update:

Checked models:

VGG16

ResNet

Yolov3

The defined models can be found in https://github.com/hanzy88/tfmodel4ncnn, to be continued.

update 20190826

The yolov3 based on full cnn/ mobilenetv2 are checked successfully by tensorflow2ncnn, which refer to https://github.com/GuodongQi/yolo3_tensorflow.

Since the model was only trained with 16 epoches AND accuracy loss after conversion, the result was not so good, as shown in follow:

Since the related flies are too big, you can download by baiduyun:

https://pan.baidu.com/s/1mQdPuTeiRw2iimcyxU0WYw code: 3hkh

= = If it's helpful for you, please give a star. And if you are interested in tf2ncnn, welcome to improve it together to make tf2ncnn better.

Tensorflow2ncnn

Thanks for the share of ncnn, refer to https://github.com/Tencent/ncnn

This repository is mainly for the converting from tensorflow to ncnn directly based on the NCNN for Tensorflow https://github.com/jiangxiluning/ncnn-tensorflow.

First, build the ncnn:

Since the related cmakefiles have changed to rebuilt on my machine, you can follow the steps in ncnn to build the ncnn.

Second, convert from tensorflow to ncnn:

please check the related files in tools/tensorflow/, tensorflow2ncnn.cpp can convert the pb file to the ncnn.param and ncnn.bin. Once the tools are built, the converter can be found in myPro(i.e., the biuld files)/tools/.. For detail of the convert, you can follow the step by:

https://github.com/hanzy88/ckpt2ncnn

For the new layers built in tensorflow2ncnn corresponding to the layers in ncnn:

you can check the related files in src/layer, the new layers are started with tf* (forgive for my poor coding in the layer, it will be updated in the future).

If you want to build new op for tf2ncnn i.e., How to build new layer for ncnn, please check:

https://github.com/Tencent/ncnn/wiki/how-to-implement-custom-layer-step-by-step

Once the new layer are added, add the layer to make in src/CMakefiles, like: ncnn_add_layer(TFReshape). and then rebuild the project.

Third, Support OP

For now, the problem caused by original bactchnorm in ncnn-tensorflow has been solved.

And layer "Shape", "StridedSlice", "Pack", "ResizeBilinear", "LeakyRelu", "Relu6", "Range", "Tile", "Reshape", "Cast", "Mean"(only support for global average pooling, i.e., tf.reduce_mean(layer, reduction_indices=[1,2])) has been added or updated for tensorflow.

The normal CNN with FC has tested, tf.flatten NOT support yet because the weight file cannot be aligned correctly, but you can use reshape op with the caculated shape like tf.reshape(max_pool, [-1, 64]).