End-to-End_ML_Architecture.txt

@ Author: Shuiling Yu (yslenjoy@gmail.com)

1. Training Data Acquisition and Storage
	
	1.1 Possible open source dataset
		https://www.plant-image-analysis.org/dataset (Plant image analysis dataset)
		https://github.com/cwfid/dataset (Crop/ Weed Field Image Dataset, with segmentation mask and plant type)

	1.2 Data storage
	Data is stored on either local disk or cloud.
	Annotation should be added if any.


2. Dataset Curation

	Possible curation tools (https://www.bitss.org/wp-content/uploads/2014/06/tools-and-resources-for-data-curation-presentation3.pdf):
	DMPTool, DataUp, Dash, WAS

3. Model development

	3.1 Dataset preprocessing
		3.1.1 Extract label information
		3.1.2 Image resizing (cropping), random sampling, normalization and augmentation. 
			  Keras has convenient API for this process. (https://keras.io/preprocessing/image/)

	3.2 Construct and train suitable model.
		3.2.1 Possible ML models: 
			  Residual Net (easy to optimize and gain accuracy from increased depth), 
			  RNN, CNN...
		3.2.2 Parameter selection: initial weights, SGD batch size, iteration time, learning rate, weight decay and momentum.

	3.3 Model performance analysis based on validation dataset
		Use validation dataset to analysis trained model performance.
		Possible metric: confusion matrix.

	3.4 Model prediction and analysis
		Get prediction results and calculate prediction accuracy.

4. Moving Models to Production
	
	4.1 Prepare documentations for publication
	4.2 Prepare pre-trained weights

5. Serving Predictions to the Client
	
	5.1 Build a web-serve to evaluate client-submitted file automatically.
	5.2 Prepare demo or instructional tutorials.