Dropout (Channel) on HIP and HOST #634
Conversation
r-abishek
commented
Nov 4, 2025
r-abishek
commented
- Adds Channel Dropout augmentation on HIP and HOST
- Adds support for U8/F32/F16/I8 bit depths and NCHW/NHWC variants with toggle support
- Adds relevant unit / qa / performance tests
…all the bitdepths
indentation modified
… as constant for better understanding, added required comments
…d the HOST backend double free fault
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## develop #634 +/- ##
===========================================
+ Coverage 88.16% 88.36% +0.20%
===========================================
Files 195 197 +2
Lines 82723 83047 +324
===========================================
+ Hits 72932 73383 +451
+ Misses 9791 9664 -127
🚀 New features to boost your workflow:
|
| rpp::Handle &handle) | ||
| { | ||
| if (roiType == RpptRoiType::LTRB) | ||
| hip_exec_roi_converison_ltrb_to_xywh(roiTensorPtrSrc, handle); |
There was a problem hiding this comment.
Please fix the typo in the function name _converison_ -> _conversion_
There was a problem hiding this comment.
This issue is prevalent across files. Will issue a common fix PR for the same. Thanks
There was a problem hiding this comment.
Why do we need a separate PR for fixing the typo?
There was a problem hiding this comment.
@rrawther The same function call is used in many instances - all are fixed in #645 now.
Thanks @AryanSalmanpour
| dstPtrImage = dstPtr + batchCount * dstDescPtr->strides.nStride; | ||
|
|
||
| uint8_t *maskPtr = scratchBuffer + batchCount * srcDescPtr->c; | ||
| int seed = randomSeed ? std::random_device{}() : DROPOUT_FIXED_SEED; // Use a true random seed if requested, otherwise use the fixed seed for deterministic QA |
There was a problem hiding this comment.
All the randomization has to happen outside the kernel from the calling function
There was a problem hiding this comment.
dropoutProbability and randomSeed have now been removed from arguments.
RPP won't be responsible for the randomization part.
| #pragma omp parallel for num_threads(numThreads) | ||
| for (int batchCount = 0; batchCount < dstDescPtr->n; batchCount++) | ||
| { | ||
| std::mt19937 gen(seed + batchCount); |
There was a problem hiding this comment.
As mentioned above this need to happen outside the kernel from the caller
rrawther
left a comment
rrawther
left a comment
There was a problem hiding this comment.
waiting for the random dropout change before merge
Channel Dropout Resolved review comments
|
@rrawther , the random dropout changes were made as discussed. Thanks |
9 participants
