Skip to content

Commit a3f1cd1

Browse files
WafaaTleslie-fang-inteldmsuehirashahbaXiaobingSuper
authored
Sync master with r2.9 (#858)
* revert bf16 changes (#488) * Add partials and spec yml for the end2end DLSA pipeline (#460) * Add partials and specs for the end2end DLSA pipeline * Add missing end line * Update name to include ipex * update specs to have use the public image as a base on one and SPR for the other * Dockerfile updates for the updated DLSA repo * Update pip install list * Rename to public * Removing partials that aren't used anymore * Fixes for 'kmp-blocktime' env var (#493) * Fixes for 'kmp-blocktime' env var Signed-off-by: Abolfazl Shahbazi <[email protected]> * update per review feedback Signed-off-by: Abolfazl Shahbazi <[email protected]> * Add 'kmp-blocktime' for mlperf-gnmt (#494) * Add 'kmp-blocktime' for mlperf-gnmt Signed-off-by: Abolfazl Shahbazi <[email protected]> * Remove duplicate parameter definition Signed-off-by: Abolfazl Shahbazi <[email protected]> * add sample_input for resnet50 training (#495) * remove the case when fragment_size not equal args.batch_size (#500) * Changed the transformer_mlperf fp32 model so that we can fuse the ops… (#389) * Changed the transformer_mlperf fp32 model so that we can fuse the ops in the model, and also minor changes for python3 * Changed the transformer_mlperf int8 model so that we can fuse the ops in the model, and also minor changes for python3 * SPR updates for WW12, 2022 (#492) * SPR updates for WW12, 2022 Signed-off-by: Abolfazl Shahbazi <[email protected]> * Update for PyTorch SPR WW2022-12 Signed-off-by: Abolfazl Shahbazi <[email protected]> * Update pytorch base for SPR too Signed-off-by: Abolfazl Shahbazi <[email protected]> * Stick with specific 'keras-nightly' version Signed-off-by: Abolfazl Shahbazi <[email protected]> * Updates per code review Signed-off-by: Abolfazl Shahbazi <[email protected]> * update maskrcnn training_multinode.sh (#502) * Fixed a bug in the transformer_mlperf model threads setting (#482) * Fixed a bug in the transformer_mlperf model threads setting * Fix failing tests Signed-off-by: Abolfazl Shahbazi <[email protected]> Co-authored-by: Abolfazl Shahbazi <[email protected]> * Added the default threads setting for transformer_mlperf inference in… (#504) * Added the default threads setting for transformer_mlperf inference in case there is no command line input * Fix unit tests Signed-off-by: Abolfazl Shahbazi <[email protected]> Co-authored-by: Abolfazl Shahbazi <[email protected]> * PyTorch Image Classification TL notebook (#490) * Adds new TL notebook with documentation * Added newline * Added to main TL README * Small fixes * Updated for review feedback * Added more models and a download limit arg * Removed py3.9 requirement and changed default model * Adds Kitti torchvision dataset to TL notebook (#512) * Adds Kitti torchvision dataset to TL notebook * Fixed citations formatting * update maskrcnn model (#515) * minor update. (#465) * Create unit-test github action workflow (#518) * Create unit-test github action workflow * Update to Horovod commit 11c1389 to fix TF v2.9 + Horovod install failure (#519) Signed-off-by: Abolfazl Shahbazi <[email protected]> * update distilbert model to 4.18 transformers and enable int8 path (#521) * rnnt: use launcher to set output file path and name (#524) * Update BareMetalSetup.md (#526) Always use the latest torchvision * Reduce memory usage for dlrm acc test (#527) * updatedistilbert with text_classification (#529) * add patch for distilbert (#530) * Update the model-builder dockerfile to use ubuntu 20.04 (#532) * Add script for coco training dataset processing (#525) * and update tensorflow ssd-resnet34 training dataset instructions * update patch (#533) Co-authored-by: Wang, Chuanqi <[email protected]> * [RNN-T training] Enable FP32 gemm using oneDNN (#531) * Update the Readme guide for distilbert (#534) * Update the Readme guide for distilbert * Fix accuracy grep bug, and grep accuracy for distilbert Co-authored-by: