Skip to content

Latest commit

 

History

History
551 lines (443 loc) · 23.6 KB

README.md

File metadata and controls

551 lines (443 loc) · 23.6 KB
Raw




🌟 STMicro x NetsPresso 🌟
STM32 model zoo


Use NetsPresso for a seamless model optimization process. NetsPresso resolves AI-related constraints in business use cases and enables cost-efficiency and enhanced performance by removing the requirement for high-spec servers and network connectivity and preventing high latency and personal data breaches.

Easily compress various models with our resources. Please browse the Docs for details, and join our Discussion Forum for providing feedback or sharing your use cases.

To get started with NetsPresso, you'll need to sign up here.

We offer a comprehensive guide to walk you through the process of optimizing an AI model using NetsPresso. A full tutorial can be found Google Colab.




Step Type Description
Train np.trainer Build and train a model.
Model Zoo
Image Classification PyTorch-CIFAR-Models
Object Detection YOLO Fastest
YOLOX
YOLOv5
YOLOv7
Semantic Segmentation PIDNet
Pose Estimation YOLOv8
Compress np.compressor Compress and optimize the user’s model.
Quantize np.quantizer Quantize the user’s model.
Convert np.converter Convert and quantize the user’s model to run efficiently on device.
Benchmark np.benchmarker Benchmark the user's model to measure model inference speed on diverse device.

Installation

Prerequisites

  • Python 3.8 | 3.9 | 3.10
  • PyTorch 1.13.0 (recommended) (compatible with: 1.11.x - 1.13.x)
  • TensorFlow 2.8.0 (recommended) (compatible with: 2.3.x - 2.8.x)

Install with PyPI (stable)

pip install netspresso

To use editable mode or docker, see INSTALLATION.md.

Getting started

Login

Log-in to your netspresso account. Please sign-up here if you need one.

from netspresso import NetsPresso

netspresso = NetsPresso(email="YOUR_EMAIL", password="YOUR_PASSWORD")

⭐⭐⭐ (New Feature) Quantizer ⭐⭐⭐

Automatic quantization

To start quantize a model, enter the model path, dataset path, and the desired quantization precision.

The quantized model will be saved to the specified output directory (output_dir).

from netspresso.enums import QuantizationPrecision, SimilarityMetric

# 1. Declare quantizer
quantizer = netspresso.quantizer()

# 2. Run automatic quantization
quantization_result = quantizer.automatic_quantization(
    input_model_path="./examples/sample_models/test.onnx",
    output_dir="./outputs/quantized/automatic_quantization",
    dataset_path="./examples/sample_datasets/pickle_calibration_dataset_128x128.npy",
    weight_precision=QuantizationPrecision.INT8,
    activation_precision=QuantizationPrecision.INT8,
    threshold=0,
)

Custom precision quantization by layer name

This method enables you to apply precision settings tailored to each layer, based on the recommendations, to optimize model.

Or, you can modify it to your desired precision and optimize it.

from netspresso.enums import QuantizationPrecision

# 1. Declare quantizer
quantizer = netspresso.quantizer()

# 2. Recommendation precision
metadata = quantizer.get_recommendation_precision(
    input_model_path="./examples/sample_models/test.onnx",
    output_dir="./outputs/quantized/recommendation",
    dataset_path="./examples/sample_datasets/pickle_calibration_dataset_128x128.npy",
    weight_precision=QuantizationPrecision.INT8,
    activation_precision=QuantizationPrecision.INT8,
    threshold=0,
)
recommendation_precisions = quantizer.load_recommendation_precision_result(metadata.recommendation_result_path)

# 2. Run quantization by layer name
quantization_result = quantizer.custom_precision_quantization_by_layer_name(
    input_model_path="./examples/sample_models/test.onnx",
    output_dir="./outputs/quantized/custom_precision_quantization_by_layer_name",
    precision_by_layer_name=recommendation_precisions.layers,
    dataset_path="./examples/sample_datasets/pickle_calibration_dataset_128x128.npy",
)

Trainer

Train

To start training a model, first select a task.

Then configure the dataset, model, augmentation, and hyperparameters.

