TPU-MLIR

TPU-MLIR is an open-source machine-learning compiler based on MLIR for TPU. This project provides a complete toolchain, which can convert pre-trained neural networks from different frameworks into binary files bmodel that can be efficiently operated on TPUs.

TPU-MLIR is originally developed by SOPHGO. This company is committed to becoming the world's leading general computing power provider. SOPHGO has inherited the technology, patents, products and customers of BITMAIN in the AI field for many years and focused on the R&D, promotion and application of artificial intelligence chips and related products. It has two major brands: SOPHON and CVITEK. With the self-developed chips as the core, SOPHGO has created a matrix of computing power products, which covers the whole scene of "cloud, edge and terminal" and provides computing power products and overall solutions for urban brains, intelligent computing centers, intelligent security, intelligent transportation, safety production, industrial quality inspection, intelligent terminals and others.

For technical details of this project, please refer to: TPU-MLIR Technical Reference Manual.

Currently, the project supports BM1684x. BM1684, CV183x, CV182x, Mars, and other chips will be supported in the future.

How to Build

After cloning the code of this project, it needs to be compiled in docker.

Download the required image from dockerhub.

docker pull sophgo/tpuc_dev:latest

# myname1234 just a example, you can set your own name
docker run --privileged --name myname1234 -v $PWD:/workspace -it sophgo/tpuc_dev:latest

After the container is created, the directory of the code in docker should be /workspace/tpu-mlir.

Building

Run the following command in the project directory:

cd tpu-mlir
source ./envsetup.sh
./build.sh

How to Test

# This project contains the yolov5s.onnx model, which can be used directly for verification
pushd regression
./run_model.sh yolov5s
popd

Options:

If you want to verify more networks, you need to clone them first. Please refer to https://github.com/sophgo/model-zoo.

After cloning, the model path should be /workspace/model-zoo, and then use the following command to verify:

# This step can also be skipped since its execution time is quite long
pushd regression
./run_all.sh
popd

Usage

Introduce the usage of TPU-MLIR by a simple example of compiling yolov5s.onnx and running it on the BM1684x TPU platform.

The model comes from the official website of yolov5: https://github.com/ultralytics/yolov5/releases/download/v6.0/yolov5s.onnx.

It has been placed in project path model/yolov5s.onnx.

Preparation

Firstly, create a model_yolov5s directory at the same level directory with this project. Then put both model and image files into it.

The operation is as follows:

mkdir model_yolov5s && cd model_yolov5s
cp ${REGRESSION_PATH}/model/yolov5s.onnx .
cp -rf ${REGRESSION_PATH}/dataset/COCO2017 .
cp -rf ${REGRESSION_PATH}/image .
mkdir workspace && cd workspace

Model to MLIR

If the model takes images as input, we need to learn its preprocessing before transforming. No preprocessing needs to be considered if the input is npz file. The preprocessing process is formulated as follows:

$$ y = （x - mean） \times scale, $$

where x represents the input.

The input of the official yolov5 is RGB image. Each value will be multiplied by 1/255. Mean and scale are 0.0, 0.0, 0.0 and 0.0039216, 0.0039216, 0.0039216 respectively.

The model conversion command:

model_transform.py \
    --model_name yolov5s \
    --model_def ../yolov5s.onnx \
    --input_shapes [[1,3,640,640]] \
    --mean 0.0,0.0,0.0 \
    --scale 0.0039216,0.0039216,0.0039216 \
    --keep_aspect_ratio \
    --pixel_format rgb \
    --output_names 350,498,646 \
    --test_input ../image/dog.jpg \
    --test_result yolov5s_top_outputs.npz \
    --mlir yolov5s.mlir

The arguments of model_transform.py:

Argument	Required？	Description
model_name	Yes	Model name
model_def	Yes	Model definition file (e.g., `.onnx`, `.tflite` or `.prototxt` files)
model_data	No	Specify the model weight file, required when it is caffe model (corresponding to the '.caffemodel' file)
input_shapes	No	The shape of the input, such as [[1,3,640,640]] (a two-dimensional array), which can support multiple inputs
resize_dims	No	The size of the original image to be adjusted to. If not specified, it will be resized to the input size of the model
keep_aspect_ratio	No	Whether to maintain the aspect ratio when resize. False by default. It will pad 0 to the insufficient part when setting
mean	No	The mean of each channel of the image. The default is 0.0,0.0,0.0
scale	No	The scale of each channel of the image. The default is 1.0,1.0,1.0
pixel_format	No	Image type, can be rgb, bgr, gray or rgbd
output_names	No	The names of the output. Use the output of the model if not specified, otherwise use the specified names as the output
test_input	No	The input file for validation, which can be an image, npy or npz. No validation will be carried out if it is not specified
test_result	No	Output file to save validation result
excepts	No	Names of network layers that need to be excluded from validation. Separated by comma
mlir	Yes	The output mlir file name (including path)

After converting to mlir file, a ${model_name}_in_f32.npz file containing preprocessed input will be generated.