Weizhuo Zhang <[email protected]> * Update end2end public dockerfile to look for IPEX in the conda directory (#535) * Notebook to script conversion example (#516) * Add notebook script conversion example * Fixed doc * Replaces custom preprocessor with built-in one * Changed tag to remove_for_custom_dataset * Add URL check prior to calling urlretrieve (#538) * Add URL check prior to calling urlretrieve Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix a typo Signed-off-by: Abolfazl Shahbazi <[email protected]> * disable for ssd since fused cat cat kernel is slow (#537) * fix bug when adding steps in rnnt inference (#528) * Fix and updates for TensorFlow WW18-2022 SPR (#542) * Fix and updates for TensorFlow WW18-2022 SPR Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix TensorFlow SPR nightly versions Signed-off-by: Abolfazl Shahbazi <[email protected]> * Update pre-trained models download URLs Signed-off-by: Abolfazl Shahbazi <[email protected]> * Intall Python 3.8 development tools Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix OpenMPI install and setup Signed-off-by: Abolfazl Shahbazi <[email protected]> * Update to Horovod commit 11c1389 to fix TF v2.9 + Horovod install failure (#519) Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix Horovod Installaion for SPR and CentOS Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix Python3.8 version for CentOS Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix a typo in TensorFlow 3d-unet partial Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix a broken partial Signed-off-by: Abolfazl Shahbazi <[email protected]> * Add TCMalloc to TF base container for SPR and remove OpenSSL Signed-off-by: Abolfazl Shahbazi <[email protected]> * Remove some repositories Signed-off-by: Abolfazl Shahbazi <[email protected]> * Add 'matplotlib' for '3d-unet' Signed-off-by: Abolfazl Shahbazi <[email protected]> * switch to build OpenMPI due to issue in Market Place provided version Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix PYTORCH_WHEEL and IPEX_WHEEL arg values Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix and updates for PyTorch WW14-2022 SPR (#543) * Fix and updates for PyTorch WW14-2022 SPR Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix and updates for TensorFlow WW18-2022 SPR Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix TensorFlow SPR nightly versions Signed-off-by: Abolfazl Shahbazi <[email protected]> * Update pre-trained models download URLs Signed-off-by: Abolfazl Shahbazi <[email protected]> * Intall Python 3.8 development tools Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix OpenMPI install and setup Signed-off-by: Abolfazl Shahbazi <[email protected]> * Update to Horovod commit 11c1389 to fix TF v2.9 + Horovod install failure (#519) Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix Horovod Installaion for SPR and CentOS Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix Python3.8 version for CentOS Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix a typo in TensorFlow 3d-unet partial Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix a broken partial Signed-off-by: Abolfazl Shahbazi <[email protected]> * Add TCMalloc to TF base container for SPR and remove OpenSSL Signed-off-by: Abolfazl Shahbazi <[email protected]> * Updates required to the base image Signed-off-by: Abolfazl Shahbazi <[email protected]> * Remove some repositories Signed-off-by: Abolfazl Shahbazi <[email protected]> * Add 'matplotlib' for '3d-unet' Signed-off-by: Abolfazl Shahbazi <[email protected]> * switch to build OpenMPI due to issue in Market Place provided version Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix PYTORCH_WHEEL and IPEX_WHEEL arg values Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix PYT resnet50 quickstart scripts for both Linux and Windows (#547) * fix quickstart scripts, detect platform type, update to run with pytorch only * Fix SPR PyTorch MaskRCNN inference documentation for CHECKPOINT_DIR (#548) * Enable bert large multi