Once setup is finished, enter the GPU number and project name for training.

from netspresso.enums import Task
from netspresso.trainer.optimizers import AdamW
from netspresso.trainer.schedulers import CosineAnnealingWarmRestartsWithCustomWarmUp
from netspresso.trainer.augmentations import Resize


# 1. Declare trainer
trainer = netspresso.trainer(task=Task.OBJECT_DETECTION)  # IMAGE_CLASSIFICATION, OBJECT_DETECTION, SEMANTIC_SEGMENTATION

# 2. Set config for training
# 2-1. Data
trainer.set_dataset_config(
    name="traffic_sign_config_example",
    root_path="/root/traffic-sign",
    train_image="images/train",
    train_label="labels/train",
    valid_image="images/valid",
    valid_label="labels/valid",
    id_mapping=["prohibitory", "danger", "mandatory", "other"],
)

# 2-2. Model
print(trainer.available_models)  # ['YOLOX-S', 'YOLOX-M', 'YOLOX-L', 'YOLOX-X']
trainer.set_model_config(model_name="YOLOX-S", img_size=512)

# 2-3. Augmentation
trainer.set_augmentation_config(
    train_transforms=[Resize()],
    inference_transforms=[Resize()],
)

# 2-4. Training
optimizer = AdamW(lr=6e-3)
scheduler = CosineAnnealingWarmRestartsWithCustomWarmUp(warmup_epochs=10)
trainer.set_training_config(
    epochs=40,
    batch_size=16,
    optimizer=optimizer,
    scheduler=scheduler,
)

# 3. Train
training_result = trainer.train(gpus="0, 1", project_name="PROJECT_TRAIN_SAMPLE")

Retrain

To start retraining a model, use hparams.yaml file which is one of the artifacts generated during the training of the original model.

Then, enter the compressed model path, which is an artifact of the compressor in fx_model_path.

Adjust the training hyperparameters as needed. (See 2-2. for detailed code.)

from netspresso.trainer.optimizers import AdamW

# 1. Declare trainer
trainer = netspresso.trainer(yaml_path="./temp/hparams.yaml")

# 2. Set config for retraining
# 2-1. FX Model
trainer.set_fx_model(fx_model_path="./temp/FX_MODEL_PATH.pt")

# 2-2. Training
optimizer = AdamW(lr=6e-3)
trainer.set_training_config(
    epochs=30,
    batch_size=16,
    optimizer=optimizer,
)

# 3. Train
retraining_result = trainer.train(gpus="0, 1", project_name="PROJECT_RETRAIN_SAMPLE")

Compressor

Compress (Automatic compression)

To start compressing a model, enter the model path to compress and the appropriate compression ratio.

The compressed model will be saved in the specified output directory (output_dir).

# 1. Declare compressor
compressor = netspresso.compressor_v2()

# 2. Run automatic compression
compression_result = compressor.automatic_compression(
    input_shapes=[{"batch": 1, "channel": 3, "dimension": [224, 224]}],
    input_model_path="./examples/sample_models/graphmodule.pt",
    output_dir="./outputs/compressed/pytorch_automatic_compression",
    compression_ratio=0.5,
)

Converter

Convert

To start converting a model, enter the model path to convert and the target framework and device name.

For NVIDIA GPUs and Jetson devices, enter the software version additionally due to the jetpack version.

The converted model will be saved in the specified output directory (output_dir).

from netspresso.enums import DeviceName, Framework, SoftwareVersion

# 1. Declare converter
converter = netspresso.converter_v2()

# 2. Run convert
conversion_result = converter.convert_model(
    input_model_path="./examples/sample_models/test.onnx",
    output_dir="./outputs/converted/TENSORRT_JETSON_AGX_ORIN_JETPACK_5_0_1",
    target_framework=Framework.TENSORRT,
    target_device_name=DeviceName.JETSON_AGX_ORIN,
    target_software_version=SoftwareVersion.JETPACK_5_0_1,
)

Benchmarker

Benchmark

To start benchmarking a model, enter the model path to benchmark and the target device name.

For NVIDIA GPUs and Jetson devices, device name and software version have to be matched with the target device of the conversion.

