Home

Awesome

openvino2tensorflow

For those who lack skills in converting from ONNX to TensorFlow, I recommend using this tool. It is a tool in the making, so there are lots of bugs, but it is much easier than going through OpenVINO.

"Self-Created Tools to convert ONNX files (NCHW) to TensorFlow format (NHWC). The purpose of this tool is to solve the massive Transpose extrapolation problem in onnx-tensorflow (onnx-tf)."

https://github.com/PINTO0309/onnx2tf

<p align="center"> <img src="https://user-images.githubusercontent.com/33194443/194713389-62826ef1-0ea5-4b0e-acd4-540b9f5cf58f.png" /> </p>
<p align="center"> <img src="https://user-images.githubusercontent.com/33194443/104584047-4e688f80-56a5-11eb-8dc2-5816487239d0.png" /> </p>

This script converts the ONNX/OpenVINO IR model to Tensorflow's saved_model, tflite, h5, tfjs, tftrt(TensorRT), CoreML, EdgeTPU, ONNX and pb. PyTorch (NCHW) -> ONNX (NCHW) -> OpenVINO (NCHW) -> openvino2tensorflow -> Tensorflow/Keras (NHWC/NCHW) -> TFLite (NHWC/NCHW). And the conversion from .pb to saved_model and from saved_model to .pb and from .pb to .tflite and saved_model to .tflite and saved_model to onnx. Support for building environments with Docker. It is possible to directly access the host PC GUI and the camera to verify the operation. NVIDIA GPU (dGPU) support. Intel iHD GPU (iGPU) support.

Special custom TensorFlow binaries and special custom TensorFLow Lite binaries are used.

Work in progress now.

Downloads GitHub PyPI CodeQL

render1629515758354

1. Environment

↥ Back to top

2. Use case

↥ Back to top

3. Supported Layers

↥ Back to top

4. Setup

4-1. [Environment construction pattern 1] Execution by Docker (strongly recommended)

You do not need to install any packages other than Docker. It consumes 23.4GB of storage.

$ docker pull ghcr.io/pinto0309/openvino2tensorflow:latest
or
# $ mv .dockerignore d
# $ docker build \
# -t ghcr.io/pinto0309/openvino2tensorflow:base.11.7.1-cudnn8-tf2.10.0-trt8.4.3-openvino2022.1.0 \
# -f Dockerfile.base .
# $ mv d .dockerignore
$ docker build --no-cache -t ghcr.io/pinto0309/openvino2tensorflow:latest .

# If you don't need to access the GUI of the HostPC and the USB camera.
$ docker run -it --rm \
  -v `pwd`:/home/user/workdir \
  ghcr.io/pinto0309/openvino2tensorflow:latest

# If conversion to TF-TRT is not required. And if you need to access the HostPC GUI and USB camera.
$ xhost +local: && \
  docker run -it --rm \
  -v `pwd`:/home/user/workdir \
  -v /tmp/.X11-unix/:/tmp/.X11-unix:rw \
  --device /dev/video0:/dev/video0:mwr \
  --net=host \
  -e XDG_RUNTIME_DIR=$XDG_RUNTIME_DIR \
  -e DISPLAY=$DISPLAY \
  --privileged \
  ghcr.io/pinto0309/openvino2tensorflow:latest

# If you need to convert to TF-TRT. And if you need to access the HostPC GUI and USB camera.
$ xhost +local: && \
  docker run --gpus all -it --rm \
  -v `pwd`:/home/user/workdir \
  -v /tmp/.X11-unix/:/tmp/.X11-unix:rw \
  --device /dev/video0:/dev/video0:mwr \
  --net=host \
  -e XDG_RUNTIME_DIR=$XDG_RUNTIME_DIR \
  -e DISPLAY=$DISPLAY \
  --privileged \
  ghcr.io/pinto0309/openvino2tensorflow:latest

# If you are using iGPU (OpenCL). And if you need to access the HostPC GUI and USB camera.
$ xhost +local: && \
  docker run -it --rm \
  -v `pwd`:/home/user/workdir \
  -v /tmp/.X11-unix/:/tmp/.X11-unix:rw \
  --device /dev/video0:/dev/video0:mwr \
  --net=host \
  -e LIBVA_DRIVER_NAME=iHD \
  -e XDG_RUNTIME_DIR=$XDG_RUNTIME_DIR \
  -e DISPLAY=$DISPLAY \
  --privileged \
  ghcr.io/pinto0309/openvino2tensorflow:latest

↥ Back to top

4-2. [Environment construction pattern 2] Execution by Host machine

To install using the Python Package Index (PyPI), use the following command.

$ pip3 install --user --upgrade openvino2tensorflow

To install with the latest source code of the main branch, use the following command.

