onnx-tool

简体中文

onnx-tool

A tool for ONNX model:

• Parse and edit: Constant folding; OPs fusion.
• Model profiling: Rapid shape inference; MACs statistics
• Compute Graph and Shape Engine.
• Model memory compression: activation compression and weight compression.
• Quantized models and sparse models are supported.

Supported Models:

• NLP: BERT, T5, GPT, LLaMa, MPT(TransformerModel)
• Diffusion: Stable Diffusion(TextEncoder, VAE, UNET)
• CV: BEVFormer, MobileNet, YOLO, ...
• Audio: sovits, LPCNet

---

Basic Parse and Edit

You can load any onnx file by onnx_tool.Model:
Change graph structure with onnx_tool.Graph;
Change op attributes and IO tensors with onnx_tool.Node;
Change tensor data or type with onnx_tool.Tensor.
To apply your changes, just call save_model method of onnx_tool.Model or onnx_tool.Graph.

Please refer benchmark/examples.py.

---

Shape Inference & Profile Model

All profiling data must be built on shape inference result.
ONNX graph with tensor shapes:

Regular model profiling table:

Sparse profiling table:

Introduction: data/Profile.md.
pytorch usage: data/PytorchUsage.md.
tensorflow usage: data/TensorflowUsage.md.
examples: benchmark/examples.py.

---

Compute Graph with Shape Engine

From a raw graph to a compute graph:

Remove shape calculation layers(created by ONNX export) to get a Compute Graph. Use Shape Engine to update tensor shapes at runtime.
Examples: benchmark/shape_regress.py. benchmark/examples.py.
Integrate Compute Graph and Shape Engine into a cpp inference engine: data/inference_engine.md

---

Memory Compression

Activation Compression

Activation memory also called temporary memory is created by each OP's output. Only the last activation marked as the model's output will be kept. So you don't have to prepare memory space for each activation tensor. They better reuse an optimized memory size.

For large language models and high-resolution CV models, the activation memory compression is a key to save memory.
The compression method achieves 5% memory compression on most models.
For example:

model	Native Memory Size(MB)	Compressed Memory Size(MB)	Compression Ratio(%)
StableDiffusion(VAE_encoder)	14,245	540	3.7
StableDiffusion(VAE_decoder)	25,417	1,140	4.48
StableDiffusion(Text_encoder)	215	5	2.5
StableDiffusion(UNet)	36,135	2,232	6.2
GPT2	40	2	6.9
BERT	2,170	27	1.25

code example: benchmark/compression.py

Weight Compression

A fp32 model with 7B parameters will take 28GB disk space and memory space. You can not even run the model if your device doesn't have that much memory space. So weight compression is critical to run large language models. As a reference, 7B model with int4 symmetric per block(32) quantization(llama.cpp's q4_0 quantization method) only has ~0.156x model size compared with fp32 model.

Current support:

• [fp16]
• [int8]x[symmetric/asymmetric]x[per tensor/per channel/per block]
• [int4]x[symmetric/asymmetric]x[per tensor/per channel/per block]

code examples:benchmark/examples.py.

---

How to install

pip install onnx-tool

OR

pip install --upgrade git+https://github.com/ThanatosShinji/onnx-tool.git

python>=3.6

If pip install onnx-tool failed by onnx's installation, you may try pip install onnx==1.8.1 (a lower version like this) first.
Then pip install onnx-tool again.

---

Known Issues

• Loop op is not supported
• Sequence type is not supported

---

Results of ONNX Model Zoo and SOTA models

Some models have dynamic input shapes. The MACs varies from input shapes. The input shapes used in these results are writen to data/public/config.py. These onnx models with all tensors' shape can be downloaded: baidu drive(code: p91k) google drive

Model Params(M) MACs(M) GPT-J 1 layer 464 173,398 MPT 1 layer 261 79,894 text_encoder 123.13 6,782 UNet2DCondition 859.52 888,870 VAE_encoder 34.16 566,371 VAE_decoder 49.49 1,271,959 SqueezeNet 1.0 1.23 351 AlexNet 60.96 665 GoogleNet 6.99 1,606 googlenet_age 5.98 1,605 LResNet100E-IR 65.22 12,102 BERT-Squad 113.61 22,767 BiDAF 18.08 9.87 EfficientNet-Lite4 12.96 1,361 Emotion 12.95 877 Mask R-CNN 46.77 92,077

Model Params(M) MACs(M) LLaMa 1 layer 618 211,801 BEVFormer Tiny 33.7 210,838 rvm_mobilenetv3 3.73 4,289 yolov4 64.33 3,319 ConvNeXt-L 229.79 34,872 edgenext_small 5.58 1,357 SSD 19.98 216,598 RealESRGAN 16.69 73,551 ShuffleNet 2.29 146 GPT-2 137.02 1,103 T5-encoder 109.62 686 T5-decoder 162.62 1,113 RoBERTa-BASE 124.64 688 Faster R-CNN 44.10 46,018 FCN ResNet-50 35.29 37,056 ResNet50 25 3,868