ncnn/tools/pnnx

• PNNX
• Rationale
• Features
• Build TorchScript to PNNX converter
• Usage
• The pnnx.param format
• • • example
    • overview
    • operator line
• The pnnx.bin format
• PNNX operator
• PNNX expression operator
• PNNX torch function operator
• PNNX module operator
• PNNX python inference
• PNNX shape propagation
• PNNX model optimization
• PNNX custom operator
• Supported PyTorch operator status

PNNX

PyTorch Neural Network eXchange(PNNX) is an open standard for PyTorch model interoperability. PNNX provides an open model format for PyTorch. It defines computation graph as well as high level operators strictly matches PyTorch.

Rationale

PyTorch is currently one of the most popular machine learning frameworks. We need to deploy the trained AI model to various hardware and environments more conveniently and easily.

Before PNNX, we had the following methods:

1. export to ONNX, and deploy with ONNX-runtime
2. export to ONNX, and convert onnx to inference-framework specific format, and deploy with TensorRT/OpenVINO/ncnn/etc.
3. export to TorchScript, and deploy with libtorch

As far as we know, ONNX has the ability to express the PyTorch model and it is an open standard. People usually use ONNX as an intermediate representation between PyTorch and the inference platform. However, ONNX still has the following fatal problems, which makes the birth of PNNX necessary:

1. ONNX does not have a human-readable and editable file representation, making it difficult for users to easily modify the computation graph or add custom operators.
2. The operator definition of ONNX is not completely in accordance with PyTorch. When exporting some PyTorch operators, glue operators are often added passively by ONNX, which makes the computation graph inconsistent with PyTorch and may impact the inference efficiency.
3. There are a large number of additional parameters designed to be compatible with various ML frameworks in the operator definition in ONNX. These parameters increase the burden of inference implementation on hardware and software.

PNNX tries to define a set of operators and a simple and easy-to-use format that are completely contrasted with the python api of PyTorch, so that the conversion and interoperability of PyTorch models are more convenient.

Features

1. Human readable and editable format
2. Plain model binary in storage zip
3. One-to-one mapping of PNNX operators and PyTorch python api
4. Preserve math expression as one operator
5. Preserve torch function as one operator
6. Preserve miscellaneous module as one operator
7. Inference via exported PyTorch python code
8. Tensor shape propagation
9. Model optimization
10. Custom operator support

Build TorchScript to PNNX converter

1. Install PyTorch and TorchVision c++ library
2. Build PNNX with cmake

Usage

1. Export your model to TorchScript

import torch
import torchvision.models as models

net = models.resnet18(pretrained=True)
net = net.eval()

x = torch.rand(1, 3, 224, 224)

mod = torch.jit.trace(net, x)
torch.jit.save(mod, "resnet18.pt")

2. Convert TorchScript to PNNX

pnnx resnet18.pt inputshape=[1,3,224,224]

Normally, you will get six files

resnet18.pnnx.param PNNX graph definition

resnet18.pnnx.bin PNNX model weight

resnet18_pnnx.py PyTorch script for inference, the python code for model construction and weight initialization

resnet18.ncnn.param ncnn graph definition

resnet18.ncnn.bin ncnn model weight

resnet18_ncnn.py pyncnn script for inference

3. Visualize PNNX with Netron

Open https://netron.app/ in browser, and drag resnet18.pnnx.param into it.

4. PNNX command line options

Usage: pnnx [model.pt] [(key=value)...]
  pnnxparam=model.pnnx.param
  pnnxbin=model.pnnx.bin
  pnnxpy=model_pnnx.py
  ncnnparam=model.ncnn.param
  ncnnbin=model.ncnn.bin
  optlevel=2
  device=cpu/gpu
  inputshape=[1,3,224,224],...
  inputshape2=[1,3,320,320],...
  customop=/home/nihui/.cache/torch_extensions/fused/fused.so,...
  moduleop=models.common.Focus,models.yolo.Detect,...
Sample usage: pnnx mobilenet_v2.pt inputshape=[1,3,224,224]
              pnnx yolov5s.pt inputshape=[1,3,640,640] inputshape2=[1,3,320,320] device=gpu moduleop=models.common.Focus,models.yolo.Detect