$ pip3 install --user --upgrade git+https://github.com/PINTO0309/openvino2tensorflow

↥ Back to top

5. Usage

5-1. openvino to tensorflow convert

usage: openvino2tensorflow
  [-h]
  --model_path MODEL_PATH
  [--model_output_path MODEL_OUTPUT_PATH]
  [--output_saved_model]
  [--output_h5]
  [--output_weight_and_json]
  [--output_pb]
  [--output_no_quant_float32_tflite]
  [--output_dynamic_range_quant_tflite]
  [--output_weight_quant_tflite]
  [--output_float16_quant_tflite]
  [--output_integer_quant_tflite]
  [--output_full_integer_quant_tflite]
  [--output_integer_quant_type OUTPUT_INTEGER_QUANT_TYPE]
  [--string_formulas_for_normalization STRING_FORMULAS_FOR_NORMALIZATION]
  [--calib_ds_type CALIB_DS_TYPE]
  [--ds_name_for_tfds_for_calibration DS_NAME_FOR_TFDS_FOR_CALIBRATION]
  [--split_name_for_tfds_for_calibration SPLIT_NAME_FOR_TFDS_FOR_CALIBRATION]
  [--download_dest_folder_path_for_the_calib_tfds DOWNLOAD_DEST_FOLDER_PATH_FOR_THE_CALIB_TFDS]
  [--tfds_download_flg]
  [--load_dest_file_path_for_the_calib_npy LOAD_DEST_FILE_PATH_FOR_THE_CALIB_NPY]
  [--output_tfjs]
  [--output_tftrt_float32]
  [--output_tftrt_float16]
  [--tftrt_maximum_cached_engines TFTRT_MAXIMUM_CACHED_ENGINES]
  [--output_coreml]
  [--output_edgetpu]
  [--edgetpu_compiler_timeout EDGETPU_COMPILER_TIMEOUT]
  [--edgetpu_num_segments EDGETPU_NUM_SEGMENTS]
  [--output_onnx]
  [--onnx_opset ONNX_OPSET]
  [--onnx_extra_opset ONNX_EXTRA_OPSET]
  [--disable_onnx_nchw_conversion]
  [--disable_onnx_optimization]
  [--output_myriad]
  [--vpu_number_of_shaves VPU_NUMBER_OF_SHAVES]
  [--vpu_number_of_cmx_slices VPU_NUMBER_OF_CMX_SLICES]
  [--replace_swish_and_hardswish]
  [--optimizing_hardswish_for_edgetpu]
  [--replace_prelu_and_minmax]
  [--replace_argmax]
  [--replace_argmax_indices_to_float32]
  [--restricted_resize_image_mode]
  [--weight_replacement_config WEIGHT_REPLACEMENT_CONFIG]
  [--disable_experimental_new_quantizer]
  [--disable_per_channel]
  [--optimizing_barracuda]
  [--layerids_of_the_terminating_output LAYERIDS_OF_THE_TERMINATING_OUTPUT]
  [--keep_input_tensor_in_nchw]
  [--input_as_ncdhw]
  [--non_verbose]

