* add 0.00001 to denominator when computing loss

* use DEFAULT_EPSILON

* support .tif .img read

* add data analyse and check tools

* upgrade to adapt new dataset

* remove global

* optimize function, file names

* update train.py

* update post_quantization.py

* add background replace

* update optical flow

* update humanseg_postprocess

* update humanseg_postprocess

* add background replace doc

* Update video_infer.py

* Update bg_replace.py
Co-authored-by: Nwuzewu <wuzewu@baidu.com>

remove the useless third_party library from C++ inference library(zlib, eigen3, boost).  (reference https://github.com/PaddlePaddle/Paddle/pull/22021) (#272)
Co-authored-by: N陈亮 <chenliang@daheng-image.com>

* modify label tool

Co-authored-by: N陈亮 <chenliang@daheng-image.com>

* update model save load

* first add

* update model save and load

* update train.py

* update LaneNet model saving and loading

* adapt slim to paddle-1.8

* update distillation save and load

* update nas model save and load

* update model load op

* update utils.py

* update load_model_utils.py

* update model saving and loading

* update copyright

* update palette.py

* update model save load

* first add

* update model save and load

* update train.py

* update LaneNet model saving and loading

* adapt slim to paddle-1.8

* update distillation save and load

* update nas model save and load

* update model load op

* update utils.py

* update load_model_utils.py

* update model saving and loading

* update *_batch_size_like ops

* remove the preprocess and postprocess of model exporting

* modify label tool

* fix label tools bug

* add RemoteSensing

* optimize data clip and normalize method for remote sensing

* fix label problem using mobile devices

* add README.md

* polish README.md, optimize main.py and transforms.py, add transforms.md

* add demo dataset

* add create_dataset_list tool, requirements.txt

* add channel-by-channel normalize and clip, remove RemoteSensing import

* update vdl to 2.0

* polish README

* add tool to randomly split dataset and generate file list

* open eval_best_metric while traing, fix typo

* remove python2.7 ci

* add hrnet, polish README

* add metrics for unet like precision,recall,etc

* save more attrs in model.yaml, fix typo

* optimize predict_demo.py

* fix hrnet dice loss bug

* polish remote sensing README

* refactor model base class

* modify train_demo.py

* Update README.md

* Update requirements.txt

* update train.py

* update post_quantization.py

* Update vdl version to 2.0.0b1.

* Update vdl version to 2.0.0b1.

Co-authored-by: Nchenguowei01 <chenguowei01@baidu.com>