stream inference (#554) * test bert multi stream module * enable input split and output concat for accuracy run * change the default num_streams batchsize cores to 56 * change ssd multi stream throughput to 1 core 1 batch * change the default parameter for rn50 ssd multi stream module * modify enable_ipex_for_squad.diff to align new multistream hint implementation * enable warmup and multi socket support * change default parameter for rn50 ssd multi stream inference * Add train-no-eval for rn50 pytorch (#555) * PyTorch SPR BERT large training updates (h5py and dataset instructions) and update LD_PRELOAD for SPR entrypoints (#550) * Add h5py install to bert training dockerfile * documentation updates * update docs, and add input_preprocessing to the wrapper package * Update LD_PRELOAD trailing : * Fix syntax * removing unnecessary change * Update DLRM entrypoint * Update docs to note that phase2 has bert_config.json in the CHECKPOINT_DIR * Fix syntax * increase shm-size to 10g * [RNN-T training] Update scripts -- run on 1S (#561) * Update maskrcnn training script to run on 1s (#562) * use single node to do ssd-rn34 training (#563) * Update training.sh (#564) * Update training.sh (#565) Use tcmalloc instead of jemalloc * use single node to do resnet50 training (#568) * add numactl -C and remove jit warm in main thread (#569) * Update unit-test.yml (#546) * re-enable int8 for api change (#579) * saperate fully convergency test from training test (#581) Co-authored-by: jianan-gu <[email protected]> * ssd enable new int8 (#580) * v1 * enable new int8 method * Revert "ssd enable new int8 (#580)" (#584) This reverts commit 9eb3211. * Revert "re-enable int8 for api change (#579)" (#583) This reverts commit 0bded92. * Update training script using 1s (#560) * Enable checkpoint during training for bert-large (#573) * minor fix * Add readme for enabling checkpoint * update phase1 to enable checkpoint by default * Update README.md * Enable ssd bf32 inference training (#589) * enable ssd bf32 inference * enable ssd bf32 train * enable RNN-T bf32 inference (#591) * Enable bf32 for bert and distilbert for inference (#593) * enable bf32 distilbert * enable bert bf32 * Enable RNN-T bf32 training (#594) * enable maskrcnn bf32 inference and training (#595) * enable resnet50 and resnext101 bf16 path (#596) * enable bert bf32 train (#600) * update resnet int8 path using new int8 api (#603) * re-enable int8 for api change (#604) Co-authored-by: jianan-gu <[email protected]> * Leslie/ssd enable new int8 (#605) * v1 * enable new int8 method * update json file * add rn50 int8 weight sharing Co-authored-by: Jiang, Xiaofei <[email protected]> * update ssd training bs to the multily of core numbers (#606) * enable bf32 for dlrm (#607) Co-authored-by: jianan-gu <[email protected]> * Update IPEX new int8 API enabling for distilbert/bert-large (#608) * enable distilbert * enable bert * fix max-ind-range and add memory info (#609) Co-authored-by: jianan-gu <[email protected]> * Remove debug code (#610) * update training steps (#611) * fix bandit scan fails (#612) * PYT Image recognition models support on Windows (#549) * fix all image recognition scripts to run on windows and linux with PYT, and only linux with IPEX * [RNN-T training] fix bandit scan fails (#614) * RNN-T inference: fix IMZ Bandit scan fails (#615) * Update unit-test.yml (#570) * MaskRCNN: fix IMZ Bandit scan fails (#623) * Fix for horovod-related failures in TF nightly runs (#613) * cpp17 horovod failure fix * minor debugging changes * minor fixes - directory name * cleanup * addressing reviewer comments * Minor fix for Horovod install and adding 'tf_slim' for SSD ResNet34 (#624) * Minor fix for Horovod install and adding 'tf_slim' for SSD ResNet34 Signed-off-by: Abolfazl Shahbazi <[email protected]> * Set 'HOROVOD_WITH_MPI=1' explicitly Signed-off-by: Abolfazl Shahbazi <[email protected]> * update GCC version to GCC 9 Signed-off-by: Abolfazl Shahbazi <[email protected]> * Add 'horovodrun --check-build' for sanity check