optional arguments:
  -h, --help
              show this help message and exit
  --model_path MODEL_PATH
              input IR model path (.xml)
  --model_output_path MODEL_OUTPUT_PATH
              The output folder path of the converted model file
  --output_saved_model
              saved_model output switch
  --output_h5
              .h5 output switch
  --output_weight_and_json
              weight of h5 and json output switch
  --output_pb
              .pb output switch
  --output_no_quant_float32_tflite
              float32 tflite output switch
  --output_dynamic_range_quant_tflite
              dynamic range quant tflite output switch
  --output_weight_quant_tflite
              weight quant tflite output switch
  --output_float16_quant_tflite
              float16 quant tflite output switch
  --output_integer_quant_tflite
              integer quant tflite output switch
  --output_full_integer_quant_tflite
              full integer quant tflite output switch
  --output_integer_quant_type OUTPUT_INTEGER_QUANT_TYPE
              Input and output types when doing Integer Quantization
              ('int8 (default)' or 'uint8')
  --string_formulas_for_normalization STRING_FORMULAS_FOR_NORMALIZATION
              String formulas for normalization. It is evaluated by
              Pythons eval() function.
              Default: '(data - [127.5,127.5,127.5]) / [127.5,127.5,127.5]'
  --calib_ds_type CALIB_DS_TYPE
              Types of data sets for calibration. tfds or numpy
              Default: numpy
  --ds_name_for_tfds_for_calibration DS_NAME_FOR_TFDS_FOR_CALIBRATION
              Dataset name for TensorFlow Datasets for calibration.
              https://www.tensorflow.org/datasets/catalog/overview
  --split_name_for_tfds_for_calibration SPLIT_NAME_FOR_TFDS_FOR_CALIBRATION
              Split name for TensorFlow Datasets for calibration.
              https://www.tensorflow.org/datasets/catalog/overview
  --download_dest_folder_path_for_the_calib_tfds DOWNLOAD_DEST_FOLDER_PATH_FOR_THE_CALIB_TFDS
              Download destination folder path for the calibration
              dataset. Default: $HOME/TFDS
  --tfds_download_flg
              True to automatically download datasets from
              TensorFlow Datasets. True or False
  --load_dest_file_path_for_the_calib_npy LOAD_DEST_FILE_PATH_FOR_THE_CALIB_NPY
              The path from which to load the .npy file containing
              the numpy binary version of the calibration data.
              Default: sample_npy/calibration_data_img_sample.npy
  --output_tfjs
              tfjs model output switch
  --output_tftrt_float32
              tftrt float32 model output switch
  --output_tftrt_float16
              tftrt float16 model output switch
  --tftrt_maximum_cached_engines
              Specifies the quantity of tftrt_maximum_cached_engines for TFTRT.
              Default: 10000
  --output_coreml
              coreml model output switch
  --output_edgetpu
              edgetpu model output switch
  --edgetpu_compiler_timeout
              edgetpu_compiler timeout for one compilation process in seconds.
              Default: 3600
  --edgetpu_num_segments
              Partition the model into 'num_segments' segments.
              Default: 1 (no partition)
  --output_onnx
              onnx model output switch
  --onnx_opset ONNX_OPSET
              onnx opset version number
  --onnx_extra_opset ONNX_EXTRA_OPSET
              The name of the onnx 'extra_opset' to enable.
              Default: ''
              'com.microsoft:1' or 'ai.onnx.contrib:1' or 'ai.onnx.ml:1'
  --disable_onnx_nchw_conversion
              Disable NCHW conversion
  --disable_onnx_optimization
              Disable onnx optimization
  --output_myriad
              myriad inference engine blob output switch
  --vpu_number_of_shaves VPU_NUMBER_OF_SHAVES
              vpu number of shaves. Default: 4
  --vpu_number_of_cmx_slices VPU_NUMBER_OF_CMX_SLICES
              vpu number of cmx slices. Default: 4
  --replace_swish_and_hardswish
              Replace swish and hard-swish with each other
  --optimizing_hardswish_for_edgetpu
              Optimizing hardswish for edgetpu
  --replace_prelu_and_minmax
              Replace prelu and minimum/maximum with each other
  --replace_argmax
              Replace 'ArgMax (TopK)' with a primitive operation.
              Optimizes 'ArgMax' to EdgeTPU. If you have 'ArgMax' at the end of your model,
              use the '--replace_argmax_indices_to_float32' option together.
  --replace_argmax_indices_to_float32
              Enabling this option may allow full mapping to EdgeTPU when 'ArgMax (TopK)'
              is at the end of the model for tasks such as SemanticSegmentation.
              If you apply it to 'ArgMax (TopK)', which is located in the middle of the model,
              the model transformation is more likely to fail.
  --restricted_resize_image_mode
              Specify this if the upsampling contains OPs that are
              not scaled by integer multiples. Optimization for
              EdgeTPU will be disabled.
  --weight_replacement_config WEIGHT_REPLACEMENT_CONFIG
              Replaces the value of Const for each layer_id defined
              in json. Specify the path to the json file.
              'weight_replacement_config.json'
  --disable_experimental_new_quantizer
              Disable MLIRs new quantization feature during INT8 quantization
              in TensorFlowLite.
  --disable_per_channel
              Disable per-channel quantization for tflite.
  --optimizing_barracuda
              Generates ONNX by replacing Barracuda unsupported layers
              with standard layers. For example, GatherND.
  --layerids_of_the_terminating_output LAYERIDS_OF_THE_TERMINATING_OUTPUT
              A comma-separated list of layerIDs to be used as output layers.
              e.g. --layerids_of_the_terminating_output 100,201,560
              Default: ''
  --keep_input_tensor_in_nchw
              Does not convert the input to NHWC, but keeps the NCHW format.
              Transpose is inserted right after the input layer, and
              the model internals are handled by NHWC. Only 4D input is supported.
  --input_as_ncdhw
              Specify when the shape of INPUT is the 5D tensor of NCDHW.
              When converting to TensorFlow, the input geometry is automatically
              converted to NDHWC format.
  --non_verbose
              Do not show all the weight information of each layer in the
              conversion log.

