Implement Sedna Python SDK support TinyMS

Motivation

Currently, Sedna only supports sklearn's xgboost, Tensorflow, Keras, PyTorch and MindSpore, but doesn’t support MindSpore’s high-level API tool TinyMS.

This proposal aims to implement Senda Python SDK support for TinyMS based on Sedna's lib development interface and implements an Incremental Learning demo.

Goals

Implement Sedna Python SDK support of TinyMS, which provides high level API to cover the entire lifecycle and workflow for AI development that ranges from data preparation to model deployment, including the following benefits:
- Compared with Keras, TinyMS designs a set of more abstract high level APIs and therefore is less complex than Keras. For example, one can complete dataset preprocessing with just one line of code in TinyMS.
- Compared with Fastai, TinyMS offers predefined MindSpore datasets which could help developers with dataset processing enormously as well as quick model deployment module.
Write user guide document of TinyMS

Proposal

Sedna Architecture Analysis

Taking the case yolo3 incremental learning in Sedna as an example, the user wraps the estimator for the specific model. when the incremental learning interface IncrementalLearning() is called, the following process takes place:

IncrementalLearning instantiates its parent class JobBase.
Call set_backend() in the constructor of JobBase to resolve the framework used by the model to TensorFlow and instantiate the corresponding TFBackend.
TFBackend inherits BackendBase, which is initialized with OS-related configuration when instantiated.
TFBackend implements the specific framework-related configuration.

Feasibility

Sedna parses the underlying framework of the model and runs the model using a different machine learning framework through a backend pair. So we need to :

implement Sedna's parsing of the underlying machining learning framework

implement TinyMSBackend

Design Details

set_backend()

Improve function set_backend() in sedna/backend/__init__.py

if backend_type == "TINYMS":
	from sedna.backend.tinyms import TinyMSBackend as REGISTER

TinyMS API

TinyMS is MindSpore's high-level API, but the TinyMS’s methods are not the same as the MindSpore framework’s , so we can not use the idea of inheritance for encapsulation.

TinyMSBackendclass

__init__(self, estimator, fine_tune=True, **kwargs) ：initialize
train(self, train_data, valid_data=None, **kwargs): train the model
predict(self, data, **kwargs) : use the model to preidct result
evaluate(self, data, **kwargs): evaluate the accuracy
load_weights(self): load model weight
get_weights(self) ：return weight
set_weights(self, weights) ：set weight

class TinyMSBackend(MSBackend):
    def __init__(self, estimator, fine_tune=True, **kwargs):
        ...
	
    def train(self, train_data, valid_data=None, **kwargs):
        ...
    
    def predict(self, data, **kwargs):
        ...
    
    def evaluate(self, valid_data, **kwargs):
        ...
  
    def save(self, model_path=None):
        ...
  
    def load(self, model_url=""):
        ...
  
    def load_weights(self):
	    ...
    
    def get_weights(self):
	    ...
	
    def set_weights(self, weights):
	    ...

Demo Based on MSBackend

implement MobileNetV2 with MindSpore
Build an estimator class based on MobileNetV2, encapsulating a series of methods such as train(),evaluate(),predict(), etc.
Call incremental learning API to accomplish model training, evaluation, and inference

Demo Based on TinyMSBackend

implement MobileNetV2 with TinyMS
Build an estimator class based on MobileNetV2, encapsulating a series of methods such as train(),evaluate(),predict(), etc.
Call incremental learning API to accomplish model training, evaluation and inference

Road Map

Implementing an incremental learning demo based on MSBackend
week1 : implement estimator
week2 : debug model
TinyMS（08/8-09/16）
week3 : learn TinyMSdocumentation
week4 : implement TinyMSBackend class
week5 : implement estimator
week6 : debug model
Write documentation（09/17-09/30）
week7 : write API documentation
week8 : write demo documentation and submit PR

5.2 kB Raw Blame History