intelliw-onnx 0.0.4

An easy to start Intelligent Workshop Algorithm Framework

Intelligent Workshop Algorithm Framework ONNX

intelliw-onnx
This tool is an ONNX tool that provides various functions, including model conversion, model quantization, model pre- and post-processing.
It can be used either through the command line or by calling it in code.

Base

Images:

# cpu onnxruntime 1.15.1
dockerhub.yonyoucloud.com/c87e2267-1001-4c70-bb2a-ab41f3b81aa3/intelliw/onnx-base:py38-1.0

# gpu cuda 11.6  onnxruntime 1.15.1
dockerhub.yonyoucloud.com/c87e2267-1001-4c70-bb2a-ab41f3b81aa3/intelliw/onnx-cuda116:py38-1.0

how to use images

docker run -p {external_port}:{internal_port} -v {external_dir_path}:{images_internal_dir_path} {image} bash
#  -p, --publish list                   Publish a container's port(s) to the host
#  -v, --volume list                    Bind mount a volume
#  e.g. docker run -p 18080:80 -v C://User//xxx//package:/home/xxx/package dockerhub.yonyoucloud.com/c87e2267-1001-4c70-bb2a-ab41f3b81aa3/intelliw/onnx-base:py38-1.0 bash

Installing

Install and update using pip:

pip install -U intelliw-onnx

Argument

model_path: [required]  Input path(model file or folder)
model_type: [required]  Input model type(ex: paddle/pytorch)
output:     [required]  Output path(ex: ./output.onnx)

op_set:             Set op_set version(default: 11)
input_shape:        [pytorch/paddle required]  Input shape for pytorch/paddle(ex: [1,3,224,224] or [1,3,224,224]/[1,3,56,56])
model_def_file:      [pytorch/paddle required]  Paddle/pytorch model definition file location(ex: --model_def_file ./cnn.py)
model_class_name:   [pytorch/paddle required]  Paddle/pytorch model class name(ex: --model_class_name CNN)
model_weights_file:  Paddle/pytorch model weights file location(ex: --model_weights_file ./0.99667.pth)
model_input_type:   Paddle/pytorch input type(default float, choice is ['float', 'float32', 'float16', 'uint8', 'int8', 'uint16', 'int16', 'uint32', 'int32', 'uint64', 'int64', 'bool'])
params_file:         Paddle/pytorch params declaration file location(ex: --params_file ./params.py)
output_num:         If output num of pytorch model > 1, you can specify it by --output_num
keep_batch:         For pytorch, if set 1, the tool will keep model batch size(if 0, set it to dynamic(-1))"
dynamic_batch:      If set 1, the tool will convert batch size to -1
simplify:           Simplify the model(0:no simplify;1:do simplify; 2:for dynamic model)
simplify_hw:        When h/w is -1, you can specify h/w as you expected(together with --simplify 2)
force_simplify:     Force simplify the model(0:no simplify;1:do simplify; 2:for dynamic model)

params_file means params.py

# params.py
param_dict = {
    "n": 3, 
    "your_params":"your_value"
    ...
}

Pytorch to ONNX

Code

from intelliw_onnx.convert import ONNXConvert, ConvertArgs

if __name__ == '__main__':
    args = ConvertArgs(model_path='./model.pth',
                       model_type='pytorch',
                       output='./test_tf_model.onnx',
                       input_shape='[1,3,10,10]',
                       model_def_file="./test_onnx_demo.py",
                       model_class_name="MyCNNModel",
                       params_file="./params.py")
    converter = ONNXConvert(args)
    converter.convert()

CMD

intelliw-onnx convert --model_path ./model.pth --model_type pytorch --output "./output.onnx" --input_shape '[1,3,10,10]' 
  --model_def_file './test_onnx_demo.py' 
  --model_class_name 'MyModel'  
  --params_file ./params.py

Other Mode

Use model_weights_file:

Contains only weight parameters, model_path can be arbitrarily specified, and will not be used

intelliw-onnx convert --model_path ./xxx --model_type pytorch --output ./output.onnx 
  --model_def_file ./unet.py
  --model_class_name Net
  --model_weights_file ./9_epoch_iou_0.9743422508239746.pth 
  --input_shape [64,3,32,32]  

Contains only weight parameters, using classes in torchvision, the model_def_file parameter is not required , model_path can be arbitrarily specified, and will not be used

intelliw-onnx convert --model_path ./xxx --model_type pytorch --output ./output.onnx 
  --model_class_name torchvision.models.resnet50 
  --model_weights_file ./0.96966957919051920.9139525532770406.pth 
  --input_shape [16,3,256,256]

Use multi input/output:

intelliw-onnx convert --model_path ./model.pth --model_type pytorch --output ./output.onnx 
  --model_def_file ./pt_multi_input.py 
  --model_class_name nettest 
  --input_shape [1,3,200,300]/[1,3,200,300] 
  --output_num 2 

or

intelliw-onnx convert --model_path ./model.pth --model_type pytorch --output ./output.onnx 
  --model_def_file ./pt_multi_input.py 
  --model_class_name nettest 
  --model_weights_file ./multi_input_state.pth 
  --input_shape [1,3,500,600]/[1,3,500,600] 
  --output_num 2

Paddle to ONNX

Cde

1 dynamic paddle model

from intelliw_onnx.convert import ONNXConvert, ConvertArgs

if __name__ == '__main__':
    args = ConvertArgs(model_path='./xxx',
                       model_type='paddle',
                       output='./paddle.onnx',
                       input_shape='[1,1,28,28] ',
                       model_def_file="./mnist.py ",
                       model_class_name="LeNet",
                       model_weights_file="./paddle_checkpoint/final.pdparams")
    converter = ONNXConvert(args)
    converter.convert()

2 static paddle model

from intelliw_onnx.convert import ONNXConvert, ConvertArgs

if __name__ == '__main__':
    args = ConvertArgs(model_path='./xxx/',
                       model_type='paddle',
                       output='./paddle.onnx')
    converter = ONNXConvert(args)
    converter.convert()

CMD

1 dynamic paddle model

intelliw-onnx --model_path ./xxx/ --model_type paddle 
  --output ./paddle.onnx --model_def_file ./mnist.py 
  --model_class_name LeNet --model_weights_file ./paddle_checkpoint/final.pdparams 
  --input_shape [1,1,28,28] 

or

intelliw-onnx --model_path ./xxx/ --model_type paddle 
  --output ./paddle.onnx --model_class_name paddle.vision.models.LeNet 
  --model_weights_file ./paddle_checkpoint/final.pdparams --input_shape [1,1,28,28] 

2 static paddle model

intelliw-onnx --model_path ./paddle_model/ 
  --model_type paddle --output ./paddle.onnx