↥ Back to top

5-2. saved_model to tflite convert

usage: saved_model_to_tflite
  [-h]
  --saved_model_dir_path SAVED_MODEL_DIR_PATH
  [--signature_def SIGNATURE_DEF]
  [--input_shapes INPUT_SHAPES]
  [--model_output_dir_path MODEL_OUTPUT_DIR_PATH]
  [--output_no_quant_float32_tflite]
  [--output_dynamic_range_quant_tflite]
  [--output_weight_quant_tflite]
  [--output_float16_quant_tflite]
  [--output_integer_quant_tflite]
  [--output_full_integer_quant_tflite]
  [--output_integer_quant_type OUTPUT_INTEGER_QUANT_TYPE]
  [--string_formulas_for_normalization STRING_FORMULAS_FOR_NORMALIZATION]
  [--calib_ds_type CALIB_DS_TYPE]
  [--ds_name_for_tfds_for_calibration DS_NAME_FOR_TFDS_FOR_CALIBRATION]
  [--split_name_for_tfds_for_calibration SPLIT_NAME_FOR_TFDS_FOR_CALIBRATION]
  [--download_dest_folder_path_for_the_calib_tfds DOWNLOAD_DEST_FOLDER_PATH_FOR_THE_CALIB_TFDS]
  [--tfds_download_flg]
  [--load_dest_file_path_for_the_calib_npy LOAD_DEST_FILE_PATH_FOR_THE_CALIB_NPY]
  [--output_tfjs]
  [--output_tftrt_float32]
  [--output_tftrt_float16]
  [--tftrt_maximum_cached_engines TFTRT_MAXIMUM_CACHED_ENGINES]
  [--output_coreml]
  [--output_edgetpu]
  [--edgetpu_compiler_timeout EDGETPU_COMPILER_TIMEOUT]
  [--edgetpu_num_segments EDGETPU_NUM_SEGMENTS]
  [--output_onnx]
  [--onnx_opset ONNX_OPSET]
  [--onnx_extra_opset ONNX_EXTRA_OPSET]
  [--disable_onnx_nchw_conversion]
  [--disable_onnx_optimization]
  [--disable_experimental_new_quantizer]
  [--disable_per_channel]

optional arguments:
  -h, --help
              show this help message and exit
  --saved_model_dir_path SAVED_MODEL_DIR_PATH
              Input saved_model dir path
  --signature_def SIGNATURE_DEF
              Specifies the signature name to load from saved_model
  --input_shapes INPUT_SHAPES
              Overwrites an undefined input dimension (None or -1).
              Specify the input shape in [n,h,w,c] format.
              For non-4D tensors, specify [a,b,c,d,e], [a,b], etc.
              A comma-separated list if there are multiple inputs.
              (e.g.) --input_shapes [1,256,256,3],[1,64,64,3],[1,2,16,16,3]
  --model_output_dir_path MODEL_OUTPUT_DIR_PATH
              The output folder path of the converted model file
  --output_no_quant_float32_tflite
              float32 tflite output switch
  --output_dynamic_range_quant_tflite
              dynamic range quant tflite output switch
  --output_weight_quant_tflite
              weight quant tflite output switch
  --output_float16_quant_tflite
              float16 quant tflite output switch
  --output_integer_quant_tflite
              integer quant tflite output switch
  --output_full_integer_quant_tflite
              full integer quant tflite output switch
  --output_integer_quant_type OUTPUT_INTEGER_QUANT_TYPE
              Input and output types when doing Integer Quantization
              ('int8 (default)' or 'uint8')
  --string_formulas_for_normalization STRING_FORMULAS_FOR_NORMALIZATION
              String formulas for normalization. It is evaluated by
              Pythons eval() function.
              Default: '(data - [127.5,127.5,127.5]) / [127.5,127.5,127.5]'
  --calib_ds_type CALIB_DS_TYPE
              Types of data sets for calibration. tfds or numpy
              Default: numpy
  --ds_name_for_tfds_for_calibration DS_NAME_FOR_TFDS_FOR_CALIBRATION
              Dataset name for TensorFlow Datasets for calibration.
              https://www.tensorflow.org/datasets/catalog/overview
  --split_name_for_tfds_for_calibration SPLIT_NAME_FOR_TFDS_FOR_CALIBRATION
              Split name for TensorFlow Datasets for calibration.
              https://www.tensorflow.org/datasets/catalog/overview
  --download_dest_folder_path_for_the_calib_tfds DOWNLOAD_DEST_FOLDER_PATH_FOR_THE_CALIB_TFDS
              Download destination folder path for the calibration
              dataset. Default: $HOME/TFDS
  --tfds_download_flg
              True to automatically download datasets from
              TensorFlow Datasets. True or False
  --load_dest_file_path_for_the_calib_npy LOAD_DEST_FILE_PATH_FOR_THE_CALIB_NPY
              The path from which to load the .npy file containing
              the numpy binary version of the calibration data.
              Default: sample_npy/calibration_data_img_sample.npy
  --output_tfjs
              tfjs model output switch
  --output_tftrt_float32
              tftrt float32 model output switch
  --output_tftrt_float16
              tftrt float16 model output switch
  --tftrt_maximum_cached_engines
              Specifies the quantity of tftrt_maximum_cached_engines for TFTRT.
              Default: 10000
  --output_coreml
              coreml model output switch
  --output_edgetpu
              edgetpu model output switch
  --edgetpu_compiler_timeout
              edgetpu_compiler timeout for one compilation process in seconds.
              Default: 3600
  --edgetpu_num_segments
              Partition the model into 'num_segments' segments.
              Default: 1 (no partition)
  --output_onnx
              onnx model output switch
  --onnx_opset ONNX_OPSET
              onnx opset version number
  --onnx_extra_opset ONNX_EXTRA_OPSET
              The name of the onnx 'extra_opset' to enable.
              Default: ''
              'com.microsoft:1' or 'ai.onnx.contrib:1' or 'ai.onnx.ml:1'
  --disable_onnx_nchw_conversion
              Disable NCHW conversion
  --disable_onnx_optimization
              Disable onnx optimization
  --disable_experimental_new_quantizer
              Disable MLIRs new quantization feature during INT8 quantization
              in TensorFlowLite.
  --disable_per_channel
              Disable per-channel quantization for tflite.

