Check the preview of 2nd version of this platform being developed by the open MLCommons taskforce on automation and reproducibility as a free, open-source and technology-agnostic on-prem platform.
package:model-tf-yolo-v3-coco (v2.5.0)
Creation date: 2019-07-22
Source: GitHub
cID: 1dc07ee0f4742028:804598aedce689d7

Don't hesitate to get in touch if you encounter any issues or would like to discuss this community project!
Please report if this CK component works: 1  or fails: 0 
Sign up to be notified when artifacts are shared or updated!

Description  

This meta package is our attempt to provide a unified Python API, CLI and JSON meta description for different package managers and building tools to automatically download and install different components (models, data sets, libraries, frameworks, tools) necessary to run portable program pipelines across evolving platforms. Our on-going project is to make the onboarding process as simple as possible via this platform. Please check this CK white paper and don't hesitate to contact us if you have suggestions or feedback!

Dependencies    

ReadMe  

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.

Input Parameter Description
image_files list with all the filenames of the image to process, with full path
iter_num integer with the loop iteration value
processed_image_ids list with the ids of all the processed images, it's an in-out parameter (the function must append to this)
params dictionary with the application parameters
Output Parameter Description
image_data NumPy array to be fed to the detection graph (input tensor)
processed_image_ids see input parameters
image_size [list of] tuple with the sizes. depends if batch is used or not, if not is a single tuple
original_image [list of] list containing the original images as read before the modification done in preprocessing. may be useless.

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.

Input Parameter Description
image_files list with all the filenames of the image to process, with full path
iter_num integer with the loop iteration value
image_size [list of] tuple with the sizes. depends if batch is used or not, if not is a single tuple
original_image [list of] list containing the original images as read before the modification done in preprocessing. may be useless.
image_data NumPy array to be fed to the detection graph (input tensor)
output_dict output tensors. dictionary containing the tensors as "name : value" couples.
category_index dictionary to identify label and categories
params dictionary with the application parameters

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

Output Parameter Description
tensor_dict dictionary with the output tensors
input_tensor input tensor

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.

Input Parameter Description
params dictionary with the application parameters

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).

Input Parameter Description
output_dict Output tensors. if TensorRT, is a list/dictionary containing the output tensors with TensorRT names.
Output Parameter Description
output_dict Output tensors. dictionary/list containing the tensors as the postprocessing function requires.

get_handles_to_tensors_RT

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

No Input Parameters

Output Parameter Description
tensor_dict dictionary with the output tensors
input_tensor input tensor

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.

Versions  

Files  

Comments  

Please log in to add your comments!
If you notice any inapropriate content that should not be here, please report us as soon as possible and we will try to remove it within 48 hours!