The pnnx.param format

example

7767517
4 3
pnnx.Input      input       0 1 0
nn.Conv2d       conv_0      1 1 0 1 bias=1 dilation=(1,1) groups=1 in_channels=12 kernel_size=(3,3) out_channels=16 padding=(0,0) stride=(1,1) @bias=(16)f32 @weight=(16,12,3,3)f32
nn.Conv2d       conv_1      1 1 1 2 bias=1 dilation=(1,1) groups=1 in_channels=16 kernel_size=(2,2) out_channels=20 padding=(2,2) stride=(2,2) @bias=(20)f32 @weight=(20,16,2,2)f32
pnnx.Output     output      1 0 2

overview

[magic]

• magic number : 7767517

[operator count] [operand count]

• operator count : count of the operator line follows
• operand count : count of all operands

operator line

[type] [name] [input count] [output count] [input operands] [output operands] [operator params]

• type : type name, such as Conv2d ReLU etc
• name : name of this operator
• input count : count of the operands this operator needs as input
• output count : count of the operands this operator produces as output
• input operands : name list of all the input blob names, separated by space
• output operands : name list of all the output blob names, separated by space
• operator params : key=value pair list, separated by space, operator weights are prefixed by @ symbol, tensor shapes are prefixed by # symbol, input parameter keys are prefixed by $

The pnnx.bin format

pnnx.bin file is a zip file with store-only mode(no compression)

weight binary file has its name composed by operator name and weight name

For example, nn.Conv2d conv_0 1 1 0 1 bias=1 dilation=(1,1) groups=1 in_channels=12 kernel_size=(3,3) out_channels=16 padding=(0,0) stride=(1,1) @bias=(16) @weight=(16,12,3,3) would pull conv_0.weight and conv_0.bias into pnnx.bin zip archive.

weight binaries can be listed or modified with any archive application eg. 7zip

PNNX operator

PNNX always preserve operators from what PyTorch python api provides.

Here is the netron visualization comparision among ONNX, TorchScript and PNNX with the original PyTorch python code shown.

import torch
import torch.nn as nn

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()

        self.attention = nn.MultiheadAttention(embed_dim=256, num_heads=32)

    def forward(self, x):
        x, _ = self.attention(x, x, x)
        return x

ONNX	TorchScript	PNNX

PNNX expression operator

PNNX trys to preserve expression from what PyTorch python code writes.

Here is the netron visualization comparision among ONNX, TorchScript and PNNX with the original PyTorch python code shown.

import torch

def foo(x, y):
    return torch.sqrt((2 * x + y) / 12)

ONNX	TorchScript	PNNX

PNNX torch function operator

PNNX trys to preserve torch functions and Tensor member functions as one operator from what PyTorch python api provides.

Here is the netron visualization comparision among ONNX, TorchScript and PNNX with the original PyTorch python code shown.

import torch
import torch.nn.functional as F

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()

    def forward(self, x):
        x = F.normalize(x, eps=1e-3)
        return x

ONNX	TorchScript	PNNX

PNNX module operator

Users could ask PNNX to keep module as one big operator when it has complex logic.

The process is optional and could be enabled via moduleop command line option.

After pass_level0, all modules will be presented in terminal output, then you can pick the intersting ones as module operators.

############# pass_level0
inline module = models.common.Bottleneck
inline module = models.common.C3
inline module = models.common.Concat
inline module = models.common.Conv
inline module = models.common.Focus
inline module = models.common.SPP
inline module = models.yolo.Detect
inline module = utils.activations.SiLU

pnnx yolov5s.pt inputshape=[1,3,640,640] moduleop=models.common.Focus,models.yolo.Detect

Here is the netron visualization comparision among ONNX, TorchScript and PNNX with the original PyTorch python code shown.

import torch
import torch.nn as nn

class Focus(nn.Module):
    # Focus wh information into c-space
    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
        super().__init__()
        self.conv = Conv(c1 * 4, c2, k, s, p, g, act)

    def forward(self, x):  # x(b,c,w,h) -> y(b,4c,w/2,h/2)
        return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))

ONNX	TorchScript	PNNX	PNNX with module operator

PNNX python inference

A python script will be generated by default when converting torchscript to pnnx.

This script is the python code representation of PNNX and can be used for model inference.

There are some utility functions for loading weight binary from pnnx.bin.

