Skip to content

Latest commit

 

History

History
98 lines (79 loc) · 2.75 KB

quantization.md

File metadata and controls

98 lines (79 loc) · 2.75 KB

Quantization

image-selection

  • File location: ${root_nestc}utils/imagenet_val_script/image_selection.py
  • image-selection generates a list file that contains paths of images for calibration.
  • Options you specify are the followings:
    • --train-images-dir
    • --num-of-images
    • --file-name
python image_selection.py --file-name=test.txt --num-of-images=1 --train-images=/home/jemin/hdd/imagenet/train_processed

Imagenet sorting by labels (Optional)

imagenet_sorting_by_labels.py arranges images depending on labels.

  • val12_label.txt contains labels of Imagenet
imagenet_sorting_by_labels.py <validation data dir> <validation labels file>

imagenet_topk_accuracy_driver_py3

imagenet_topk_accuracy_driver_py3.py provides the following features:

  • image preprocessing to fit 224x224x3 size of input tensor.
--validation-images-dir=/ssd/imagenet/train
--only-resize-and-save
--resize-mode=0

#output
Saving centered cropped input images: /ssd/imagenet/train/processed
0 / 1000
1 / 1000
2 / 1000
3 / 1000
4 / 1000
5 / 1000
6 / 1000
  • Accuracy measurement with validation-set.
--verbose # log 저장 + resize 이미지 저장
--batch-size # 성능향상 없음
--validation-images-dir=/Users/jeminlee/fake_imagenet/
# 검증 데이터셋 위치. 레이블 오름차순으로 디텍토리 안에 정렬
--image-classifier-cmd="/Users/jeminlee/development/forked_glow/NEST_Compiler2/nest_compiler/cmake-build-release/bin/image-classifier -m=/Users/jeminlee/development/onnx_models/resnet-v1/resnet18v1.onnx -model-input-name=data -image-mode=0to1 -use-imagenet-normalization -backend=CPU -compute-softmax -topk=5 -”
# - streaming option을 반드시 줌.
# image에서 normalization을 하지 않았으므로 여기서 줘야함.
# image-mode는 0~1
# -topk=5를 해서 top5 error까지 계산 하도록 함.

To easily use imagenet_topk_accuracy_driver_py3.py and handle outputs, the following shell scripts are provided.

Calibration

# resnet 18
cd /home/jemin/development/nest_compiler/utils/profile
./profile.sh

Parameter Search

cd /home/jemin/development/nest_compiler/utils/imagenet_val_script

# 1,000, Max, non-fusion
./resnet18.sh 3

# 10,000, KL, fusion
./resnet18.sh 4

# 10,000, KL, Fusion
./resnet18.sh 7

Show explored results

# All search results of resnet18 
ls | grep resnet18

# 3
ls | grep -E "nonfusion.*resnet18.*_1000_.*none" | xargs cat
# 4
ls | grep -E "nonfusion.*resnet18.*_10000_.*KL" | xargs cat
# 7
ls | grep -E "vtainterpreter-fusion.*resnet18.*_10000_.*KL" | xargs cat

# Print all results
cd /home/jemin/development/nest_compiler/utils/imagenet_val_script/quant_result
ls | grep resnet18 | xargs more