Repository navigation
[TFTRT] Add Dynamic Shape Testing and fix Explicit Batch Mode for ConvertUnpack - #48049
copybara-service[bot] merged 1 commit into
Conversation
| if (! params->validation_only){ | ||
| } |
There was a problem hiding this comment.
Indeed debug leftovers. Removing
| if (! params->validation_only){ | ||
| } | ||
| } | ||
| if (params->validation_only) return Status::OK(); |
There was a problem hiding this comment.
I should remove this for these reasons:
The SqueezeTensor converter can handle validataion_only mode.
The SqueezeTensor converter may return error, in that case, we should not convert and returning Status::OK() here premature and wrong.
There was a problem hiding this comment.
SqueezeTensor seems to handle validation_only mode, but in practice we do not use it for validation in any other converter. I believe we cannot use it here either:
- The first arg (
params->outputs->at(i)) is not created byConvertStridedSliceHelperduring validation. - I am not sure if
params->converterexists invalidation_mode.
| #endif | ||
| // We need to skip error testing for Dynamic Shape mode, as it is impossible | ||
| // to convert Unpack in Dynamic Shape Mode. | ||
| // `Unimplemented: Strided slice op not implemented for dynamic shape input` |
There was a problem hiding this comment.
Sorry, I don't understand this line. What is the purpose of this line?
There was a problem hiding this comment.
Error expected from the converter if you try dynamic shape inputs.
Removing it
google-ml-butler
Bot
added
kokoro:force-run
bixia1
left a comment
bixia1
left a comment
There was a problem hiding this comment.
The added test didn't pass:
third_party/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc:129
No Diff
Inline Diff
Line Diff
Token Diff
Char Diff
Ignore White Space
Value of: status.error_message()
Expected: has substring "Dimension 3 of size 3 is not evenly divisble by 2, at my_split"
Actual: "Dimension 3 of size 3 is not evenly divisible by 2, at my_split"
Invalid argument: Dimension 3 of size 3 is not evenly divisible by 2, at my_split
Stack trace:
0x55fb2b36f5ad: tensorflow::tensorrt::convert::ExpectStatus(tensorflow::Status, tensorflow::error::Code, char const*) @ ../sysdeps/x86_64/start.S:121
0x55fb2b385435: tensorflow::tensorrt::convert::OpConverterTest::RunValidationAndConversion(tensorflow::NodeDef const&, tensorflow::error::Code, char const*, bool) @ ??:??
0x55fb2b3da77e: tensorflow::tensorrt::convert::OpConverterTest_ConvertSplit_Test::TestBody() @ ??:??
0x55fb2c886521: testing::Test::Run() @ ??:??
0x55fb2c8877ee: testing::TestInfo::Run() @ ??:??
... Google Test internal frames ...
third_party/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc:129
No Diff
Inline Diff
Line Diff
Token Diff
Char Diff
Ignore White Space
Value of: status.error_message()
Expected: has substring "Dimension 3 of size 3 is not evenly divisble by 4, at my_split"
Actual: "Dimension 3 of size 3 is not evenly divisible by 4, at my_split"
Invalid argument: Dimension 3 of size 3 is not evenly divisible by 4, at my_split
Stack trace:
0x55fb2b36f5ad: tensorflow::tensorrt::convert::ExpectStatus(tensorflow::Status, tensorflow::error::Code, char const*) @ ../sysdeps/x86_64/start.S:121
0x55fb2b385435: tensorflow::tensorrt::convert::OpConverterTest::RunValidationAndConversion(tensorflow::NodeDef const&, tensorflow::error::Code, char const*, bool) @ ??:??
0x55fb2b3da8f0: tensorflow::tensorrt::convert::OpConverterTest_ConvertSplit_Test::TestBody() @ ??:??
0x55fb2c886521: testing::Test::Run() @ ??:??
0x55fb2c8877ee: testing::TestInfo::Run() @ ??:??
... Google Test internal frames ...
Thanks Bixia |
Fixed ConvertSplitHelper to make ConvertUnpack work for explicit batch mode.
Add dynamic shape testing for ConvertUnpack.
@bixia1 @tfeher for review
Feature Tracker: #45481