Enstriksyon dokimantasyon Qualcomm Aimet Efficiency Toolkit

KBA-231226181840

1. Enstalasyon Envirnment

1.1. Enstale Nvidia Driver ak CUDA

1.2. Enstale Bibliyotèk Python ki gen rapò

python3 -m pip install –upgrade –ignore-enstale pip
python3 -m pip install –ignore-enstale gdown
python3 -m pip install –ignore-installed opencv-python
python3 -m pip install –ignore-installed torch==1.9.1+cu111 torchvision==0.10.1+cu111 torchaudio==0.9.1 -f https://download.pytorch.org/whl/torch_stable.html
python3 -m pip install –ignore-installed jax
python3 -m pip install –ignore-installed ftfy
python3 -m pip install –ignore-installed torchiinfo
python3 -m pip install –ignore-installed https://github.com/quic/aimet/releases/download/1.25.0/AimetCommon-torch_gpu_1.25.0-cp38-cp38-linux_x86_64.whl
python3 -m pip install –ignore-installed https://github.com/quic/aimet/releases/download/1.25.0/AimetTorch-torch_gpu_1.25.0-cp38-cp38-linux_x86_64.whl
python3 -m pip install –ignore-installed numpy==1.21.6
python3 -m pip install –ignore-enstale psutil

1.3. Klone aimet-model-zoo

git klon https://github.com/quic/aimet-model-zoo.git
cd aimet-model-zoo
git checkout d09d2b0404d10f71a7640a87e9d5e5257b028802
ekspòte PYTHONPATH=${PYTHONPATH}:${PWD}

1.4. Telechaje Set14

wget https://uofi.box.com/shared/static/igsnfieh4lz68l926l8xbklwsnnk8we9.zip
unzip igsnfieh4lz68l926l8xbklwsnnk8we9.zip

1.5. Modifye liy 39 aimt-model-zoo/aimet_zoo_torch/quicksrnet/dataloader/utils.py

chanje
pou img_path nan glob.glob(os.path.join(test_images_dir, "*")):
pou
pou img_path nan glob.glob(os.path.join(test_images_dir, "*_HR.*")):

1.6. Kouri evalyasyon.

# kouri anba YOURPATH/aimet-model-run
# Pou quicksrnet_small_2x_w8a8
python3 aimet_zoo_torch/quicksrnet/evaluators/quicksrnet_quanteval.py \
–model-config quicksrnet_small_2x_w8a8 \
–dataset-path ../Set14/image_SRF_4

# Pou quicksrnet_small_4x_w8a8
python3 aimet_zoo_torch/quicksrnet/evaluators/quicksrnet_quanteval.py \
–model-config quicksrnet_small_4x_w8a8 \
–dataset-path ../Set14/image_SRF_4

# Pou quicksrnet_medium_2x_w8a8
python3 aimet_zoo_torch/quicksrnet/evaluators/quicksrnet_quanteval.py \
–model-config quicksrnet_medium_2x_w8a8 \
–dataset-path ../Set14/image_SRF_4

# Pou quicksrnet_medium_4x_w8a8
python3 aimet_zoo_torch/quicksrnet/evaluators/quicksrnet_quanteval.py \
–model-config quicksrnet_medium_4x_w8a8 \
–dataset-path ../Set14/image_SRF_4

sipoze w ap jwenn valè PSNR pou modèl simulation. Ou ka chanje modèl-config pou diferan gwosè QuickSRNet, opsyon a se underaimet-modelzoo/aimet_zoo_torch/quicksrnet/model/model_cards/.

2 Ajoute Patch

2.1. Louvri "Export to ONNX Steps REVISED.docx"

2.2. Sote id komèt git

2.3. Seksyon 1 Kòd

Ajoute tout 1. kòd anba dènye liy (apre liy 366) aimet-model-zoo/aimet_zoo_torch/quicksrnet/model/models.py

2.4. Seksyon 2 ak 3 Kòd

Ajoute tout kòd 2, 3 anba liy 93 aimt-model-zoo/aimet_zoo_torch/quicksrnet/evaluators/quicksrnet_quanteval.py

2.5. Paramèt kle nan Fonksyon load_model

modèl = load_model (MODEL_PATH_INT8,

MODEL_NAME,
MODEL_ARGS.get(MODEL_NAME).get(MODEL_CONFIG),
use_quant_sim_model=Vre,
encoding_path=ENCODING_PATH,
quantsim_config_path=CONFIG_PATH,
calibration_data=IMAGES_LR,
use_cuda=Se vre,
before_quantization=Vre,
convert_to_dcr=Vre)

MODEL_PATH_INT8 = aimet_zoo_torch/quicksrnet/model/weights/quicksrnet_small_2x_w8a8/pre_opt_weights
MODEL_NAME = QuickSRNetSmall
MODEL_ARGS.get(MODEL_NAME).get(MODEL_CONFIG) = {'scaling_factor': 2}
ENCODING_PATH = aimet_zoo_torch/quicksrnet/model/weights/quicksrnet_small_2x_w8a8/adaround_encodings
CONFIG_PATH = aimet_zoo_torch/quicksrnet/model/weights/quicksrnet_small_2x_w8a8/aimet_config

Tanpri ranplase varyab yo pou diferan gwosè QuickSRNet

2.6 Modifikasyon gwosè modèl

1. "input_shape" nan aimet-model-zoo/aimet_zoo_torch/quicksrnet/model/model_cards/*.json
2. Anndan fonksyon load_model(…) nan aimet-model-zoo/aimet_zoo_torch/quicksrnet/model/inference.py
3. Paramèt andedan fonksyon export_to_onnx(…, input_height, input_width) soti nan "Ekspòtasyon pou ONNX Etap REVISED.docx"

2.7 Re-Run 1.6 ankò pou ekspòte modèl ONNX

3. Konvèti nan SNPE

3.1. Konvèti

${SNPE_ROOT}/bin/x86_64-linux-clang/snpe-onnx-to-dlc \
–input_network model.onnx \
–quantization_overrides ./model.encodings

3.2. (Si ou vle) Ekstrè sèlman DLC quantized

(opsyonèl) snpe-dlc-quant –input_dlc model.dlc –float_fallback –override_params

3.3. (ENPÒTAN) I/O ONNX nan lòd NCHW; DLC konvèti a se nan lòd NHWC

Dokiman / Resous

Dokimantasyon Qualcomm Aimet Efficiency Toolkit [pdfEnstriksyon yo quicksrnet_small_2x_w8a8, quicksrnet_small_4x_w8a8, quicksrnet_medium_2x_w8a8, quicksrnet_medium_4x_w8a8, Aimet Efficiency Toolkit Dokimantasyon, Efikasite Toolkit Dokimantasyon, Toolkit Dokimantasyon, Dokimantasyon

Referans

• Manyèl itilizatè