Process FP16 output and output to Console and CSV file #176
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
added 11 commits
February 19, 2019 18:25
…r tensor comparisons
fdwr
reviewed
Feb 26, 2019
| std::getline(actualFileStream, actualValue, '\n'); | ||
| float actualValueNum = 0; | ||
| float expectedValueNum = 0; | ||
| try |
fdwr
reviewed
Feb 26, 2019
| } | ||
| //If the string values aren't floats then ignore | ||
| } | ||
| if (std::abs(actualValueNum - expectedValueNum) > epsilon * std::abs(expectedValueNum) + epsilon) //Check if the values are too different. |
Contributor
There was a problem hiding this comment.
relative tolerance is the typical term.
fdwr
reviewed
Feb 26, 2019
Tools/WinMLRunner/src/OutputHelper.h
Outdated
|
|
||
| template<typename T> | ||
| void WriteTensorResultToCSV(T* m_Res, uint32_t iterationNum, const CommandLineArgs &args, uint32_t &capacity, std::string &featureName) | ||
| void ProcessTensorResult(com_ptr<ITensorNative> itn, |
Contributor
There was a problem hiding this comment.
Try to detemplatize so there isn't code repetition. You can just use a switch statement in the innermost loop.
Contributor
Author
There was a problem hiding this comment.
Removed out most of the code that's unnecessary for being inside the template. Case switch statement for tensorkind will have to exist somewhere. I chose to have it in the helper method that calls ProcessTensorResult and leave ProcessTensorResult as template for using typename T for looping through the buffer.
fdwr
approved these changes
Feb 26, 2019
Contributor
fdwr
left a comment
fdwr
left a comment
There was a problem hiding this comment.
Looks fine. Add some desired changes, but nothing blocking. Thanks for unblocking Intel and Indy for float16.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This change primarily addresses these main points:
Before this change, we access the output tensor twice and it was inefficient and had code bloat. It accessed it once to dump the output tensor to CSV and another time to print the max value to console. With this change, there is only one method to access the output tensor once. PrintOrSaveEvaluationResults is a helper method which controls the cases for the output tensorkind and it calls ProcessTensorResult which is a templated method that handles the access of the output tensor depending on the data type.
This change fixes the CompareTensor helper method inside the tests so that the comparing or the tensor values is more fair. Epsilon is used here to determine the threshold where the expected tensor value and actual tensor value are too different.
This change adds squeezenet fp16 model. It was converted with WinMLTools 1.3.0.
Output Tensor for FP16 tests and test data were added as well.