Skip to main content

pnnx

python wrapper of pnnx, only support python 3.7+ now.

Install from pip

pnnx is available as wheel packages for macOS, Windows and Linux distributions, you can install with pip:

pip install pnnx

Build & Install from source

Prerequisites

On Unix (Linux, OS X)

  • A compiler with C++14 support
  • CMake >= 3.4

On Mac

  • A compiler with C++14 support
  • CMake >= 3.4

On Windows

  • Visual Studio 2015 or higher
  • CMake >= 3.4

Build & install

  1. clone ncnn.
git clone https://github.com/Tencent/ncnn.git
  1. install pytorch

install pytorch according to https://pytorch.org/ . Anaconda is strongly recommended for example:

conda install pytorch
  1. install
cd /pathto/ncnntools/pnnx/python
python setup.py install

Note: If torchvision and pnnx2onnx are needed, you can set the following environment variables before 'python setup.py install' to enable them. e.g. on ubuntu:

export TORCHVISION_INSTALL_DIR="/project/torchvision"
export PROTOBUF_INCLUDE_DIR="/project/protobuf/include"
export PROTOBUF_LIBRARIES="/project/protobuf/lib64/libprotobuf.a"
export PROTOBUF_PROTOC_EXECUTABLE="/project/protobuf/bin/protoc" 

To do these, you must install Torchvision and Protobuf first.

Tests

cd /pathto/ncnn/tools/pnnx/python
pytest tests

Usage

  1. export model to pnnx
import torch
import torchvision.models as models
import pnnx

net = models.resnet18(pretrained=True)
x = torch.rand(1, 3, 224, 224)

# You could try disabling checking when torch tracing raises error
# opt_net = pnnx.export(net, "resnet18.pt", x, check_trace=False)
opt_net = pnnx.export(net, "resnet18.pt", x)
  1. convert existing model to pnnx
import torch
import pnnx

x = torch.rand(1, 3, 224, 224)
opt_net = pnnx.convert("resnet18.pt", x)

API Reference

  1. pnnx.export

model (torch.nn.Model): model to be exported.

ptpath (str): the torchscript name.

inputs (torch.Tensor of list of torch.Tensor) expected inputs of the model.

inputs2 (torch.Tensor of list of torch.Tensor) alternative inputs of the model. Usually, it is used with input_shapes to resolve dynamic shape.

input_shapes (Optional, list of int or list of list with int type inside) shapes of model inputs. It is used to resolve tensor shapes in model graph. for example, [1,3,224,224] for the model with only 1 input, [[1,3,224,224],[1,3,224,224]] for the model that have 2 inputs.

input_types (Optional, str or list of str) types of model inputs, it should have the same length with input_shapes. for example, "f32" for the model with only 1 input, ["f32", "f32"] for the model that have 2 inputs.

typename torch type
f32 torch.float32 or torch.float
f64 torch.float64 or torch.double
f16 torch.float16 or torch.half
u8 torch.uint8
i8 torch.int8
i16 torch.int16 or torch.short
i32 torch.int32 or torch.int
i64 torch.int64 or torch.long
c32 torch.complex32
c64 torch.complex64
c128 torch.complex128

input_shapes2 (Optional, list of int or list of list with int type inside) shapes of alternative model inputs, the format is identical to input_shapes. Usually, it is used with input_shapes to resolve dynamic shape (-1) in model graph.

input_types2 (Optional, str or list of str) types of alternative model inputs.

device (Optional, str, default="cpu") device type for the input in TorchScript model, cpu or gpu.

customop (Optional, str or list of str) list of Torch extensions (dynamic library) for custom operators. For example, "/home/nihui/.cache/torch_extensions/fused/fused.so" or ["/home/nihui/.cache/torch_extensions/fused/fused.so",...].

moduleop (Optional, str or list of str) list of modules to keep as one big operator. for example, "models.common.Focus" or ["models.common.Focus","models.yolo.Detect"].

optlevel (Optional, int, default=2) graph optimization level

option optimization level
0 do not apply optimization
1 do not apply optimization
2 optimization more for inference

pnnxparam (Optional, str, default="*.pnnx.param", * is the model name): PNNX graph definition file.

