Image Classification Based on Resnet18
1. Configure Docker development environment
Refer to here. After configuring the Docker development environment, return here to continue the next step.
2. Prepare the working directory in Docker
Create and enter the resnet18
working directory, note that it is a directory at the same level as tpu-mlir
.
# mkdir resnet18 && cd resnet18
Get the original model
# wget https://github.com/onnx/models/raw/main/vision/classification/resnet/model/resnet18-v1-7.tar.gz
Extract resnet18-v1-7.tar.gz
# tar -zxvf resnet18-v1-7.tar.gz
After extraction is complete, the resnet18-v1-7
folder will be generated in the current directory, which contains the resnet18-v1-7.onnx
model file.
Copy test image:
# cp -rf ${TPUC_ROOT}/regression/dataset/ILSVRC2012/ .
# cp -rf ${TPUC_ROOT}/regression/image/ .
${TPUC_ROOT}
here is an environment variable, corresponding to the tpu-mlir
directory, which is loaded in the source ./tpu-mlir/envsetup.sh
step in the previous configuration of the Docker development environment.
Create and enter the work
working directory to store compiled files such as MLIR
and cvimodel
# mkdir work && cd work
3. ONNX Model Conversion
The Duo development board is equipped with the CV1800B chip, which supports the ONNX series and Caffe models. Currently, it does not support TFLite models. In terms of quantized data types, it supports quantization in BF16 format and asymmetric quantization in INT8 format.
The steps for model conversion are as follows:
- Convert ONNX model to MLIR
- Generate calibration tables required for quantification
- MLIR quantization into INT8 asymmetric cvimodel
ONNX model converted to MLIR
The model in this example is RGB input, mean
and scale
are 123.675
,116.28
,103.53
and 0.0171
,0.0175
,0.0174
respectively.
The command to convert an ONNX model to an MLIR model is as follows:
model_transform.py \
--model_name resnet18 \
--model_def ../resnet18-v1-7/resnet18-v1-7.onnx \
--test_input ../image/cat.jpg \
--input_shapes [[1,3,224,224]] \
--resize_dims 256,256 \
--mean 123.675,116.28,103.53 \
--scale 0.0171,0.0175,0.0174 \
--pixel_format rgb \
--test_result resnet18_top_outputs.npz \
--mlir resnet18.mlir
Example of successful operation
After converting to the MLIR model, a resnet18.mlir
file will be generated, which is the MLIR model file. A resnet18_in_f32.npz
file and a resnet18_top_outputs.npz
file will also be generated, which are the input files for subsequent model conversion.