MLIR to F32 bmodel

Convert the mlir file to the F32 bmodel by the following command:

model_deploy.py \
  --mlir yolov5s.mlir \
  --quantize F32 \
  --chip bm1684x \
  --test_input yolov5s_in_f32.npz \
  --test_reference yolov5s_top_outputs.npz \
  --model yolov5s_1684x_f32.bmodel

The arguments of model_deploy.py:

Argument	Required？	Description
mlir	Yes	Mlir file
quantize	Yes	Quantization type (F32/F16/BF16/INT8)
chip	Yes	The platform that the model will use. Currently only bm1684x is supported. More TPU platforms will be supported in the future
calibration_table	No	The quantization table path. Required when it is INT8 quantization
tolerance	No	Tolerance for the minimum similarity between MLIR quantized and MLIR fp32 inference results
correctnetss	No	Tolerance for the minimum similarity between simulator and MLIR quantized inference results. 0.99,0.90 by default
excepts	No	Names of network layers that need to be excluded from validation. Separated by comma
model	Yes	Name of output model file (including path)

MLIR to INT8 bmodel

Before converting to the INT8 model, you need to run calibration to get the calibration table. The number of input data is about 100 to 1000 according to the situation.

Then use the calibration table to generate a symmetric or asymmetric bmodel. It is generally not recommended to use the asymmetric one if the symmetric one already meets the requirements, because the performance of the asymmetric model will be slightly worse than the symmetric model.

Here is an example of the existing 100 images from COCO2017 to perform calibration:

run_calibration.py yolov5s.mlir \
  --dataset ../COCO2017 \
  --input_num 100 \
  -o yolov5s_cali_table

Execute the following command to convert to the INT8 symmetric quantized model:

model_deploy.py \
  --mlir yolov5s.mlir \
  --quantize INT8 \
  --calibration_table yolov5s_cali_table \
  --chip bm1684x \
  --test_input yolov5s_in_f32.npz \
  --test_reference yolov5s_top_outputs.npz \
  --tolerance 0.85,0.45 \
  --model yolov5s_1684x_int8_sym.bmodel

To the INT8 asymmetric quantized model:

model_deploy.py \
  --mlir yolov5s.mlir \
  --quantize INT8 \
  --asymmetric \
  --calibration_table yolov5s_cali_table \
  --chip bm1684x \
  --test_input yolov5s_in_f32.npz \
  --test_reference yolov5s_top_outputs.npz \
  --tolerance 0.90,0.55 \
  --model yolov5s_1684x_int8_asym.bmodel

Results Comparison

This project has a yolov5 sample written in python (path: python/samples/detect_yolov5.py) for object detection. Read the code to learn how the model is used:

preprocess the input
model inference to get output
post-process the output

The following code is used to verify the output of onnx/f32/int8 model respectively:

ONNX model:

detect_yolov5.py \
  --input ../image/dog.jpg \
  --model ../yolov5s.onnx \
  --output dog_onnx.jpg

F32 bmodel:

detect_yolov5.py \
  --input ../image/dog.jpg \
  --model yolov5s_1684x_f32.bmodel \
  --output dog_f32.jpg

INT8 symmetric quantized bmodel:

detect_yolov5.py \
  --input ../image/dog.jpg \
  --model yolov5s_1684x_int8_sym.bmodel \
  --output dog_int8_sym.jpg

INT8 asymmetric quantized bmodel:

detect_yolov5.py \
  --input ../image/dog.jpg \
  --model yolov5s_1684x_int8_asym.bmodel \
  --output dog_int8_asym.jpg

Outputs of four different models are compared below:

Auxiliary Tools

Model Inference Tool `model_runner.py`

Supports bmodel/mlir/onnx/tflite.

model_runner.py \
  --input resnet18_in_f32.npz \
  --model resnet18_1684x_f32.bmodel \
  --output resnet18_output.npz

Tool for `bmodel`

The bmodel file can be viewed and edited by model_tool:

  model_tool
    --info model_file : show brief model info
    --print model_file : show detailed model info
    --extract model_file : extract one multi-net bmodel to multi one-net bmodels
    --combine file1 .. fileN -o new_file: combine bmodels to one bmodel by filepath
    --combine_dir dir1 .. dirN -o new_dir: combine bmodels to one bmodel by directory path
    --dump model_file start_offset byte_size out_file: dump binary data to file from bmodel

For example, to get basic information of bmodel:

model_tool --info resnet18_1684x_f32.bmodel

eat4toast/tpu-mlir