TensorFlow Library:

checkpoint_explore

explore the checkpoint's parameter's name and shape

useage:

python checkpoint_explore.py --model_checkpoint *.ckpt-*

tensorboard_demo

a demo show how to add scalar and graph to tensorboard

finetune_with_multi_models

if you add some layer to one trained model ,and you want to load the trained model's para and only want to train the other parts of the model,you need this,include 2 functions,one for load trained model ,one for only train the the other parts of the whole model

pb_graph_explore

explore pb's parameter's name

useage:

python pb_graph_explore.py

import_pb_to_tensorboard

load pb's graph to tensorboard,only use a .pb file

useage:

python import_pb_to_tensorboard.py --model_dir=./froze_graph.pb --log_dir=./log
tensorboard --logdir=./log

finetune_import_arbitrarily_op_with_tf

finetune with offical model,and changed class num,so the model shape is different,this code can help import arbitrarily op.

caffe_optimise_quantize

change caffe_root to your dir where installed caffe

python3 merge_bn_scale_droupout.py --model deploy.prototxt --weights yolov3.caffemodel --output_model deploy_mergebn.prototxt --output_weights yolov3_mergebn.caffemodel
python3 caffe-int8-convert-tool-dev-weight.py --proto=deploy_mergebn.prototxt --model=yolov3_mergebn.caffemodel --mean 127.5 127.5 127.5 --norm=0.007843 --images=./testimgs/ --output=yolov3.table
caffe2ncnn deploy_mergebn.prototxt yolov3_mergebn.caffemodel yolov3_int8.param yolov3_int8.bin 256 yolov3.table

then use the .param and .bin as usual,ncnn can rec int8 mode auto.