[i] != -1 && dataDims.d[i] != -1) { ASSERT(indicesDims.d[i] <= dataDims.d[i] && "Indices dimensions must be less than data dimensions!", ErrorCode::kUNSUPPORTED_NODE); } if (updatesDims.d[i] != -1 && dataDims.d[i] != -1) { ASSERT(updatesDims.d[i] <= dataDims.d[i] && "Updates dimensions must be less than data dimensions!", ErrorCode::kUNSUPPORTED_NODE); } }In this section, the assertion
indicesDims.d[i] <= dataDims.d[i]should be changed intoindices.max() <= dataDims.d[i]. the assertionupdatesDims.d[i] <= dataDims.d[i]should be changed intoupdates.max() <= dataDims.d[i].This is because that scatter cares about the index instead of the values, hence, dataDims can be irrelevant with indicesDims/updatesDims, besides that their ranks should be the same.
Environment
TensorRT Version: 8.6.1.6 ONNX-TensorRT Version / Branch: 1.16 GPU Type: RTX 4070 Nvidia Driver Version: 535.146.02 CUDA Version: 12.2 CUDNN Version: 11.8 Operating System + Version: Ubuntu 18.4 Python Version (if applicable): 3.10 TensorFlow + TF2ONNX Version (if applicable): PyTorch Version (if applicable): 2.2.1 Baremetal or Container (if container which image + tag):
Relevant Files
No relevant files.
Steps To Reproduce
You could reproduce it by converting the following into an onnx and then into a trt engine:
y = torch.scatter_add(input=y_, dim=1, index=x, src=x_)where y is of the shape of (1, 20), x is of the shape of (1, 4096) with the maximum value <= 19, x is of the shape of (1, 4096) and filled with 1's.You will see that it can be converted into onnx but fails at trt engine.