Signed-off-by: Abolfazl Shahbazi <[email protected]> * removo force install inside Docker Signed-off-by: Abolfazl Shahbazi <[email protected]> * [RNN-T training] Fix ddp sample number issue (#625) * update BF32 usage (#627) * resnet50 training: add warm up before collecting time (#628) * image to bf16 (#629) * Update end2end DLSA dockerfile due to SPR wheel path update and removing int8 patch (#631) * Update mlpc path for SPR wheels * remove patch * Update Horovod commit id for BareMetal, Docker will be updated next (#630) Signed-off-by: Abolfazl Shahbazi <[email protected]> * fix dlrm convergence and change training performance BS to 32K (#633) Co-authored-by: jianan-gu <[email protected]> * [RNN-T training] Merge sh files to one (#635) * update torch-ccl into 1.12 (#636) * Liangan1/update torch ccl version (#637) * Update torch_ccl version * resnet50_distributed_training: don't set MASTER_ADDR by user (#638) * Update torch_ccl in script (#639) * Enable offline download distilbert (#632) * enable offline download distilbert * add convert * Update README.md * add accuracy.py * add file * refine download * refine path * refine path * add license * Update dlrm_s_pytorch.py (#643) * Update README.md (#649) * init pytorch T5 language model (#648) * init pytorch T5 language model * update README.md * update doc * update fpn models (#650) * pytorch resnet50: directly call ipex.quantization (#653) * fix int8 accuracy (#655) Co-authored-by: Zhang, Weizhuo <[email protected]> * Made fixes to the broken links (#652) * Made fixes to the broken links * Changed the ResNet50v1_5 version back to v2_7_0 * Modified the setup AI kit instructions Co-authored-by: msalopan <[email protected]> * Update Security Center URL (#657) Signed-off-by: Abolfazl Shahbazi <[email protected]> * Weizhuoz/fix for pt 1.12 (#656) * fix vgg11_bn accuracy syntax error * remove exact_match from roberta-base * modify maskrcnn BS to 2*num_cores * Update dlrm_s_pytorch.py (#660) * Update dlrm_s_pytorch.py Reduce int8 memory usage. * Update dlrm_s_pytorch.py * Update dlrm_s_pytorch.py * Update dlrm_s_pytorch.py * Update dlrm_s_pytorch.py * Add BF32 DDP for bert-large (#663) * Update run_ddp_bert_pretrain_phase1.sh * Update run_ddp_bert_pretrain_phase2.sh * Update README.md * move OMP_NUM_THREADS=1 into dlrm_s_pytorch.py (#664) minor changes * remove rn50 ao (#665) * Re-organize models list to be grouped by framework (#654) * re-organize models list to be grouped by framework * update tensorflow ssd-resnet34 training dataset * add T5 in benchmark/README.md * mannuel set torch num threads only for int8 (#666) * Update inference_performance.sh (#669) * improve ssdrn34 perf. (#671) * improve ssdrn34 perf. * minor update. * Fix linting Signed-off-by: Abolfazl Shahbazi <[email protected]> * Fix unit tests too Signed-off-by: Abolfazl Shahbazi <[email protected]> Co-authored-by: Abolfazl Shahbazi <[email protected]> * Use IPEX Pytorch whls instead of building IPEX from source (#674) Co-authored-by: Clayne Robison <[email protected]> * Lpot2inc (#446) Co-authored-by: ltsai1 <[email protected]> * Sriniva2/ssd rn34 (#682) * improve ssdrn34 perf. * minor update. * enabling synthetic data. * Update base_benchmark_util.py * Fix linting error Signed-off-by: Abolfazl Shahbazi <[email protected]> Co-authored-by: Abolfazl Shahbazi <[email protected]> * Add doc updates for '--synthetic-data' option (#683) Signed-off-by: Abolfazl Shahbazi <[email protected]> * Change checkpoint setting for Bert train phase 1 (#602) * Change checkpoint setting for Bert train phase 1 * fix model and config saving * fix error when runing gpu path (#686) * fix load pretrained model error when using torch_ccl (#688) * update py version in base spec (#678) (#690) * TF addons upgrade to 0.17.1 (#689) (#691) * updated tf adons version * remove comment * Update Dockerfiles prior to IMZ 2.8 release (#693) Signed-off-by: Abolfazl Shahbazi <[email protected]> * Update Documents prior to IMZ 2.8 release (#694) Signed-off-by: Abolfazl Shahbazi <[email