TensorRT Model has strong dependency with the device type and its jetpack version.

from netspresso.enums import DeviceName, SoftwareVersion

# 1. Declare benchmarker
benchmarker = netspresso.benchmarker_v2()

# 2. Run benchmark
benchmark_result = benchmarker.benchmark_model(
    input_model_path="./outputs/converted/TENSORRT_JETSON_AGX_ORIN_JETPACK_5_0_1/TENSORRT_JETSON_AGX_ORIN_JETPACK_5_0_1.trt",
    target_device_name=DeviceName.JETSON_AGX_ORIN,
    target_software_version=SoftwareVersion.JETPACK_5_0_1,
)
print(f"model inference latency: {benchmark_result.benchmark_result.latency} ms")
print(f"model gpu memory footprint: {benchmark_result.benchmark_result.memory_footprint_gpu} MB")
print(f"model cpu memory footprint: {benchmark_result.benchmark_result.memory_footprint_cpu} MB")
Supported options for Converter & Benchmarker

Frameworks that support conversion for model's framework

Target / Source Framework ONNX TENSORFLOW_KERAS TENSORFLOW
TENSORRT ✔️
DRPAI ✔️
OPENVINO ✔️
TENSORFLOW_LITE ✔️ ✔️ ✔️

Devices that support benchmarks for model's framework

Device / Framework ONNX TENSORRT TENSORFLOW_LITE DRPAI OPENVINO
RASPBERRY_PI_5 ✔️ ✔️
RASPBERRY_PI_4B ✔️ ✔️
RASPBERRY_PI_3B_PLUS ✔️ ✔️
RASPBERRY_PI_ZERO_W ✔️ ✔️
RASPBERRY_PI_ZERO_2W ✔️ ✔️
ARM_ETHOS_U_SERIES ✔️(only INT8)
ALIF_ENSEMBLE_E7_DEVKIT_GEN2 ✔️(only INT8)
RENESAS_RA8D1 ✔️(only INT8)
NXP_iMX93 ✔️(only INT8)
ARDUINO_NICLA_VISION ✔️(only INT8)
RENESAS_RZ_V2L ✔️ ✔️
RENESAS_RZ_V2M ✔️ ✔️
JETSON_NANO ✔️ ✔️
JETSON_TX2 ✔️ ✔️
JETSON_XAVIER ✔️ ✔️
JETSON_NX ✔️ ✔️
JETSON_AGX_ORIN ✔️ ✔️
JETSON_ORIN_NANO ✔️ ✔️
AWS_T4 ✔️ ✔️
INTEL_XEON_W_2233 ✔️

Software versions that support conversions and benchmarks for specific devices

Software Versions requires for Jetson Device. If you are using a different device, you do not need to enter it.

Software Version / Device JETSON_NANO JETSON_TX2 JETSON_XAVIER JETSON_NX JETSON_AGX_ORIN JETSON_ORIN_NANO
JETPACK_4_4_1 ✔️
JETPACK_4_6 ✔️ ✔️ ✔️ ✔️
JETPACK_5_0_1 ✔️
JETPACK_5_0_2 ✔️
JETPACK_6_0 ✔️

The code below is an example of using software version.

conversion_result = converter.convert_model(
    input_model_path=INPUT_MODEL_PATH,
    output_dir=OUTPUT_DIR,
    target_framework=Framework.TENSORRT,
    target_device_name=DeviceName.JETSON_AGX_ORIN,
    target_software_version=SoftwareVersion.JETPACK_5_0_1,
)
benchmark_result = benchmarker.benchmark_model(
    input_model_path=CONVERTED_MODEL_PATH,
    target_device_name=DeviceName.JETSON_AGX_ORIN,
    target_software_version=SoftwareVersion.JETPACK_5_0_1,
)

Hardware type that support benchmarks for specific devices

Benchmark and compare models with and without Arm Helium.

RENESAS_RA8D1 and ALIF_ENSEMBLE_E7_DEVKIT_GEN2 are available for use.

The benchmark results with Helium can be up to twice as fast as without Helium.

The code below is an example of using hardware type.

benchmark_result = benchmarker.benchmark_model(
    input_model_path=CONVERTED_MODEL_PATH,
    target_device_name=DeviceName.RENESAS_RA8D1,
    target_data_type=DataType.INT8,
    target_hardware_type=HardwareType.HELIUM
)

Guide to Credit Consumption by Module

Module Feature Credit
Compressor Automatic compression 25
Advanced compression 50
Converter Convert 50
Benchmarker Benchmark 25

Contact

Join our Discussion Forum for providing feedback or sharing your use cases, and if you want to talk more with Nota, please contact us here.
Or you can also do it via email(netspresso@nota.ai) or phone(+82 2-555-8659)!