pnnxbin (Optional, str, default="*.pnnx.bin"): PNNX model weight.

pnnxpy (Optional, str, default="*_pnnx.py"): PyTorch script for inference, including model construction and weight initialization code.

pnnxonnx (Optional, str, default="*.pnnx.onnx"): PNNX model in onnx format.

ncnnparam (Optional, str, default="*.ncnn.param"): ncnn graph definition.

ncnnbin (Optional, str, default="*.ncnn.bin"): ncnn model weight.

ncnnpy (Optional, str, default="*_ncnn.py"): pyncnn script for inference.

  1. pnnx.convert

ptpath (str): torchscript model to be converted.

Other parameters are consistent with pnnx.export

Metadata

Release files for pnnx 20261020

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Built distributions (wheels)

Table of built distributions (wheels) for pnnx 20261020
File Actions
pnnx-20261020-py3-none-win_amd64.whl Python 3 none Windows x86-64 Details for pnnx-20261020-py3-none-win_amd64.whl
pnnx-20261020-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl Python 3 none Linux glibc 2.17+ x86-64 Details for pnnx-20261020-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl
pnnx-20261020-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl Python 3 none Linux glibc 2.17+ ARM64 Details for pnnx-20261020-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl
pnnx-20261020-py3-none-macosx_10_9_universal2.macosx_10_9_x86_64.macosx_11_0_arm64.whl Python 3 none macOS 10.9+ universal2 (ARM64, x86-64), macOS 10.9+ x86-64, macOS 11.0+ ARM64 Details for pnnx-20261020-py3-none-macosx_10_9_universal2.macosx_10_9_x86_64.macosx_11_0_arm64.whl

Total release size: 127.7 MB

Release files / pnnx-20261020-py3-none-win_amd64.whl

Download URL pnnx-20261020-py3-none-win_amd64.whl
Size 18.7 MB
Tags Python 3 Windows x86-64
SHA-256 checksum
How to use checksums
cde5381cdebbef5d3e0e796e824ed02e2e1dfd075731921a3b3653514c518309
BLAKE2b-256 checksum
How to use checksums
e8511b30523f85f9482f55f3f4f11ad8f4a37db7c8773d66625cceb2973c5c60
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.14

Release files / pnnx-20261020-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl

Download URL pnnx-20261020-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl
Size 29.4 MB
Tags Linux glibc 2.17+ x86-64 Python 3
SHA-256 checksum
How to use checksums
d50d921e5897293d616444191c1a2dc5366423c81bdb4bd22a985fe8071cf17c
BLAKE2b-256 checksum
How to use checksums
5cb79665ea73f6503fea4f30abbd3eebcea1c8500095496167761a15482354bc
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.14

Release files / pnnx-20261020-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl

Download URL pnnx-20261020-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl
Size 24.7 MB
Tags Linux glibc 2.17+ ARM64 Python 3
SHA-256 checksum
How to use checksums
13727adb291bf8061a4cc95c8b911ff2459b406144ad3c86e755b6e3f8953427
BLAKE2b-256 checksum
How to use checksums
0a85ba5d0b9b632822b63354fa03d6f84e1948528068897bb2cef64398c99ba9
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.14

Release files / pnnx-20261020-py3-none-macosx_10_9_universal2.macosx_10_9_x86_64.macosx_11_0_arm64.whl

Download URL pnnx-20261020-py3-none-macosx_10_9_universal2.macosx_10_9_x86_64.macosx_11_0_arm64.whl
Size 54.8 MB
Tags Python 3 macOS 10.9+ universal2 (ARM64, x86-64) macOS 10.9+ x86-64 macOS 11.0+ ARM64
SHA-256 checksum
How to use checksums
f4d60b7df5d1f7329f60cf75018ab4e25006db6d2299cf5ef8d56c051044660b
BLAKE2b-256 checksum
How to use checksums
8ce798373b217be926b9a6895c9316c350f97fad8bedeb64e5f43cc99c0f498e
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.14

Security for pnnx 20261020

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release

20261020 This release

4 release files No provenance

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page