后续是否会支持将 tensorflow pb模型转为ncnn的模型呢?

[dangbo]

No description provided.

[nihui]

我也很想要这个转换器 qwq

[nihui]

TODO
reopen

[BiranLi]

And Mxnet?

[Wenstery]

现在tools目录下的tensorflow，是否已经开始支持tensorflow pb转ncnn?

[Daoctor]

@nihui hi, 我编译安装了 tensorflow2ncnn, 但是我尝试解析一个 tf 的 pb 文件时, 提示了这个错误

[libprotobuf FATAL /usr/local/include/google/protobuf/stubs/hash.h:204] Should never be called.
terminate called after throwing an instance of 'google::protobuf::FatalException'
  what():  Should never be called.
Aborted (core dumped)

跟进去看了下, 是函数 read_proto_from_binary中的这一句报错了 message->ParseFromCodedStream(&codedstr).
不知道是我的环境问题 还是别的原因 ...

[elva7073]

@Daoctor 我也是报这个问题 我用的是3.4.0的protobuf

[nihui]

以下repo里的 mnist 和 inception graph 可以正常转换使用了
https://github.com/miyosuda/TensorFlowAndroidMNIST
https://github.com/miyosuda/TensorFlowAndroidDemo

mnist expert-graph.pb 输入 28x28 灰度图黑底白字 norm_vals[1] = { 1/255.f }

inception tensorflow_inception_graph.pb 输入 224x224 RGB图 mean_vals[3] = { 117.f }

[kenttong]

write_graph导出的pb文件用tensorflow2ncnn转换失败，格式解析不对，找不到node里面的input tensor，因为input tensor的类型不是const。是需要用freeze_graph进行固化吗？

tf.train.write_graph(tf.get_default_graph().as_graph_def, './checkpoints/', 'graph.pb', False)

[Cv9527]

when compile tensorflow2nccc by qt with protobuf 3.3.0, I encounter the problems as follow:

op_def.pb.cc:(.text+0x3884)：对‘google::protobuf::internal::WireFormatLite::VerifyUtf8String(char const*, int, google::protobuf::internal::WireFormatLite::Operation, char const*)’未定义的引用
CMakeFiles/tensorflow2ncnn.dir/op_def.pb.cc.o：在函数‘tensorflow::OpDef_ArgDef::OpDef_ArgDef()’中：
op_def.pb.cc:(.text+0x3973)：对‘google::protobuf::internal::fixed_address_empty_string[abi:cxx11]’未定义的引用
CMakeFiles/tensorflow2ncnn.dir/op_def.pb.cc.o：在函数‘tensorflow::OpDef_AttrDef::OpDef_AttrDef()’中：
op_def.pb.cc:(.text+0x3a63)：对‘google::protobuf::internal::fixed_address_empty_string[abi:cxx11]’未定义的引用
CMakeFiles/tensorflow2ncnn.dir/op_def.pb.cc.o：在函数‘tensorflow::OpDeprecation::OpDeprecation()’中：
op_def.pb.cc:(.text+0x3b63)：对‘google::protobuf::internal::fixed_address_empty_string[abi:cxx11]’未定义的引用
CMakeFiles/tensorflow2ncnn.dir/op_def.pb.cc.o：在函数‘tensorflow::OpDef_AttrDef::~OpDef_AttrDef()’中：
op_def.pb.cc:(.text+0x3cff)：对‘google::protobuf::internal::fixed_address_empty_string[abi:cxx11]’未定义的引用
CMakeFiles/tensorflow2ncnn.dir/op_def.pb.cc.o：在函数‘tensorflow::OpDef_AttrDef::~OpDef_AttrDef()’中：
op_def.pb.cc:(.text+0x3e37)：对‘google::protobuf::internal::fixed_address_empty_string[abi:cxx11]’未定义的引用
CMakeFiles/tensorflow2ncnn.dir/op_def.pb.cc.o:op_def.pb.cc:(.text+0x3fe3): 跟着更多未定义的参考到 google::protobuf::internal::fixed_address_empty_string[abi:cxx11]

how could I solve them?

[wanghuijing]

用源码编译出来的tensorflow2ncnn转化了tensorflow_inception_graph.pb，修改源码里带的android demo加载model和param，修改squeezencnn_jni.cpp中为ex.input(0, in);和ex.extract(329, out);测试结果是物体类别识别都是0。请问支持tensorflow转ncnn吗？支持的话，是out设置的不对吗？

[mrlonely001]

你好， 我按照你这句 “mnist expert-graph.pb 输入 28x28 灰度图黑底白字 norm_vals[1] = { 1/255.f }”，修改了squeezenet.cpp 文件中的这几行，改为如下：
squeezenet.load_param("/home/pi/ncnn-master/tools/tensorflow/build/ncnn.proto");
squeezenet.load_model("/home/pi/ncnn-master/tools/tensorflow/build/ncnn.bin");
ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_GRAY, bgr.cols, bgr.rows, 28, 28);
const float mean_vals[1] = {1/255.f};
重新编译之后，运行：squeezenet /home/pi/1.jpg
报错：
find_blob_index_by_name data failed
find_blob_index_by_name prob failed
段错误

[WenguoLi]

tensorflow2ncnn 这个工具应该该怎么编译出来？
我编译的时候出错了：
versions.proto:1:10: Unrecognized syntax identifier "proto3". This parser only recognizes "proto2".
make[2]: *** [tools/tensorflow/versions.pb.cc] Error 1
make[1]: *** [tools/tensorflow/CMakeFiles/tensorflow2ncnn.dir/all] Error 2
make: *** [all] Error 2