You can even export the model torchscript AGAIN from this generated code!

import torch
import torch.nn as nn
import torch.nn.functional as F

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()

        self.linear_0 = nn.Linear(in_features=128, out_features=256, bias=True)
        self.linear_1 = nn.Linear(in_features=256, out_features=4, bias=True)

    def forward(self, x):
        x = self.linear_0(x)
        x = F.leaky_relu(x, 0.15)
        x = self.linear_1(x)
        return x

import os
import numpy as np
import tempfile, zipfile
import torch
import torch.nn as nn
import torch.nn.functional as F

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()

        self.linear_0 = nn.Linear(bias=True, in_features=128, out_features=256)
        self.linear_1 = nn.Linear(bias=True, in_features=256, out_features=4)

        archive = zipfile.ZipFile('../../function.pnnx.bin', 'r')
        self.linear_0.bias = self.load_pnnx_bin_as_parameter(archive, 'linear_0.bias', (256), 'float32')
        self.linear_0.weight = self.load_pnnx_bin_as_parameter(archive, 'linear_0.weight', (256,128), 'float32')
        self.linear_1.bias = self.load_pnnx_bin_as_parameter(archive, 'linear_1.bias', (4), 'float32')
        self.linear_1.weight = self.load_pnnx_bin_as_parameter(archive, 'linear_1.weight', (4,256), 'float32')
        archive.close()

    def load_pnnx_bin_as_parameter(self, archive, key, shape, dtype):
        return nn.Parameter(self.load_pnnx_bin_as_tensor(archive, key, shape, dtype))

    def load_pnnx_bin_as_tensor(self, archive, key, shape, dtype):
        _, tmppath = tempfile.mkstemp()
        tmpf = open(tmppath, 'wb')
        with archive.open(key) as keyfile:
            tmpf.write(keyfile.read())
        tmpf.close()
        m = np.memmap(tmppath, dtype=dtype, mode='r', shape=shape).copy()
        os.remove(tmppath)
        return torch.from_numpy(m)

    def forward(self, v_x_1):
        v_7 = self.linear_0(v_x_1)
        v_input_1 = F.leaky_relu(input=v_7, negative_slope=0.150000)
        v_12 = self.linear_1(v_input_1)
        return v_12

PNNX shape propagation

Users could ask PNNX to resolve all tensor shapes in model graph and constify some common expressions involved when tensor shapes are known.

The process is optional and could be enabled via inputshape command line option.

pnnx shufflenet_v2_x1_0.pt inputshape=[1,3,224,224]

def channel_shuffle(x: Tensor, groups: int) -> Tensor:
    batchsize, num_channels, height, width = x.size()
    channels_per_group = num_channels // groups

    # reshape
    x = x.view(batchsize, groups, channels_per_group, height, width)

    x = torch.transpose(x, 1, 2).contiguous()

    # flatten
    x = x.view(batchsize, -1, height, width)

    return x

without shape propagation	with shape propagation

PNNX model optimization

ONNX	TorchScript	PNNX without optimization	PNNX with optimization

PNNX custom operator

import os

import torch
from torch.autograd import Function
from torch.utils.cpp_extension import load, _import_module_from_library

module_path = os.path.dirname(__file__)
upfirdn2d_op = load(
    'upfirdn2d',
    sources=[
        os.path.join(module_path, 'upfirdn2d.cpp'),
        os.path.join(module_path, 'upfirdn2d_kernel.cu'),
    ],
    is_python_module=False
)

def upfirdn2d(input, kernel, up=1, down=1, pad=(0, 0)):
    pad_x0 = pad[0]
    pad_x1 = pad[1]
    pad_y0 = pad[0]
    pad_y1 = pad[1]

    kernel_h, kernel_w = kernel.shape
    batch, channel, in_h, in_w = input.shape

    input = input.reshape(-1, in_h, in_w, 1)

    out_h = (in_h * up + pad_y0 + pad_y1 - kernel_h) // down + 1
    out_w = (in_w * up + pad_x0 + pad_x1 - kernel_w) // down + 1

    out = torch.ops.upfirdn2d_op.upfirdn2d(input, kernel, up, up, down, down, pad_x0, pad_x1, pad_y0, pad_y1)

    out = out.view(-1, channel, out_h, out_w)

    return out

#include <torch/extension.h>

torch::Tensor upfirdn2d(const torch::Tensor& input, const torch::Tensor& kernel,
                        int64_t up_x, int64_t up_y, int64_t down_x, int64_t down_y,
                        int64_t pad_x0, int64_t pad_x1, int64_t pad_y0, int64_t pad_y1) {
    // operator body
}