protected]> * Update README.md (#697) * change numpy version requirement (#703) * Remove MiniGo training from IMZ (#644) * remove MiniGo training scripts and unit test * [RNN-T] [Inference] optimize the batch decoder (#711) * reduce fill_ OP in rnnt embedding kernel * optimize add between int and log to reduce dtype conversion * rnnt: support dump tracing file and print profile table (#712) * add support for open SUSE leap operating system (#708) * rnnt inference: pre convert data to bf16 (#713) * remove squeeze/slice/transpose (#714) * update resnet50 training code (#710) * update resnet50 training code * not using ipex optimize for resnet50 training * use ipex.optimize() on the whole model (#718) * resnet50 bf32: calling ipex.optimize to enable bf32 path (#719) * Added batch size as an env variable to the quickstart scripts (#676) Co-authored-by: Clayne Robison <[email protected]> * Added batchsize as an env variable to quickstart scripts (#680) * updated readme: nit fix (#723) Co-authored-by: Rahul Nair <[email protected]> * compute throughput by test_mini_batch_size (#740) * pytorch resnet50: fix bf32 training path error (#739) * Fix a subtle 'E275' style issue that causes unknown behavior (#742) Signed-off-by: Abolfazl Shahbazi <[email protected]> Signed-off-by: Abolfazl Shahbazi <[email protected]> * rearrange the paragraphs and fix Markdown headers (#744) * Align Transformers version for BERT models (#738) * align transformer version(4.18) for bert models * change scripts to legacy * redo calibration * patch fix * Update README.md (#746) * Add support for stock PYT- object detection models (#732) * stock PYT and windows support for object detection models * Weizhuoz/reduce model zoo steps (#762) * reduce steps for bert-base, roberta, fpn models * modify max_iter for fpn models * reduce all img classification models steps * update new config for bert models (#763) * Addin Scipy for TensorFlow serving SSD-MobileNet model (#764) Signed-off-by: Abolfazl Shahbazi <[email protected]> Signed-off-by: Abolfazl Shahbazi <[email protected]> * Update TF ResNet50v1.5 inference for SPR (baremetal) (#749) * Added matplotlib dependency to image_segmentation requirements (#768) * Update readmes for the path to output directory (#769) * update wide & deep readme for the path to pretrained model directory (#771) * add a check for ubuntu 22.04 support (#721) * Changes to add bfloat16 support for DIEN training (#679) * Changes to add bfloat16 support for DIEN training * Some for for reporting performance * Fixes for dien training and unit tests * updated tpp file withr2.8 approvals (#773) * Add Windows stock PyTorch support for TransNet v2 (#779) * update TransNet v2 to work with stock pytorch * update Windows.md path in all relevant docs * add P99 metric for LZ models (#780) Co-authored-by: Weizhuo Zhang <[email protected]> * Rn50 training multiple epoches output 1 KPI and add training_steps argument. (#775) * enable --training_steps and 1 training KPI output with multiple epoches * add prefix * update print freq * fix display bug * enable PyTorch resnet50 fp16 path (#783) * enable PyTorch resnet50 fp16 path * fix conflict * Extract p99 metric from log to summary (#784) * enable fp16 bert train and inference (#782) * Vruddarr/pt update windows readmes (#778) * remove bfloat16 experimental support note (#786) * Update IPEX installation path (#788) * Clean up _pycache_ files, remove symlinks, and add license headers for dien training bf16 (#787) * update readme for jemalloc and iomp path (#789) * update readme for jemalloc and iomp path * Updated IOMP path as path to the intel-openmp directory * PyTorch: fix resnext101 running script (#795) * Update 3dunet mlperf bash scripts and README (#797) * update 3dunet mlperf doc to use quickstart scripts, rename quickstart scripts for multi-instance * fix tests job (#803) * rnnt inference: align replace lstm API due to IPEX change (#802) * Adding quick start scripts to MobileNetV1 bfloat16 precision (#793) * Adding quick start scripts to