7u5 / aiface-tf2 Goto Github PK

original Forked from MrZhousf/tf-slim-inception ,i haven't fork inplace becoz i both have tf1 and tf2 version, are incompatible.

Modify the original code to support tensorflow2.3, change some of the codes structure that easier to modify.

you should install tf-slim in tensorflow 2 by "pip install tf-slim"

please run export PYTHONPATH="" first if there has PYTHONPATH value in Linux env. then source path.sh

(1) remove models-master dir, research move forward

(2) remove conf, config.py and train_inception.py move to root dir. config.py now can control model-name, steps, almost training config related configuration.

(3) remove duplicate {root_dir}/research/slim/datasets/{classname}.py, merge into a common.py

no longer to need create many files for new class creation. there has a common.py within {root_dir}/research/slim/datasets directory.

if you read Chinese, viewing Mrzhou's comparation amongs them ：InceptionV3、InceptionV4 image classification comparison

{classname} is class name of sub task classification， related to ../data/{classname} sample directory

if you serving a binary-classification: you can name pos_{classname} is positive sample, neg_{classname} is negative samples if you serving multi-classfication: you can try a classname as a taskname, then directory under ../data/{classname}/{sub-classname1-N}

image2tfrecord.py

// demo class is "flowers"

• train.py {classname}

• eval.py {classname}

• show_train.py // default using 6006 port. can be cutomized and view by http://{ip}:6006 // if you have relay machine, you need to open 6006 port mapping in tunnel config.
• show_eval.py // default using 6007 port. can be cutomized and view by http://{ip}:6006

• export.py {classname}

eval_single_img.py {classname} // single image eval_single_dir.py {classname} // a directory with many images

• config.py