Attention: make sure you've pull this repo before run commands below
conda create -n legodnn python==3.6
conda activate legodnn then enter root path of this repo
pip install --upgrade pip setuptools
pip install -r environments.txt model supported are listed below :
- vgg16
- mobilenetV2
- resnet18
- inceptionV3
dataset supported :
- cifar10
- cifar100 ( default )
cd cv_task/image_classification/cifar train the teacher model to get model's checkpoint (which can instruct block train in LegoDNN)
Attention: You must customize the root_path & dataset_path, all your train set and report (each epoch's acc & loss) are saved in root_path
if there are ValueError (OOM) , you may need to decrease num of workers via --num_workers=7/6/5/4/3/2/1/0 (num_workers=0 always fix the error)
# train vgg16 on cifar100 for example
python cv_task/image_classification/cifar/main.py --model=vgg16 --root_path=$TEACHER_MODEL_PATH$ --dataset_path=$DATASET_PATH$ --dataset=cifar100 if you'd like to set other argumets , check via python cv_task/image_classification/cifar/main.py --help
what we need is the .pth file of the teacher model under $TEACHER_MODEL_PATH$/$DATASET$/$MODEL$/YYYY-MM-DD/HH-MM-SS/.pth
e.g. $TEACHER_MODEL_PATH$/cifar100/vgg16/2023-07-01/10-00-47/vgg16.pth
You can check teacher model's acc in $TEACHER_MODEL_PATH$/$DATASET$/$MODEL$/YYYY-MM-DD/HH-MM-SS/report.txt
The time spent in this process is logged in time_log/teacher_model_train_time
demo of models are listed below:
what you need to set is dataset_root_dir, checkpoint_path, num_workers
checkpoint_path is the '.pth' file path after teacher model training in last step
all the other settings are remained as default
if there are ValueError (OOM) , you may need to decrease num of workers via reset num_workers or use GPU with higher memory
before run the program, make sure the checkpoint's dataset is suitable with dataset_name in demo
To mesure the time LegoDNN use , I've write an additional component time_logger.py
you can use it in other project as well.
usage:
import cv_task.log_on_time.time_logger import time_logger
# set log_dir & log's tittle when init
time_logger_obj = time_logger(log_dir='./log', title='time log demo')
# 1. start the timer
time_logger_obj.start()
"""
program part 1
"""
# 2. lap (for several times)
time_logger_obj.lap(time_delta_name='step 1') # name segment of each part
"""
program part 2
"""
# 3. end (must do at the end of timer)
time_logger_obj.end(time_delta_name='step 2') # name segment of the last part Time log of LegoDNN's demo are saved under time_log/legodnn_execute_time
You can find the log & result under experiments/image_classification/log & experiments/image_classification/result