ssd-mobilenet bfloat16 precision (#798) * Update T5 model with windows quick start scripts (#790) * Update T5 model with windows quick start scripts * Updated Readme by specifying values to environment variables * Update inference int8 readme and script of 4 CV models using INC (#698) * update docs to add INC int8 models as an option * add instructions for how to quantize a fp32 model using INC * rnnt: fix stft due to PyTorch API change (#811) * rnnt training: fix stft due to PyTorch API change (#813) * Update BareMetalSetup.md (#817) * Gerardod/build container (#807) First phase of GHA WF to build the image of a Model Zoo workload container and push it to CAAS. * Sharvils/tf workload (#808) * TFv2.10 support added. Horovod version updated. * Vruddarr/tf add language translation bert fp32 quick start scripts (#804) * Adding quick start scripts to language translation BERT FP32 model * Updated TL notebooks for SPR Launch (#810) * Updates for TL PyTorch notebook * Edits for two more TL notebooks * Reverting previous change for virtualenv * Removed --no-deps and some nonexistent links * Added TFHub cache dir * Updated TL notebook README for legal/branding * Update typo in Readme (#821) Co-authored-by: veena.mounika.ruddarraju <[email protected]> * PyTorch: using ipex.optimize for bf16 training (#824) * Fix CVEs for Pillow and notebook packages (#831) Signed-off-by: Abolfazl Shahbazi <[email protected]> Signed-off-by: Abolfazl Shahbazi <[email protected]> * add intel-alphafold2 optimized w/ IPEX from realm of AIDD (#737) * add alphafold2 from AIDD realm * Remove unused variable in mlperf 3DUnet performance run (#832) * Update Model Zoo name, Python version and message for IPEX (#833) * Update instruction for Miniconda, Jemalloc, PyTorch and IPEX and updt… (#830) * Update models main tables (#836) *update main readmes * Adding jemalloc instructions and environment variables (#838) * Add support for dGPU models (#840) * add support for dGPU support * remove spr dockerfiles and spec files (#842) * delete links to 3dunet mlperf and bert large int8 (#841) * update tbb files (#843) * fix vulnerability issues reported by snyk scans (#848) * update for new precision (#849) * upgrade for ipex 1.13 * delete workflows Signed-off-by: Abolfazl Shahbazi <[email protected]> Co-authored-by: leslie-fang-intel <[email protected]> Co-authored-by: Dina Suehiro Jones <[email protected]> Co-authored-by: Abolfazl Shahbazi <[email protected]> Co-authored-by: XiaobingZhang <[email protected]> Co-authored-by: Xiaoming (Jason) Cui <[email protected]> Co-authored-by: jiayisunx <[email protected]> Co-authored-by: Melanie Buehler <[email protected]> Co-authored-by: Srini511 <[email protected]> Co-authored-by: Sean-Michael Riesterer <[email protected]> Co-authored-by: jianan-gu <[email protected]> Co-authored-by: Chunyuan WU <[email protected]> Co-authored-by: zhuhaozhe <[email protected]> Co-authored-by: Wang, Chuanqi <[email protected]> Co-authored-by: YanbingJiang <[email protected]> Co-authored-by: Weizhuo Zhang <[email protected]> Co-authored-by: xiaofeij <[email protected]> Co-authored-by: liangan1 <[email protected]> Co-authored-by: blzheng <[email protected]> Co-authored-by: Om Thakkar <[email protected]> Co-authored-by: mahathis <[email protected]> Co-authored-by: Clayne Robison <[email protected]> Co-authored-by: root <[email protected]> Co-authored-by: Neo Zhang Jianyu <[email protected]> Co-authored-by: ltsai1 <[email protected]> Co-authored-by: Jitendra Patil <[email protected]> Co-authored-by: Kanvi Khanna <[email protected]> Co-authored-by: Rahul Nair <[email protected]> Co-authored-by: Veena2207 <[email protected]> Co-authored-by: jojivk-intel-nervana <[email protected]> Co-authored-by: xiangdong <[email protected]> Co-authored-by: Huang, Zhiwei <[email protected]> Co-authored-by: gera-aldama <[email protected]> Co-authored-by: Sharvil Shah <[email protected]> Co-authored-by: wyang2 <[email protected]> Co-authored-by: Yimei Sun <[email protected]>
1 parent b63c719 commit a3f1cd1

