[ Model ] Enable Mixed Precision Training #2628
Merged
Conversation
This file contains 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
We will add Var32 Tensor if the Variable Weight is not Full precision (FP32). This eables the Weight Update with full precision and only Apply Gradient Process ueses this Tensor. Therefore, the lifespan of this tensor should be "ApplyGradient". . Modify TensorPool to generate Weigth considering Mixed Precsion. **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This pr create the variable fp32 tensor when we create the Weight and Optimizer Weight. . update the manager to create Weight with var32 tensor which requested to weight pool. . update the weight requests with Weight Spec and var, grad and var32 tensors which created already. . add clone Tensor with specific type in tensor.h Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR enables the FP16 support for the layers below: . input layer . mse loss layer Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR includes the mixed precision test case. . Input - FC - MSE : "batch_size=2", "model_tensor_type=FP16-FP16", "loss_scale=128" **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This commit modify apply gradient in optimizer. We do not need to save optimizer variables in weight type. Only Optimizer needs the optimizer variables and we should update the weight with full precision to maintain the accuracy. Therefore, remove the var32 tensors for optimizer variables. Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR add is_NaN function to check if the tensor has NaN value. This is for the check NaN during mixed precision training. **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR add loss scale parameter in runcontext and use it to update mse loss. . Add Loss Scale Parameter in RunLayerContext Constructor . Add applyLossScale func to update return derivitive in Loss Layer . Change MSE Loss Layer to apply the loss scale to return derivitive **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR enables the Mixed Precision Training. For now only FP16-FP32 is considered. Additional Test cases will be added. . add getSortedLayerIdx to set the graph order for fowarding. . change clip_weights to lazy_apply_weights to use both cases. . add fowarding_op to run forwarding from that layer which has a gradient with nan. . add while loop for re-run backwarding after reset the loss scale. . add setLossScale in RunLayerContext . add check the gradient if mixed precsion enable. **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR add inifinity value check in Tensor data. . rename the hasNaN to isValid . add infinity check in isValid Function and now it check NaN and Inf . modify to check the blas_avx and blas_neon . modify graph and model check is_valid rather than has_nan . add unittest of isValid Function **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR chage the loss computation using full precsion rather than half precsion to maintain accuracy. **Changes proposed in this PR:** - Added TOC generator for README.md Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR enables the Mixed Precsion Unittest with Torch Model. Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR add torch mixed precsion golden data generation and input and output for test. . some fixes to test. Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR includes more unittest and fixes for mixed precsion. . Model Unittest . 2 fc layer which generate NaN or Inf Gradient from Troch. . MSE Loss and Check whole procedure of the mixed precsion training. . Even if the FC model only have one weight, but it is good enough to validate the mixed precsion. . Torch model also work similar way of NNTrainer. . Some fixes about the exeuction order of apply gradient when the mixed precision is on. . Update SGD to support Mixed Precision training **Changes proposed in this PR:** - Added TOC generator for README.md Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR update the conv2D Layer to support Mixed Precision (FP16). It is based on the PR #2579 Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This commit enables mixed precision support for LSTM Layer. Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR add Execution Mode parameter when we compile. The default is ml::train::ExeuctionMode::TRAIN. Currently we do not support compiler optimization for inference mode such as batch normalization fusing, etc. But we will add more optimization depending on the exeuction mode. Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR includes Mixed Precision support for batch normalization layer. When the training, BN layer should run full precsion with FP16 Weight data. Therefore, Reading the FP16 data read and data coversion of the current Weight and Activation are required. For the Inference, we do need compiler optimization like bn fusing. So it also includes execution mode parameters for compile. Because of compilcate data conversion of bn layer, test case generation also needs to update, so that taking the fp16 input,output tensors and weights and converting FP32 weight for computation. For veification, we do need convert FP32 to FP16. **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
