Adjusting the model the input and output format of `forward`,`cal_loss`
    and the `infer` function to unify the interfaces between the
  `model`,`preprocessor`,`postprocessor` and `visualizer`
modules.(LwOpenpose models finished)
2. Custom Model Class
    Provide a `BasicModel` class for users to inherit to make the model.
  customization process easier
3. Metric Manager
    Introducing MetricManager class to manage the loss record and
    statistic. The MetricManager object is passed to the `cal_loss` function
    automatically in the training procedure,thus also to make it convenient for user to record the
    training process when using the custom model.

  1.1 warp all the domain adaptation data pipeline into the
    Domainadapt_dataset class, to make the data pipeline clearer.
  1.2 debug the domain adaptation loss calculation and optimization
    problem in the tensorflow tape scope.
2. Standard logging info
  2.1 split the logging information of hyperpose into
    [DATA],[MODEL],[TRAIN] 3 parts to make the logging more clear
  2.2 using standard file stream and stdout stream in logging module
    to output log, and standardize the format of logging massages

  1.1 fit Kungfu new APIs

  1.1 warp all pre-processing logic into `PreProcessor` class
  1.2 warp all post-processing logic into `PostProcessor` class
  1.3 warp all visualizing logic into `Visualizer` class
  1.4 warp all plots procedure into `PltDrawer` class
  1.5 provide abstract class `BasicPreProcessor`, `BasicPostProcessor`,
    `BasicVisualizer` for users to inherite and plug in custom code
2. Unify model forwarding APIs
  2.1 unifying all model `forward` function to output a dict
    `predict_x`
  2.2 unifying all model `cal_loss` function to input two dicts
    `predict_x` and `target_x` to calculate loss
  above change aims to convenient the users when implementating custom model

  1.1 pifpaf model fixed.
2.Add help information when initialize configuration.
  2.1 variable definition
  2.2 object definition
  2.3 development platform basic usage
  2.4 development platform custom usage
  2.5 additional features

2. Dataset generation debug
3. Uniform channel format

2. pifpaf debuged

  unify model weight save format and load format to be `npz_dict`.
  (notes: `npz` format weights are ordered arrays, which are difficult to
   check and convert. `npz_dict` format weights are dict of arrays,
   which are convenient to examine and convert.)
2. Model weight examine API
2.1 add `Model.exam_model_weights(model)` function to check model weight
2.2 add `Model.exam_npz_dict_weights(npz_dict_path)` function to check
  npz_dict weight
2.3 add `Model.exam_npz_weights(npz_path)` function to check npz weight.

  1.1 adapt unified processor in evaluation pipeline
  1.2 fix tf eager tensor and numpy ndarray compatibility problem in
    post-processor
2. ImageProcessor Added
  2.1 add ImageProcessor that integrated image reading, scale and
    padding for user to easily convert the images to the model
    input format.
3. Python Demon
  3.1 add python_demo.py as an example for users to perform pose
    estimation demo using python module APIs with the provided `npz_dict` model
    weight.

  1.1 Pifpaf postprocessor debuged.
  1.2 PoseProposal postprocessor debuged.
2. Python demon debug
3. Human class description tidy up.
  3.1 formating human body joints output string format

  1.1 pb model conversion compatibility.
  1.2 tflite model conversion compatibility.