↥ Back to top

5-3. pb to saved_model convert

usage: pb_to_saved_model
  [-h]
  --pb_file_path PB_FILE_PATH
  --inputs INPUTS
  --outputs OUTPUTS
  [--model_output_path MODEL_OUTPUT_PATH]

optional arguments:
  -h, --help
              show this help message and exit
  --pb_file_path PB_FILE_PATH
              Input .pb file path (.pb)
  --inputs INPUTS
              (e.g.1) input:0,input:1,input:2
              (e.g.2) images:0,input:0,param:0
  --outputs OUTPUTS
              (e.g.1) output:0,output:1,output:2
              (e.g.2) Identity:0,Identity:1,output:0
  --model_output_path MODEL_OUTPUT_PATH
              The output folder path of the converted model file

↥ Back to top

5-4. pb to tflite convert

usage: pb_to_tflite
  [-h]
  --pb_file_path PB_FILE_PATH
  --inputs INPUTS
  --outputs OUTPUTS
  [--model_output_path MODEL_OUTPUT_PATH]

optional arguments:
  -h, --help
              show this help message and exit
  --pb_file_path PB_FILE_PATH
              Input .pb file path (.pb)
  --inputs INPUTS
              (e.g.1) input,input_1,input_2
              (e.g.2) images,input,param
  --outputs OUTPUTS
              (e.g.1) output,output_1,output_2
              (e.g.2) Identity,Identity_1,output
  --model_output_path MODEL_OUTPUT_PATH
              The output folder path of the converted model file

↥ Back to top

5-5. saved_model to pb convert

usage: saved_model_to_pb
  [-h]
  --saved_model_dir_path SAVED_MODEL_DIR_PATH
  [--model_output_dir_path MODEL_OUTPUT_DIR_PATH]
  [--signature_name SIGNATURE_NAME]

optional arguments:
  -h, --help
              show this help message and exit
  --saved_model_dir_path SAVED_MODEL_DIR_PATH
              Input saved_model dir path
  --model_output_dir_path MODEL_OUTPUT_DIR_PATH
              The output folder path of the converted model file (.pb)
  --signature_name SIGNATURE_NAME
              Signature name to be extracted from saved_model

↥ Back to top

5-6. Extraction of IR weight

usage: ir_weight_extractor
  [-h]
  -m MODEL
  -o OUTPUT_PATH

optional arguments:
  -h, --help
              show this help message and exit
  -m MODEL, --model MODEL
              input IR model path
  -o OUTPUT_PATH, --output_path OUTPUT_PATH
              weights output folder path

↥ Back to top

6. Execution sample

6-1. Conversion of OpenVINO IR to Tensorflow models

OutOfMemory may occur when converting to saved_model or h5 when the file size of the original model is large, please try the conversion to a pb file alone.

$ openvino2tensorflow \
  --model_path openvino/448x448/FP32/Resnet34_3inputs_448x448_20200609.xml \
  --output_saved_model \
  --output_pb \
  --output_weight_quant_tflite \
  --output_float16_quant_tflite \
  --output_no_quant_float32_tflite

↥ Back to top

6-2. Convert Protocol Buffer (.pb) to saved_model

This tool is useful if you want to check the internal structure of pb files, tflite files, .h5 files, coreml files and IR (.xml) files. https://lutzroeder.github.io/netron/

$ pb_to_saved_model \
  --pb_file_path model_float32.pb \
  --inputs inputs:0 \
  --outputs Identity:0

↥ Back to top

6-3. Convert Protocol Buffer (.pb) to tflite

$ pb_to_tflite \
  --pb_file_path model_float32.pb \
  --inputs inputs \
  --outputs Identity,Identity_1,Identity_2

↥ Back to top

6-4. Convert saved_model to Protocol Buffer (.pb)

$ saved_model_to_pb \
  --saved_model_dir_path saved_model \
  --model_output_dir_path pb_from_saved_model \
  --signature_name serving_default

↥ Back to top

6-5. Converts saved_model to OpenVINO IR

$ python3 ${INTEL_OPENVINO_DIR}/deployment_tools/model_optimizer/mo_tf.py \
  --saved_model_dir saved_model \
  --output_dir openvino/reverse

↥ Back to top

6-6. Checking the structure of saved_model

$ saved_model_cli show \
  --dir saved_model \
  --tag_set serve \
  --signature_def serving_default

↥ Back to top

6-7. Replace weights or constant values in Const OP, and add Transpose or Reshape or Cast or Squeeze or Unsqueeze or Add or Multiply just before/after the operation specified by layer_id

6-7-1. Overview

If the transformation behavior of Reshape, Transpose, etc. does not go as expected, you can force the Const content to change by defining weights and constant values in a JSON file and having it read in. Alternatively, Transpose or Reshape or Cast or Squeeze or Unsqueeze or Add or Multiply can be added just before the operation specified by layer_id. After changing the structure, you need to carefully check the consistency of Reshape, Transpose and Interpolate before and after. Even if the model is successfully transformed, there is a possibility that the dimension that should be changed is transformed incorrectly. In particular, Reshape and Interpolate are often able to transform the model even if the number of elements in the dimension is messed up.

$ openvino2tensorflow \
  --model_path xxx.xml \
  --output_saved_model \
  --output_pb \
  --output_weight_quant_tflite \
  --output_float16_quant_tflite \
  --output_no_quant_float32_tflite \
  --weight_replacement_config weight_replacement_config_sample.json

Structure of JSON sample

{
    "format_version": 2,
    "layers": [
        {
            "layer_id": "659",
            "type": "Const",
            "replace_mode": "direct",
            "values": [
                0,
                1,
                2
            ]
        },
        {
            "layer_id": "660",
            "type": "Reshape",
            "replace_mode": "insert_after",
            "values": [
                2100,
                85
            ]
        },
        {
            "layer_id": "680",
            "type": "Cast",
            "replace_mode": "insert_after",
            "values": "i64"
        },
        {
            "layer_id": "442",
            "type": "Concat",
            "replace_mode": "change_axis",
            "values": 4
        },
        {
            "layer_id": "450",
            "type": "SoftMax",
            "replace_mode": "change_axis",
            "values": 2
        },
        {
            "layer_id": "500",
            "type": "StridedSlice",
            "replace_mode": "change_attributes",
            "values": [
                0,
                0,
                0,
                0,
                0
            ]
        },
        {
            "layer_id": "550",
            "type": "StridedSlice",
            "replace_mode": "replace",
            "values": [
                [0,0,0,8],
                [2,7,11,16],
                [1,1,1,1],
                0,
                0,
                0,
                0,
                0
            ]
        },
        {
            "layer_id": "600",
            "type": "MaxPool",
            "replace_mode": "change_padding_mode",
            "values": "REFLECT"
        },
        {
            "layer_id": "720",
            "type": "PReLU",
            "replace_mode": "change_shared_axes",
            "values": [
                1,
                2
            ]
        },
        {
            "layer_id": "800",
            "type": "ReverseSequence",
            "replace_mode": "change_seq_axis",
            "values": 2
        },
        {
            "layer_id": "850",
            "type": "Squeeze",
            "replace_mode": "insert_after",
            "values": 1
        },
        {
            "layer_id": "900",
            "type": "Unsqueeze",
            "replace_mode": "insert_before",
            "values": 2
        },
        {
            "layer_id": "1000",
            "type": "Einsum",
            "replace_mode": "change_equation",
            "values": "vu,nctu->nctv"
        },
        {
            "layer_id": "1005",
            "type": "Add",
            "replace_mode": "insert_after",
            "values": [
                0,
                0,
                0,
                2
            ]
        },
        {
            "layer_id": "1010",
            "type": "Multiply",
            "replace_mode": "insert_after",
            "values": [
                1.0,
                1.0,
                -0.5,
                1.0
            ]
        }
    ]
}
No.ElementsDescription
1format_versionFormat version of weight_replacement_config. Values less than or equal to 2.
2layersA list of layers. Enclose it with "[ ]" to define multiple layers to child elements.
2-1layer_idID of the Const layer whose weight/constant parameter is to be swapped. The important thing to note is that you cannot create multiple settings for a single layer_id. There should always be a single setting for a single layer_id. For example, specify "1123" for layer id="1123" for type="Const" in .xml.<br>Screenshot 2021-02-08 01:06:30
2-2typeFixed value replacement or type of operation to be added. "Const" or "Transpose" or "Reshape" or "Cast" or "Concat" or "SoftMax" or "StridedSlice" or "MaxPool" or "PReLU" or "ReverseSequence" or "Squeeze" or "Unsqueeze" or "LogSoftmax" or "Einsum" or "Add" or "Multiply"
2-3replace_mode"direct" or "npy" or "insert_before" or "insert_after" or "change_axis" or "change_attributes".<br>"direct": Specify the values of the Numpy matrix directly in the "values" attribute. Ignores the values recorded in the .bin file and replaces them with the values specified in "values".<br>Screenshot 2021-08-10 23:16:05<br>"npy": Load a Numpy binary file with the matrix output by np.save('xyz', a). The "values" attribute specifies the path to the Numpy binary file.<br>Screenshot 2021-08-10 23:17:22<br>"insert_before": Add Transpose or Reshape or Cast or Squeeze or Unsqueeze or Add or Multiply just before the operation specified by layer_id. Note that when Squeeze and Unsqueeze are specified, the value to set for "values" is the axis of the dimension operation target.<br>Screenshot 2021-09-16 14:17:22<br>"insert_after": Add Transpose or Reshape or Cast or Squeeze or Unsqueeze or Add or Multiply just after the operation specified by layer_id. Note that when Squeeze and Unsqueeze are specified, the value to set for "values" is the axis of the dimension operation target.<br>Screenshot 2021-08-10 23:12:52<br>"change_axis": Changes the axis of the Concat or SoftMax or ShuffleChannels or LogSoftmax attribute value.<br>Screenshot 2021-10-17 01:16:22<br>"change_attributes": Changes the ATTRIBUTES of the StridedSlice attribute value. Specify five values in numerical list format in the order of begin_mask, end_mask, ellipsis_mask, new_axis_mask, shrink_axis_mask.<br>Screenshot 2021-11-19 11:54:27<br>"replace": Replaces OP by specifying parameters directly in TensorFlow Strided_Slice specification. begin, end, strides, begin_mask, end_mask, ellipsis_mask, new_axis_mask, shrink_axis_mask https://www.tensorflow.org/api_docs/python/tf/strided_slice<br>image<br>"change_padding_mode": Change the padding mode of MaxPool.<br>Screenshot 2021-12-04 01:37:53<br>"change_shared_axes": Changed shared_axes in PReLU.<br>Screenshot 2021-12-04 22:22:31<br>"change_batch_axis","change_seq_axis": Changed axis in ReverseSequence.<br>Screenshot 2021-12-12 13:43:30<br>"change_equation": Changed equation in Einsum.<br>20220131183853
2-4valuesSpecify the value or the path to the Numpy binary file to replace the weight/constant value recorded in .bin. The way to specify is as described in the description of 'replace_mode'. The table below shows the correspondence between the strings that can be specified for the "Cast" operation and the TensorFlow types. In most cases, you will probably only use "i32", "i64", "f32", and "f16".<br>Screenshot 2021-08-22 01:48:43<br>change_padding_mode: "ZERO" or "SYMMETRIC" or "REFLECT". https://www.tensorflow.org/api_docs/python/tf/pad<br>change_shared_axes: https://www.tensorflow.org/api_docs/python/tf/keras/layers/PReLU<br>change_batch_axis, change_seq_axis: https://docs.openvino.ai/2021.4/openvino_docs_ops_movement_ReverseSequence_1.html<br>"change_equation": https://numpy.org/doc/stable/reference/generated/numpy.einsum.html

↥ Back to top

6-7-2. Example

  1. Let's assume that you don't need Transpose in the final layer of the model. Here you have [1, 85, 2100] as input, and the original OpenVINO model transposes [0, 2, 1] in that order to obtain the tensor [1, 2100, 85]. This figure shows the visualization of a yolox_nano_320x320.xml file using Netron. The number shown in the OUTPUTS - output - name: is the layer ID of Transpose. The layer ID 660 is the number in the part before the colon. The number in the part after the colon is called the port number 2. However, what you are trying to change is the transposition parameter of the INPUTS - custom - name: part. The name of the parameter you are trying to change is 625. Note that 625 is not a layer ID, just a name. Screenshot 2021-08-04 23:45:15
  2. Check the model structure as recorded in .xml. First, open yolox_nano_320x320.xml in your favorite IDE. Screenshot 2021-08-05 00:00:38 Screenshot 2021-08-05 00:08:50
  3. Search for to-layer="660" (Transpose) in the IDE. In the figure below, Layer ID 658 and Layer ID 659 are represented as input values connected to Layer ID 660. Screenshot 2021-08-05 00:17:31

In the figure below, one of them is 658 and one of them is 659. It is difficult to determine exactly what it is from the image alone. You must again note that 658:3 in the image is only a name, not a layer ID. It is worth noting here that the type of value you want to replace is Const.