File tree

879 files changed

+36203
-18945
lines changed

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

879 files changed

+36203
-18945
lines changed

.gitmodules

+3
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,3 @@
1+
[submodule "models/aidd/pytorch/alphafold2/inference/alphafold"]
2+
path = models/aidd/pytorch/alphafold2/inference/alphafold
3+
url = https://github.com/deepmind/alphafold

README.md

+65-58
Large diffs are not rendered by default.

benchmarks/README.md

+52-52
Large diffs are not rendered by default.

benchmarks/common/base_benchmark_util.py

+21-6
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
11
#
22
# -*- coding: utf-8 -*-
33
#
4-
# Copyright (c) 2018-2019 Intel Corporation
4+
# Copyright (c) 2018-2021 Intel Corporation
55
#
66
# Licensed under the Apache License, Version 2.0 (the "License");
77
# you may not use this file except in compliance with the License.
@@ -68,8 +68,8 @@ def _define_args(self):
6868

6969
self._common_arg_parser.add_argument(
7070
"-p", "--precision",
71-
help="Specify the model precision to use: fp32, int8, or bfloat16",
72-
required=required_arg, choices=["fp32", "int8", "bfloat16"],
71+
help="Specify the model precision to use: fp32, fp16, int8, or bfloat16",
72+
required=required_arg, choices=["fp32", "fp16", "int8", "bfloat16"],
7373
dest="precision")
7474

7575
self._common_arg_parser.add_argument(
@@ -172,7 +172,7 @@ def _define_args(self):
172172
dest="weight_sharing", action="store_true")
173173
self._common_arg_parser.add_argument(
174174
"--synthetic-data",
175-
help="Enables synthetic data layer for some models where support exists",
175+
help="Enables synthetic data layer for some models like SSD-ResNet34 where support exists",
176176
dest="synthetic_data", action="store_true")
177177

178178
self._common_arg_parser.add_argument(
@@ -274,6 +274,12 @@ def _define_args(self):
274274
help="Additional command line arguments (prefix flag start with"
275275
" '--').")
276276

277+
# Check if GPU is enabled.
278+
self._common_arg_parser.add_argument(
279+
"--gpu",
280+
help="Run the benchmark script using GPU",
281+
dest="gpu", action="store_true")
282+
277283
def _validate_args(self):
278284
"""validate the args and initializes platform_util"""
279285
# check if socket id is in socket number range
@@ -304,8 +310,9 @@ def _validate_args(self):
304310
format(system_num_cores))
305311

306312
if args.output_results and ((args.model_name != "resnet50" and
307-
args.model_name != "resnet50v1_5") or args.precision != "fp32"):
308-
raise ValueError("--output-results is currently only supported for resnet50 FP32 inference.")
313+
args.model_name != "resnet50v1_5") or
314+
(args.precision != "fp32" and args.precision != "fp16")):
315+
raise ValueError("--output-results is currently only supported for resnet50 FP32 or FP16 inference.")
309316
elif args.output_results and (args.mode != "inference" or not args.data_location):
310317
raise ValueError("--output-results can only be used when running inference with a dataset.")
311318

@@ -348,6 +355,14 @@ def _validate_args(self):
348355
"This is less than the number of cores per socket on the system ({})".
349356
format(args.socket_id, cpuset_len_for_socket, self._platform_util.num_cores_per_socket))
350357

358+
if args.gpu:
359+
if args.socket_id != -1:
360+
raise ValueError("--socket-id cannot be used with --gpu parameter.")
361+
if args.num_intra_threads is not None:
362+
raise ValueError("--num-intra-threads cannot be used with --gpu parameter.")
363+
if args.num_inter_threads is not None:
364+
raise ValueError("--num-inter-threads cannot be used with --gpu parameter.")
365+
351366
def initialize_model(self, args, unknown_args):
352367
"""Create model initializer for the specified model"""
353368
model_initializer = None

benchmarks/common/platform_util.py

+1-1
Original file line numberDiff line numberDiff line change
@@ -229,7 +229,7 @@ def _get_list_from_string_ranges(self, str_ranges):
229229
start, end = section.split("-")
230230
section_list = range(int(start), int(end) + 1)
231231
result_list += section_list
232-
elif(len(section)):
232+
elif len(section):
233233
# This section is either empty or just a single number and not a range
234234
result_list.append(int(section))
235235

0 commit comments

Comments
 (0)