请问前面的几位大神是怎么编译出来的，能够指导一下，多谢！

我安装了下载了最新的protobuf， 编译可以通过，但是make check failed, 我还是安装了libprotoc 3.5.1，

重新编译上面的tensorflow2ncnn的工具，还是报和上面一样的错误。

求大神们指点迷津啊！

[delongqilinksprite]

Hi,
我使用tensorflow2ncnn来转我的模型，遇到了以下的错误：
./tensorflow2ncnn frozen_model.pb
FusedBatchNorm not supported yet !
nepsilon #0
T #1
data_format #0
is_training #0
Elu not supported yet !
nT #1
FusedBatchNorm not supported yet !
nepsilon #0
T #1
data_format #0
is_training #0
Elu not supported yet !
nT #1
FusedBatchNorm not supported yet !
nepsilon #0
T #1
data_format #0
is_training #0
Elu not supported yet !
nT #1
FusedBatchNorm not supported yet !
nepsilon #0
T #1
data_format #0
is_training #0
Elu not supported yet !
nT #1
FusedBatchNorm not supported yet !
nepsilon #0
T #1
data_format #0
is_training #0
Elu not supported yet !
nT #1
FusedBatchNorm not supported yet !
nepsilon #0
T #1
data_format #0
is_training #0
Elu not supported yet !
nT #1
FusedBatchNorm not supported yet !
nepsilon #0
T #1
data_format #0
is_training #0
Elu not supported yet !
nT #1
FusedBatchNorm not supported yet !
nT #1
data_format #0
is_training #0
epsilon #0
Elu not supported yet !
nT #1
FusedBatchNorm not supported yet !
nT #1
data_format #0
is_training #0
epsilon #0
Elu not supported yet !
nT #1
FusedBatchNorm not supported yet !
nT #1
data_format #0
is_training #0
epsilon #0
Elu not supported yet !
nT #1
FusedBatchNorm not supported yet !
nT #1
data_format #0
is_training #0
epsilon #0
Elu not supported yet !
nT #1
FusedBatchNorm not supported yet !
nepsilon #0
T #1
data_format #0
is_training #0
Elu not supported yet !
nT #1
FusedBatchNorm not supported yet !
nT #1
data_format #0
is_training #0
epsilon #0
Elu not supported yet !
nT #1
FusedBatchNorm not supported yet !
nT #1
data_format #0
is_training #0
epsilon #0
FusedBatchNorm not supported yet !
nT #1
data_format #0
is_training #0
epsilon #0

这是不是意味着我需要修改tensorflow的网络模型呢？
@nihui 谢谢

[WanchaoYao]

补充一些不支持转换的op
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
Print not supported yet !
nT #3
U #0
first_n #0
message #0
summarize #0
Cast not supported yet !
nDstT #1
SrcT #3
ResizeBilinear not supported yet !
nT #1
align_corners #0
SpaceToBatchND not supported yet !
nT #1
Tblock_shape #3
Tpaddings #3
BatchToSpaceND not supported yet !
nT #1
Tblock_shape #3
Tcrops #3
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
StridedSlice not supported yet !
nIndex #3
T #3
begin_mask #0
ellipsis_mask #0
end_mask #0
new_axis_mask #0
shrink_axis_mask #0
ResizeBilinear not supported yet !
nT #1
align_corners #0
Pack not supported yet !
nN #0
T #3
axis #0
Slice not supported yet !
nIndex #3
T #3
finished

[ujsyehao]

期待resizebilinear operation支持！

[HudsonHuang]

I found an nice blog about this topic:
https://blog.csdn.net/u012477435/article/details/80765005

[lvchigo]

@Daoctor @elva7073 这个问题解决了吗？
尝试解析一个 tf 的 pb 文件时, 提示了这个错误：
[libprotobuf FATAL /usr/local/include/google/protobuf/stubs/hash.h:204] Should never be called. terminate called after throwing an instance of 'google::protobuf::FatalException' what(): Should never be called. Aborted (core dumped)

[laobadao]

Can you supported BasicLSTMCell op? please.

[nihui]

https://github.com/hanzy88/tensorflow2ncnn
https://github.com/hanzy88/ckpt2ncnn
note: this is a community developed tensorflow model converter

[ArtyZe]

@hunto Hello, I followed similar approach. but my param file is not loading in ncnn.

Here is starting of my param file generated by onnx2ncnn.
`7767517
199 199
Input input1 0 1
Padding padding 1 1 input1 conv1_pad_output 0=0 1=0 2=0 3=0 4=0 5=0.000000
Convolution convolution 1 1 conv1_pad_output conv1_output 0=8 1=3 11=3 2=1 12=1 3=2 13=2 5=0 6=216

`

it is not loading in ncnn library (load params).

1> as seen below, in input layer, there is no shape/channel declared. is it a problem?
well, I tried by adding this to first input layer: 0 1 input1 0=3 1=128 2=128
still params are not loading!

2> In normal param file I see layer has same name in parameter list , but here
padding-> conv1_pad_output
convolution -> conv1_output
and on...

does this makes difference?
Do I need to change 'padding' to 'conv1_pad_output' or
'conv1_pad_output' to 'padding'?

Hello, have you already solved this input shape problem? I have met the same problem

[Zrufy]

during the conversion i have this error

Unpack not supported yet !
naxis #0
T #3
num #0
Transpose not supported yet !
nT #1
Tperm #3

any solution???

[nihui]

针对onnx模型转换的各种问题，推荐使用最新的pnnx工具转换到ncnn

pip install pnnx
pnnx model.onnx inputshape=[1,3,224,224]

详细参考文档
https://github.com/pnnx/pnnx
https://github.com/Tencent/ncnn/wiki/use-ncnn-with-pytorch-or-onnx#how-to-use-pnnx