-
Notifications
You must be signed in to change notification settings - Fork 86
[WIP] Add the config file of parameter tuner for object detection #275
base: master
Are you sure you want to change the base?
Conversation
Optimizer Selector へのパッチ
xml parser use xml.etree.ElementTree instead of BeautifulSoup4
Added pattern matching and new optimization passes
Optimizer Selector の PR コメント対応
* add new file for ray tune * add config file for ray tune * add optuna-based hp optimization contributed by @djKooks * improve print for more trial detail * casual clean-up on tune_optuna * adapt more hyper-parameters for tuning * add search for weight decay rate parameter * make optimizer tunable * reloading the config file every time new trial start * clear kwargs before assigning new ones * add perimagestandardization to tune config file * add search option for lr scheduler of polynomial decay * move tune_spec to config file * remove redundant imports * add dynamic allocation of resources * fix incompatible problem of easydict * make compatible with data iterator * auto detect number of gpus to train * test lmnet_v1_quantize_cifar10 * add options for saving directory * saving directory defaults to ~/ray_results * add config file for tuning segmentation * add optional id for different tuning sessions * oops, no weight decay parameter for lm_segnet * remove redundant lines * remove more redundant lines * test various values in config file * update package requirement for ray * separate core scripts and config files * remove un-used imports * remove optuna implementation * quick update comments * update copyright year * add some comments * Rename trial to shorter string
* Add try catch to ensure queue import compatibility * fix queue to Queue typo * use from instead of direct import
* add build dependency to pep8-dlk * fix to follow pep8
* create automated_testing dir on FPGA for dlk test * change to use default python(2.7.12) on FPGA for dlk test * remove Type Hints for python 2.7 * enable to set fpga host by environment variables * add FPGA_HOST from environment variable * add StrictHostKeyChecking option instead of ssh configuration via volume mount of /root/.ssh * fix mount.err to mount.out * remove comment 'only available on Jenkins' * enable to connect to FPGA by user's ssh config and keys * fix 'Bad owner or permissions on /root/.ssh/config'
* Update docs for CUDA10 * fix > to >=
…s possible even if the source code is changed. (blue-oil#273)
* allow dataset prefetch only for training * remove to-do flag * add comments
Super Nice PR!! |
|
|
Motivation and Context
Previously, there were not config file provided for object detection of the experimental parameter tuner. This PR provides it.
Description
lm_yolo
model inlmnet
with namelm_fyolo_quantize_pascalvoc_2007_2012_tune.py
.soft_start
is provided to support the 'warm-up' step.How has this been tested?
Screenshots (if appropriate):
Types of changes
Checklist: