Converting TensorFlow YOLO Models

Danger

The code described here has been deprecated! Do not use it to avoid working with a legacy solution. It will be kept for some time to ensure backwards compatibility, but you should not use it in contemporary applications.

This guide describes a deprecated conversion method. The guide on the new and recommended method can be found in the Python tutorials.

This document explains how to convert real-time object detection YOLOv1, YOLOv2, YOLOv3 and YOLOv4 public models to the Intermediate Representation (IR). All YOLO models are originally implemented in the DarkNet framework and consist of two files:

  • The .cfg file with model configurations

  • The .weights file with model weights

Depending on a YOLO model version, the convert_model() method converts it differently:

  • YOLOv4 must be first converted from Keras to TensorFlow 2.

  • YOLOv3 has several implementations. This tutorial uses a TensorFlow implementation of YOLOv3 model, which can be directly converted to an IR.

  • YOLOv1 and YOLOv2 models must be first converted to TensorFlow using DarkFlow.

Converting a YOLOv4 Model to IR

This section explains how to convert the YOLOv4 Keras model from the repository to an IR. To convert the YOLOv4 model, follow the instructions below:

  1. Download YOLOv4 weights and associated with it cfg file:

  2. Clone the repository with the YOLOv4 model:

    git clone https://github.com/david8862/keras-YOLOv3-model-set
    
  3. Convert the model to the TensorFlow 2 format:

    • for YOLOv4:

      python keras-YOLOv3-model-set/tools/model_converter/convert.py <path_to_cfg_file>/yolov4.cfg <path_to_weights>/yolov4.weights <saved_model_dir>
      
    • for YOLOv4-tiny:

      python keras-YOLOv3-model-set/tools/model_converter/convert.py <path_to_cfg_file>/yolov4-tiny.cfg <path_to_weights>/yolov4-tiny.weights <saved_model_dir>
      
  4. Run model conversion for from the TensorFlow 2 format to an IR:

    Note

    Before you run the conversion, make sure you have installed all the model conversion API dependencies for TensorFlow 2.

    mo --saved_model_dir yolov4 --output_dir models/IRs --input_shape [1,608,608,3] --model_name yolov4
    

Converting YOLOv3 Model to the OpenVINO format

There are several public versions of TensorFlow YOLOv3 model implementation available on GitHub. This section explains how to convert YOLOv3 model from the repository (commit ed60b90) to an IR , but the process is similar for other versions of TensorFlow YOLOv3 model.

Overview of YOLOv3 Model Architecture

Originally, YOLOv3 model includes feature extractor called Darknet-53 with three branches at the end that make detections at three different scales. These branches must end with the YOLO Region layer.

Region layer was first introduced in the DarkNet framework. Other frameworks, including TensorFlow, do not have the Region implemented as a single layer, so every author of public YOLOv3 model creates it using simple layers. This badly affects performance. For this reason, the main idea of YOLOv3 model conversion to IR is to cut off these custom Region -like parts of the model and complete the model with the Region layers where required.

Dumping a YOLOv3 TensorFlow Model

To dump TensorFlow model out of GitHub repository (commit ed60b90), follow the instructions below:

  1. Clone the repository:

    git clone https://github.com/mystic123/tensorflow-yolo-v3.git
    cd tensorflow-yolo-v3
    
  2. (Optional) Checkout to the commit that the conversion was tested on:

    git checkout ed60b90
    
  3. Download coco.names file from the DarkNet website OR use labels that fit your task.

  4. Download the yolov3.weights (for the YOLOv3 model) or yolov3-tiny.weights (for the YOLOv3-tiny model) file OR use your pre-trained weights with the same structure.

  5. Install PIL, which is used by the conversion script in the repo:

    pip install pillow
    
  6. Run a converter:

    Note

    This converter works with TensorFlow 1.x and numpy 1.19 or lower.

    • For YOLO-v3:

      python3 convert_weights_pb.py --class_names coco.names --data_format NHWC --weights_file yolov3.weights
      
    • For YOLOv3-tiny:

      python3 convert_weights_pb.py --class_names coco.names --data_format NHWC --weights_file yolov3-tiny.weights --tiny
      

    At this step, you may receive a warning like WARNING:tensorflow:Entity <...> could not be transformed and will be executed as-is.. To work around this issue, switch to gast 0.2.2 with the following command:

    pip3 install --user gast==0.2.2
    

If you have YOLOv3 weights trained for an input image with the size different from 416 (320, 608 or your own), provide the --size key with the size of your image specified while running the converter. For example, run the following command for an image with size 608:

python3 convert_weights_pb.py --class_names coco.names --data_format NHWC --weights_file yolov3_608.weights --size 608

Converting a YOLOv3 TensorFlow Model to the OpenVINO format

To solve the problems explained in the YOLOv3 architecture overview section, use the yolo_v3.json or yolo_v3_tiny.json (depending on a model) configuration file with custom operations located in the <OPENVINO_INSTALL_DIR>/tools/model_optimizer/extensions/front/tf repository.

It consists of several attributes:

[
  {
    "id": "TFYOLOV3",
    "match_kind": "general",
    "custom_attributes": {
      "classes": 80,
      "anchors": [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326],
      "coords": 4,
      "num": 9,
      "masks":[[6, 7, 8], [3, 4, 5], [0, 1, 2]],
      "entry_points": ["detector/yolo-v3/Reshape", "detector/yolo-v3/Reshape_4", "detector/yolo-v3/Reshape_8"]
    }
  }
]

where:

  • id and match_kind are parameters that you cannot change.

