• Automation framework: CK
  
• Development repository: ck-object-detection
  
• Source: GitHub
  
• Tags: object-detection,model,tf,tensorflow,yolo,yolo-v3,non-quantized,vcoco
  
• Support for host OS: linux
  
• Support for target OS: linux
  
• Version: reference
  
• Software detection plugin: model.tensorflow.object-detection
  
• How to get the stable version from this portal via the CK client:
  
  pip install cbench
  cb download package:model-tf-yolo-v3-coco --version=2.5.0 --all
  ck install package:model-tf-yolo-v3-coco
  ck help package
  
• How to get the development version:
  
  pip install ck
  ck pull repo:ck-object-detection
  ck install package:model-tf-yolo-v3-coco
  ck help package
  
• Useful CK automation actions (CLI with the Python CK API and JSON IO):
  
  ck show env
  ck find env:{ID from above}
  ck virtual env:{ID from above}
  
• CLI and Python API: module:package
  

Object Detection - TensorFlow YOLO-v3 Model Package

This package contains an example of a "custom" model for the object-detection-tf-py application. This application has function hooks that allow anyone to integrate into the application a model that is not structured in the same way as models in the TensorFlow model zoo. These functions are in two Python files in this package: custom_hooks.py and custom_tensorRT.py.

custom_hooks.py

The file has to expose 5 functions: - ck_custom_preprocess, ck_custom_preprocess_batch - ck_custom_postprocess, ck_custom_postprocess_batch - ck_custom_get_tensors

NB: The batch processing functions (ck_custom_*process_batch) have the same interface as the non-batch ones (ck_custom_*process). The main difference is in the first dimension of input/output: for example, if the non-batch functions work with tensors of shape [1, H, W, C], the batch ones work with tensors of shape [N, H, W,C].

TODO: Describe how the --env.CK_ENABLE_BATCH flag enables the choice of batch and non-batch functions.

In more detail, the function descriptions and parameters have to follow the following scheme.

ck_custom_preprocess, ck_custom_preprocess_batch

These functions are in charge of preparing the input image for the detection. They must produce the input tensor and some other helper data.

ck_custom_postprocess, ck_custom_postprocess_batch

These functions are in charge of producing the output of the detection. They must read output tensors and produce a txt file with the detections, and, if requested, output images with the boxes.

No Output Parameters

ck_custom_get_tensors

These function is in charge of getting the input and output tensors from the model graph.

No Input Parameters

custom_tensorRT.py

This file contains 3 functions required to support the TensorRT backend:

• load_graph_tensorrt_custom
• convert_from_tensorrt
• get_handles_to_tensors_RT

TODO: Describe how the --env.CK_ENABLE_TENSORRT flag enables the choice the backend.

In more detail, function must support the interfaces as follows:

load_graph_tensorrt_custom:

This function is in charge of loading the graph from a frozen model.

No Output Parameters

convert_from_tensorrt

This function is in charge of converting the output to a dictionary (since TensorRT outputs a list, not a dictionary).

get_handles_to_tensors_RT

This function is in charge of getting the input and output tensors from the model graph.

No Input Parameters

The internal tensor representation is strictly linked to the model, and the application is completely agnostic in this aspect. The programmer is in charge to keep the coherency between the preprocess, get tensor and postprocess functions.

• 2.5.0 (2020-09-04)
• 2.0.3 (2020-09-03)
• 1.0.1 (2020-09-02)
• 1.0.0 (2019-10-08)

• .cm/desc.json (3 bytes)
• .cm/info.json (317 bytes)
• .cm/meta.json (1418 bytes)
• README.md (5531 bytes)
• custom_hooks.py (16772 bytes)
• custom_tensorRT.py (1623 bytes)