Screenshot 2021-08-05 00:26:29

  1. Now you will search for layer ID "658" in the IDE. The type is "Concat", so the desired layer was not this one. What you are looking for is "Const". Screenshot 2021-08-05 01:02:00
  2. Now, search for layer ID 659 in the IDE. The type is "Const". Now you can finally identify that the layer ID of the layer you want to replace is 659. Screenshot 2021-08-05 01:05:33
  3. Create a JSON file to replace the constants [0, 2, 1] with [0, 1, 2], and you can use any name for the JSON file. Suppose you save the file with the name replace.json. If you want to replace it with a numpy matrix, specify "npy" for "replace_mode": and the path to the .npy file for "values":.
{
  "format_version": 2,
  "layers": [
      {
          "layer_id": "659",
          "type": "Const",
          "replace_mode": "direct",
          "values": [
              0,
              1,
              2
          ]
      }
  ]
}

{
  "format_version": 2,
  "layers": [
      {
          "layer_id": "659",
          "type": "Const",
          "replace_mode": "npy",
          "values": "path/to/your/xxx.npy"
      }
  ]
}
  1. Specify the created JSON file as the argument of the --weight_replacement_config parameter of the conversion command and execute it. This is the end of the explanation of how to replace weights and constants.
$ openvino2tensorflow \
--model_path yolox_nano_320x320.xml \
--output_saved_model \
--output_pb \
--output_no_quant_float32_tflite \
--weight_replacement_config replace.json

↥ Back to top

6-8. Check the contents of the .npy file, which is a binary version of the image file

$ view_npy --npy_file_path sample_npy/calibration_data_img_sample.npy

Press the Q button to display the next image. calibration_data_img_sample.npy contains 20 images extracted from the MS-COCO data set.

aaa

↥ Back to top

6-9. Sample image of a conversion error message

Since it is very difficult to mechanically predict the correct behavior of Transpose and Reshape, errors like the one shown below may occur. Using the information in the figure below, try several times to force the replacement of constants and weights using the --weight_replacement_config option #6-7-replace-weights-or-constant-values-in-const-op-and-add-transpose-or-reshape-or-cast-or-squeeze-or-unsqueeze-or-add-or-multiply-just-beforeafter-the-operation-specified-by-layer_id. This is a very patient process, but if you take the time, you should be able to convert it correctly. error_sample2

↥ Back to top

6-10. Ability to specify an output layer for debugging the output values of the model

If you want to debug the output values of each layer, specify multiple layer IDs separated by commas in the --layerids_of_the_terminating_output option to check the output values. For example, if you want to debug the output values of two layers, LayerID=1007 (Add) and LayerID=1214 (Sigmoid), as shown in the figure below, specify as --layerids_of_the_terminating_output 1007,1214. Screenshot 2021-09-06 21:33:19 Screenshot 2021-09-06 21:33:28 When you convert a model, the output will be censored at the two specified layer IDs, and the model will be generated with the output of the model available for review. Note that if you specify a layer ID for an operation that has multiple outputs, such as Split, VariadicSplit, TopK, or NonMaxSuppression, all output values will be used as outputs. Screenshot 2021-09-06 21:43:17

↥ Back to top

7. Output sample

Screenshot 2020-10-16 00:08:40

↥ Back to top

8. Model Structure

https://digital-standard.com/threedpose/models/Resnet34_3inputs_448x448_20200609.onnx

ONNX (NCHW)OpenVINO (NCHW)TFLite (NHWC)
Resnet34_3inputs_448x448_20200609 onnx_Resnet34_3inputs_448x448_20200609 xmlmodel_float32 tflite

↥ Back to top

9. My article

↥ Back to top

10. Conversion Confirmed Models

  1. u-2-net
  2. mobilenet-v2-pytorch
  3. midasnet
  4. footprints
  5. efficientnet-b0-pytorch
  6. efficientdet-d0
  7. dense_depth
  8. deeplabv3
  9. colorization-v2-norebal
  10. age-gender-recognition-retail-0013
  11. resnet
  12. arcface
  13. emotion-ferplus
  14. mosaic
  15. retinanet
  16. shufflenet-v2
  17. squeezenet
  18. version-RFB-320
  19. yolov4
  20. yolov4x-mish
  21. ThreeDPoseUnityBarracuda - Resnet34_3inputs_448x448
  22. efficientnet-lite4
  23. nanodet
  24. yolov4-tiny
  25. yolov5s
  26. yolact
  27. MiDaS v2
  28. MODNet
  29. Person Reidentification
  30. DeepSort
  31. DINO (Transformer)

↥ Back to top