TORCH_LIBRARY(upfirdn2d_op, m) {
    m.def("upfirdn2d", upfirdn2d);
}

Supported PyTorch operator status

torch.nn	Is Supported	Export to ncnn
nn.AdaptiveAvgPool1d	✔️	✔️
nn.AdaptiveAvgPool2d	✔️	✔️
nn.AdaptiveAvgPool3d	✔️	✔️
nn.AdaptiveMaxPool1d	✔️	✔️
nn.AdaptiveMaxPool2d	✔️	✔️
nn.AdaptiveMaxPool3d	✔️	✔️
nn.AlphaDropout	✔️	✔️
nn.AvgPool1d	✔️	✔️*
nn.AvgPool2d	✔️	✔️*
nn.AvgPool3d	✔️	✔️*
nn.BatchNorm1d	✔️	✔️
nn.BatchNorm2d	✔️	✔️
nn.BatchNorm3d	✔️	✔️
nn.Bilinear
nn.CELU	✔️
nn.ChannelShuffle	✔️	✔️
nn.ConstantPad1d	✔️	✔️
nn.ConstantPad2d	✔️	✔️
nn.ConstantPad3d	✔️	✔️
nn.Conv1d	✔️	✔️
nn.Conv2d	✔️	✔️
nn.Conv3d	✔️	✔️
nn.ConvTranspose1d	✔️
nn.ConvTranspose2d	✔️	✔️
nn.ConvTranspose3d	✔️
nn.CosineSimilarity
nn.Dropout	✔️	✔️
nn.Dropout2d	✔️	✔️
nn.Dropout3d	✔️	✔️
nn.ELU	✔️	✔️
nn.Embedding	✔️	✔️
nn.EmbeddingBag
nn.Flatten	✔️
nn.Fold
nn.FractionalMaxPool2d
nn.FractionalMaxPool3d
nn.GELU	✔️	✔️
nn.GroupNorm	✔️	✔️
nn.GRU	✔️	✔️
nn.GRUCell
nn.Hardshrink	✔️
nn.Hardsigmoid	✔️	✔️
nn.Hardswish	✔️	✔️
nn.Hardtanh	✔️	✔️
nn.Identity
nn.InstanceNorm1d	✔️
nn.InstanceNorm2d	✔️	✔️
nn.InstanceNorm3d	✔️
nn.LayerNorm	✔️	✔️
nn.LazyBatchNorm1d
nn.LazyBatchNorm2d
nn.LazyBatchNorm3d
nn.LazyConv1d
nn.LazyConv2d
nn.LazyConv3d
nn.LazyConvTranspose1d
nn.LazyConvTranspose2d
nn.LazyConvTranspose3d
nn.LazyLinear
nn.LeakyReLU	✔️	✔️
nn.Linear	✔️	✔️
nn.LocalResponseNorm	✔️	✔️
nn.LogSigmoid	✔️
nn.LogSoftmax	✔️
nn.LPPool1d	✔️
nn.LPPool2d	✔️
nn.LSTM	✔️	✔️
nn.LSTMCell
nn.MaxPool1d	✔️	✔️
nn.MaxPool2d	✔️	✔️
nn.MaxPool3d	✔️	✔️
nn.MaxUnpool1d
nn.MaxUnpool2d
nn.MaxUnpool3d
nn.Mish	✔️	✔️
nn.MultiheadAttention	✔️	✔️*
nn.PairwiseDistance
nn.PixelShuffle	✔️	✔️
nn.PixelUnshuffle	✔️	✔️
nn.PReLU	✔️	✔️
nn.ReflectionPad1d	✔️	✔️
nn.ReflectionPad2d	✔️	✔️
nn.ReLU	✔️	✔️
nn.ReLU6	✔️	✔️
nn.ReplicationPad1d	✔️	✔️
nn.ReplicationPad2d	✔️	✔️
nn.ReplicationPad3d	✔️
nn.RNN	✔️	✔️*
nn.RNNBase
nn.RNNCell
nn.RReLU	✔️
nn.SELU	✔️	✔️
nn.Sigmoid	✔️	✔️
nn.SiLU	✔️	✔️
nn.Softmax	✔️	✔️
nn.Softmax2d
nn.Softmin	✔️
nn.Softplus	✔️
nn.Softshrink	✔️
nn.Softsign	✔️
nn.SyncBatchNorm
nn.Tanh	✔️	✔️
nn.Tanhshrink	✔️
nn.Threshold	✔️
nn.Transformer
nn.TransformerDecoder
nn.TransformerDecoderLayer
nn.TransformerEncoder
nn.TransformerEncoderLayer
nn.Unflatten
nn.Unfold
nn.Upsample	✔️	✔️
nn.UpsamplingBilinear2d	✔️	✔️
nn.UpsamplingNearest2d	✔️	✔️
nn.ZeroPad2d	✔️	✔️