  • custom_attributes is a parameter that stores all the YOLOv3 specific attributes:

    • classes, coords, num, and masks are attributes that you should copy from the configuration file that was used for model training. If you used DarkNet officially shared weights, you can use yolov3.cfg or yolov3-tiny.cfg configuration file from GitHub repository. Replace the default values in custom_attributes with the parameters that follow the [yolo] titles in the configuration file.

    • anchors is an optional parameter that is not used while inference of the model, but it used in a demo to parse Region layer output

    • entry_points is a node name list to cut off the model and append the Region layer with custom attributes specified above.

To generate an IR of the YOLOv3 TensorFlow model, run:

mo                                                   \
--input_model /path/to/yolo_v3.pb                                  \
--transformations_config front/tf/yolo_v3.json \
--batch 1                                                          \
--output_dir <OUTPUT_MODEL_DIR>

To generate an IR of the YOLOv3-tiny TensorFlow model, run:

mo                                                        \
--input_model /path/to/yolo_v3_tiny.pb                                  \
--transformations_config front/tf/yolo_v3_tiny.json \
--batch 1                                                               \
--output_dir <OUTPUT_MODEL_DIR>

where:

  • batch defines shape of model input. In the example, batch is equal to 1, but you can also specify other integers larger than 1.

  • transformations_config adds missing Region layers to the model. In the IR, the Region layer has name RegionYolo.

Note

The color channel order (RGB or BGR) of an input data should match the channel order of the model training dataset. If they are different, perform the RGB<->BGR conversion specifying the command-line parameter: reverse_input_channels. Otherwise, inference results may be incorrect. For more information about the parameter, refer to the When to Reverse Input Channels section of the Converting a Model to Intermediate Representation (IR) guide.

OpenVINO toolkit provides a demo that uses YOLOv3 model. Refer to the Object Detection C++ Demo for more information.

Converting YOLOv1 and YOLOv2 Models to the IR

Before converting, choose a YOLOv1 or YOLOv2 model version that best suits your task. Download model configuration file and corresponding weight file:

  • From DarkFlow repository : configuration files are stored in the cfg directory, links to weight files are given in the README.md file. The files from this repository are adapted for conversion to TensorFlow using DarkFlow.

  • From DarkNet website and repository: configuration files are stored in the cfg directory of the repository, links to weight files are given on the YOLOv1 and YOLOv2 websites.

To convert DarkNet YOLOv1 and YOLOv2 models to the OpenVINO format, follow these steps:

  1. Install DarkFlow

  2. Convert DarkNet YOLOv1 or YOLOv2 model to TensorFlow using DarkFlow

  3. Convert TensorFlow YOLOv1 or YOLOv2 model to IR

You need DarkFlow to convert YOLOv1 and YOLOv2 models to TensorFlow. To install DarkFlow:

  1. Install DarkFlow required dependencies.

  2. Clone DarkFlow git repository:

    git clone https://github.com/thtrieu/darkflow.git
    
  3. Go to the root directory of the cloned repository:

    cd darkflow
    
  4. Install DarkFlow, using the instructions from the README.md file in the DarkFlow repository.

To convert YOLOv1 or YOLOv2 model to TensorFlow, go to the root directory of the cloned DarkFlow repository, place the previously downloaded *.cfg and *.weights files in the current directory and run the following command:

  • For YOLOv1:

    python3 flow --model yolov1.cfg --load yolov1.weights --savepb
    
  • For YOLOv2 with VOC dataset --labels argument should be specified and additional changes in the original exporting script are required. In the file change line 121 from self.offset = 16 to self.offset = 20. Then run:

    python3 flow --model yolov2-voc.cfg --load yolov2-voc.weights --labels voc-labels.txt --savepb
    

VOC labels can be found on the following link

General conversion command is:

python3 flow --model <path_to_model>/<model_name>.cfg --load <path_to_model>/<model_name>.weights --labels <path_to_dataset_labels_file> --savepb

For YOLOv1, the --labels argument can be skipped. If the model was successfully converted, you can find the <model_name>.meta and <model_name>.pb files. in built_graph subdirectory of the cloned DarkFlow repository.

File <model_name>.pb is a TensorFlow representation of the YOLO model.

Converted TensorFlow YOLO model is missing Region layer and its parameters. Original YOLO Region layer parameters are stored in the configuration <path_to_model>/<model_name>.cfg file under the [region] title.

To recreate the original model structure, use the corresponding yolo .json configuration file with custom operations and Region layer parameters when converting the model to the IR. This file is located in the <OPENVINO_INSTALL_DIR>/tools/model_optimizer/extensions/front/tf directory.

If chosen model has specific values of these parameters, create another configuration file with custom operations and use it for conversion.

To generate the IR of the YOLOv1 model, provide TensorFlow YOLOv1 or YOLOv2 model to model conversion API with the following parameters:

mo
--input_model <path_to_model>/<model_name>.pb       \
--batch 1                                       \
--scale 255 \
--transformations_config front/tf/<yolo_config>.json

where:

  • batch defines shape of model input. In the example, batch is equal to 1, but you can also specify other integers larger than 1.

  • scale specifies scale factor that input values will be divided by. The model was trained with input values in the range [0,1]. OpenVINO toolkit samples read input images as values in [0,255] range, so the scale 255 must be applied.

  • transformations_config adds missing Region layers to the model. In the IR, the Region layer has name RegionYolo. For other applicable parameters, refer to the Convert Model from TensorFlow guide.

Note

The color channel order (RGB or BGR) of an input data should match the channel order of the model training dataset. If they are different, perform the RGB<->BGR conversion specifying the command-line parameter: reverse_input_channels. Otherwise, inference results may be incorrect. For more information about the parameter, refer to the When to Reverse Input Channels section of the Converting a Model to Intermediate Representation (IR) guide.