Support non-spatial mode in BatchNormalization #2092
Conversation
| @@ -532,7 +532,6 @@ int real_main(int argc, char* argv[], Ort::Env& env) { | |||
| #endif | |||
|
|
|||
| #ifdef USE_CUDA | |||
| broken_tests.insert({"mxnet_arcface", "result mismatch"}); | |||
There was a problem hiding this comment.
Added to the exclusion list once above
| @@ -319,7 +319,7 @@ private void TestMultiThreads() | |||
| private static Dictionary<string, string> GetSkippedModels() | |||
| { | |||
| var skipModels = new Dictionary<string, string>() { | |||
| { "mxnet_arcface", "Model not supported by CPU execution provider" } , | |||
| { "mxnet_arcface", "Model is an invalid ONNX model", | |||
There was a problem hiding this comment.
See reply here on why it is invalid:
onnx/models#156
| //constexpr int kMinCudaNumDims = 4; | ||
| //constexpr int kMaxCudaNumDims = 5; | ||
|
|
||
| if (X->Shape().GetDims().empty()) { |
There was a problem hiding this comment.
Removed this check in favor of the more stricter check in line 20.
| const Tensor* var, | ||
| bool is_spatial = true) { | ||
| const auto& x_dims = X->Shape().GetDims(); | ||
| if (x_dims.size() < 2) { |
There was a problem hiding this comment.
Even though, the spec doesn't explicitly state it, it makes sense if the minimum rank of 'X' is atleast 3 - N,C, and atleast one data dimension (D1). However, when we impose that, it seems to break some successfully running models that have input rank 2 - [N, C]. The implicit shape is [N, C, 1] in these cases and the results seem ok. Hence, modifying the minimum rank criteria to just 2. This is necessary to add as in the next step we parse the number of channels from the index position 1.
Description: Older opset spec of BatchNormalization supports non-spatial mode of the operator supported by some frameworks. This change supports that mode of operation.
Motivation and Context
Keep ORT backwards compatible