torch.nn.functional	Is Supported	Export to ncnn
F.adaptive_avg_pool1d	✔️	✔️
F.adaptive_avg_pool2d	✔️	✔️
F.adaptive_avg_pool3d	✔️	✔️
F.adaptive_max_pool1d	✔️	✔️
F.adaptive_max_pool2d	✔️	✔️
F.adaptive_max_pool3d	✔️	✔️
F.affine_grid	✔️	✔️
F.alpha_dropout	✔️	✔️
F.avg_pool1d	✔️	✔️*
F.avg_pool2d	✔️	✔️*
F.avg_pool3d	✔️	✔️*
F.batch_norm	✔️	✔️
F.bilinear
F.celu	✔️
F.conv1d	✔️	✔️
F.conv2d	✔️	✔️
F.conv3d	✔️	✔️
F.conv_transpose1d	✔️
F.conv_transpose2d	✔️	✔️
F.conv_transpose3d	✔️
F.cosine_similarity
F.dropout	✔️	✔️
F.dropout2d	✔️	✔️
F.dropout3d	✔️	✔️
F.elu	✔️	✔️
F.elu_	✔️	✔️
F.embedding	✔️	✔️
F.embedding_bag
F.feature_alpha_dropout	✔️	✔️
F.fold
F.fractional_max_pool2d
F.fractional_max_pool3d
F.gelu	✔️	✔️
F.glu
F.grid_sample	✔️
F.group_norm	✔️	✔️
F.gumbel_softmax
F.hardshrink	✔️
F.hardsigmoid	✔️	✔️
F.hardswish	✔️	✔️
F.hardtanh	✔️	✔️
F.hardtanh_	✔️	✔️
F.instance_norm	✔️	✔️
F.interpolate	✔️	✔️
F.layer_norm	✔️	✔️
F.leaky_relu	✔️	✔️
F.leaky_relu_	✔️	✔️
F.linear	✔️	✔️*
F.local_response_norm	✔️	✔️
F.logsigmoid	✔️
F.log_softmax	✔️
F.lp_pool1d	✔️
F.lp_pool2d	✔️
F.max_pool1d	✔️	✔️
F.max_pool2d	✔️	✔️
F.max_pool3d	✔️	✔️
F.max_unpool1d
F.max_unpool2d
F.max_unpool3d
F.mish	✔️	✔️
F.normalize	✔️	✔️
F.one_hot
F.pad	✔️	✔️
F.pairwise_distance
F.pdist
F.pixel_shuffle	✔️	✔️
F.pixel_unshuffle	✔️	✔️
F.prelu	✔️	✔️
F.relu	✔️	✔️
F.relu_	✔️	✔️
F.relu6	✔️	✔️
F.rrelu	✔️
F.rrelu_	✔️
F.selu	✔️	✔️
F.sigmoid	✔️	✔️
F.silu	✔️	✔️
F.softmax	✔️	✔️
F.softmin	✔️
F.softplus	✔️
F.softshrink	✔️
F.softsign	✔️
F.tanh	✔️	✔️
F.tanhshrink	✔️
F.threshold	✔️
F.threshold_	✔️
F.unfold
F.upsample	✔️	✔️
F.upsample_bilinear	✔️	✔️
F.upsample_nearest	✔️	✔️