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 ZooImage ClassificationPyTorch-CIFAR-ModelsSemantic SegmentationPIDNetPose EstimationYOLOv8 | ||
| 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. |
3.8 | 3.9 | 3.101.13.0 (recommended) (compatible with: 1.11.x - 1.13.x)2.8.0 (recommended) (compatible with: 2.3.x - 2.8.x)pip install netspresso
To use editable mode or docker, see INSTALLATION.md.
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")
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,
)
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",
)
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")
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")
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,
)
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,
)
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")
| Target / Source Framework | ONNX | TENSORFLOW_KERAS | TENSORFLOW |
|---|---|---|---|
| TENSORRT | ✔️ | ||
| DRPAI | ✔️ | ||
| OPENVINO | ✔️ | ||
| TENSORFLOW_LITE | ✔️ | ✔️ | ✔️ |
| 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 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,
)
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
)
| Module | Feature | Credit |
|---|---|---|
| Compressor | Automatic compression | 25 |
| Advanced compression | 50 | |
| Converter | Convert | 50 |
| Benchmarker | Benchmark | 25 |
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)!
FAQs
PyNetsPresso
We found that netspresso demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 2 open source maintainers collaborating on the project.
Did you know?
Socket for GitHub automatically highlights issues in each pull request and monitors the health of all your open source dependencies. Discover the contents of your packages and block harmful activity before you install or update your dependencies.
Security News
Research
The Socket Research Team breaks down a malicious wrapper package that uses obfuscation to harvest credentials and exfiltrate sensitive data.
Research
Security News
Attackers used a malicious npm package typosquatting a popular ESLint plugin to steal sensitive data, execute commands, and exploit developer systems.
Security News
The Ultralytics' PyPI Package was compromised four times in one weekend through GitHub Actions cache poisoning and failure to rotate previously compromised API tokens.