Created by: shippingwang

Doc & Folder Structure

• refine readme doc, remove the defunct pics, comments, etc.
• refine readme_cn doc, translate args explanation.
• remove models/ to legacy/, rename models_name/ to models/, and revise associated codes in the train/eval/infer python files.
• remove PIL reader to legacy/, keep the cv2 reader as default reader.
• move utility.py to utils
• Add Stage3 release pretrained models.

Train.py:

• Add args optimizer , etcs , add comments to explain how to train ImageNet when using RMSProp For now, we do not add this args.
• polish learning rate calculation, Using "ceil"
• polish log format.
• print learning rate while traing models.
• use cosine_decay api instead of python implementation.

Reader.py

• reader.py reader_cv2.py add cv2 resize parameters.

Others

• polish infer.py, use "==" to compare instead of "is", etc.
• fix the conflicts causing by redundacy "softmax" activation codes both in the model architecutre and train/eval/infer python files.
• Add save_inference args in infer.py
• add cosine_decay_with_warmup to support shufflenetv2 in utility.py
• Update run.sh