enable mixed precision on reshape layer - reshape layer only change dim, so change dimensions and check datatype **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: Donghak PARK <[email protected]>
Enable Mixed precision on Pooling 2D Layer - I modified it to properly cast for the case of FP16 so that the mixed precision function can be activated on the existing pooling 2d layer. **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: Donghak PARK <[email protected]>
In this PR, when we compute the l2norm of gradient tensor, it converts to full precsion and computes the l2norm for gradient clipping. Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR add the mu and var backup tensor ( mu_b, var_b ) to restore the previous moving mean and moving variance for mixed precsion training. Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
In order to resotore previous iteration data, this pr disable randomnization of mask if it need restore previous data. Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
This PR enable the check if it need restore previous data. By doing this, we can remove the NaN or Inf data in Tensor for the mixed precsion training. **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
jijoongmoon
requested review from
myungjoo,
again4you,
jaeyun-jung,
leemgs,
wooksong,
helloahn and
kparichay
as code owners
jijoongmoon
requested review from
anyj0527,
zhoonit,
lhs8928,
songgot,
jihochu,
DonghakPark,
SeoHyungjun,
baek2sm,
skykongkong8,
djeong20,
EunjuYang and
a team
as code owners
|
📝 TAOS-CI Version: 1.5.20200925. Thank you for submitting PR #2628. Please a submit 1commit/1PR (one commit per one PR) policy to get comments quickly from reviewers. Your PR must pass all verificiation processes of cibot before starting a review process from reviewers. If you are new member to join this project, please read manuals in documentation folder and wiki page. In order to monitor a progress status of your PR in more detail, visit http://ci.nnstreamer.ai/. |
|
|
cibot:
jijoongmoon
force-pushed
the
mixed_precision_training
branch
from
88d84e4 to
8e06368
Compare
|
|
jijoongmoon
force-pushed
the
mixed_precision_training
branch
from
8e06368 to
54bd73d
Compare
|
|
We do need to remove the Nan or Inf value in Tensor by call setZero(). However, if we using sscal, then Nan or Inf values are remain still. This PR change the sscal to memset. Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
jijoongmoon
force-pushed
the
mixed_precision_training
branch
from
54bd73d to
afff553
Compare
taos-ci
approved these changes
Jun 10, 2024
There was a problem hiding this comment.
@jijoongmoon, 💯 All CI checkers are successfully verified. Thanks.
myungjoo
reviewed
Jun 10, 2024
| @@ -64,9 +64,19 @@ warning_c_flags = [ | |||
| '-Wno-error=varargs' | |||
| ] | |||
|
|
|||
| arch = host_machine.cpu_family() | |||
|
|
|||
| if get_option('enable-avx') | |||
There was a problem hiding this comment.
Please don't enable-avx if arch is not x64. We need to emit erros or force-disable avx if it's not x64 in meson script.
You may use tri-state (meson feature) instead of bool for the convinience of avx option.
Currenty, you are enabling avx in Ubuntu-arm/risc-v.
myungjoo
reviewed
Jun 10, 2024
| @@ -40,7 +40,7 @@ option('enable-fp16', type: 'boolean', value: false) | |||
| option('enable-cublas', type: 'boolean', value: false) | |||
| option('enable-openmp', type: 'boolean', value: true) | |||
| option('enable-neon', type: 'boolean', value: false) | |||
| option('enable-avx', type: 'boolean', value: false) | |||
| option('enable-avx', type: 'boolean', value: true) | |||
There was a problem hiding this comment.
For those who depend on arch, I recommend "feature" instead of "boolean".
If this is to be kept "boolean" with "true", meson should be able to check the arch and turn avx off if it's not x64.
myungjoo
approved these changes
Jun 10, 2024
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.
In this PR
This PR modifies codes related to Mixed Precision Training.
Commits to be reviewed in this PR
[ Model ] Fix the gradient clipping for the FP16 or Low bit Gradient
In this PR, when we compute the l2norm of gradient tensor, it converts
to full precsion and computes the l2norm for gradient clipping.
Resolves:
Self evaluation:
Signed-off-by: jijoong.moon [email protected]
[ Layer ] Add mu and var backup up tensor
This PR add the mu and var backup tensor ( mu_b, var_b ) to restore
the previous moving mean and moving variance for mixed precsion
training.
Resolves:
Self evaluation:
Signed-off-by: jijoong.moon [email protected]
[ Layer ] prevent randomize when it restore the data
In order to resotore previous iteration data, this pr disable randomnization of mask if it need restore previous data.
Resolves:
Self evaluation:
Signed-off-by: jijoong.moon [email protected]
[ Context ] add check if it needs restore previous data
This PR enable the check if it need restore previous data. By doing this, we can remove the NaN or Inf data in Tensor for the mixed precsion training.
Self evaluation:
Signed-off-by: jijoong.moon [email protected]
[ Tensor ] remove sscal to set zero.
We do need to remove the Nan or Inf value in Tensor by call setZero(). However, if we using sscal, then Nan or Inf values are remain still. This PR change the sscal to memset.
Resolves:
Self evaluation:
Signed-off-by: jijoong.moon [email protected]