Comments (15)
moraval
commented on July 18, 2024
8
I solved it by "downgrading" fblualib (commit: bdcf94bb835c41b5d822371fb6f7b75c1d53efbc) and thpp (commit: ebb4fcbb4a9b3310ac0a36eb37775a45f111448b).
I have a question concerning the receptive field. What is the width of the column, which is send to the convolutional layer. Is it really just 1 pixel as was stated in the paper?
Thanks in advance.
@bgshih
from crnn.
mpech
commented on July 18, 2024
7
with a simple function as
template <class T>
int naiveDecoding(lua_State* L) {
fblualib::luaPushTensor(L, thpp::Tensor<T>({1,1}));
fblualib::luaPushTensor(L, thpp::Tensor<T>({1,1}));
return 2;
}
we do get a core so I don't know how to fix the core (I'm all new to lua and fb*) so hopefully someone more competent will give a help
However constructs such as
const thpp::Tensor<T> input = *fblualib::luaGetTensorChecked<T>(L, 1);
are wrong because
fblualib::luaGetTensorChecked<T>(L, 1)
actually instanciate a THType*
template <class Tensor>
TensorPtr<Tensor>::TensorPtr(THType* th) noexcept
: hasTensor_(th) {
if (hasTensor_) {
construct(th, true);
}
}
and return a TensorPtr BUT when you dereference the TensorPtr
Tensor& operator*() const noexcept { return *get(); }
you get the rawpointer (I presume linking to the very same THType*)
Since your TensorPtr is a temporary variable (not used by anyone) it may be disposed and its destructor is called
template <class Tensor>
TensorPtr<Tensor>::~TensorPtr() {
destroy();
}
template <class Tensor>
void TensorPtr<Tensor>::destroy() noexcept {
if (hasTensor_) {
tensor_.~Tensor();
hasTensor_ = false;
}
}
then your input variable has an underlying THType* pointing to a wrong area
Better (at least somehow working) code is:
const auto& input = fblualib::luaGetTensorChecked<T>(L, 1);//construct
const int nFrame = input->size(0);
const int inputLength = input->size(1);
or
const auto& dummy = fblualib::luaGetTensorChecked<T>(L, 1);//simply increase the life duration of the underlying thtype* instance
const thpp::Tensor<T>& input = *dummy;
long story short below diff works for me
diff --git a/crnn/src/cpp/ctc.cpp b/crnn/src/cpp/ctc.cpp
index 97f9231..a752d2f 100644
--- a/crnn/src/cpp/ctc.cpp
+++ b/crnn/src/cpp/ctc.cpp
@@ -19,17 +19,19 @@ const int blankLabel = 0;
template <class T>
int forwardBackward(lua_State* L) {
- const thpp::Tensor<T> input = fblualib::luaGetTensorChecked<T>(L, 1);
- const thpp::Tensor<int> targets = fblualib::luaGetTensorChecked<int>(L, 2);
+ const auto& input = fblualib::luaGetTensorChecked<T>(L, 1);
+ const auto& dumtargets = fblualib::luaGetTensorChecked<int>(L, 2);
+ const thpp::Tensor<int>& targets = *dumtargets;
const bool forwardOnly = lua_toboolean(L, 3);
- thpp::Tensor<T> gradInput = fblualib::luaGetTensorChecked<T>(L, 4);
+ const auto& dumgrad = fblualib::luaGetTensorChecked<T>(L, 4);
+ thpp::Tensor<T> gradInput = *dumgrad;
- const int nFrame = input.size(0);
- const int inputLength = input.size(1);
- const int nClasses = input.size(2);
+ const int nFrame = input->size(0);
+ const int inputLength = input->size(1);
+ const int nClasses = input->size(2);
const int maxTargetlength = targets.size(1);
if (!forwardOnly) {
- gradInput.resizeAs(input);
+ gradInput.resizeAs(*input);
gradInput.fill(LogMath<T>::logZero);
}
@@ -41,7 +43,7 @@ int forwardBackward(lua_State* L) {
#pragma omp parallel for
for (int i = 0; i < nFrame; ++i) {
- const thpp::Tensor<T>& input_i = input[i];
+ const thpp::Tensor<T>& input_i = (*input)[i];
const int* targetData_i = targets[i].data();
const int targetLength = zeroPadArrayLength(targetData_i, maxTargetlength);
const int nSegment = 2 * targetLength + 1;
@@ -156,11 +158,11 @@ int forwardBackward(lua_State* L) {
template <class T>
int naiveDecoding(lua_State* L) {
- const thpp::Tensor<T>& input = fblualib::luaGetTensorChecked<T>(L, 1);
- const int nFrame = input.size(0);
- const int inputLength = input.size(1);
+ const auto& input = fblualib::luaGetTensorChecked<T>(L, 1);
+ const int nFrame = input->size(0);
+ const int inputLength = input->size(1);
- thpp::Tensor<long> rawPred_ = input.max(2).second; // [nFrame x inputLength]
+ thpp::Tensor<long> rawPred_ = input->max(2).second; // [nFrame x inputLength]
thpp::Tensor<int> rawPred({nFrame, inputLength});
from crnn.
umdreamer
commented on July 18, 2024
3
The compile error is because the type "thpp::Tensor&" is different with the type "thpp::Ptr<Tensor >", that is, the reference type and the pointer type.
For example:
const thpp::Tensor<T> input = fblualib::luaGetTensorChecked<T>(L, 1);
The fblualib::luaGetTensorChecked is defined as follows in the torch/install/include/fblualib/LuaUtils.h
template <class NT>
typename thpp::Tensor<NT>::Ptr luaGetTensorChecked(lua_State* L, int ud);
So the modification is the change all the fblualib::luaGetTensorChecked to *fblualib::luaGetTensorChecked, according to the file in thpp/thpp/TensorPtr.h
73 Tensor& operator*() const noexcept { return *get(); }
74 Tensor* operator->() const noexcept { return get(); }
75 Tensor* get() const noexcept;
The ctc.cpp modification is as follows:
--- a/crnn/src/cpp/ctc.cpp
+++ b/crnn/src/cpp/ctc.cpp
@@ -19,10 +19,10 @@ const int blankLabel = 0;
template <class T>
int forwardBackward(lua_State* L) {
- const thpp::Tensor<T> input = fblualib::luaGetTensorChecked<T>(L, 1);
- const thpp::Tensor<int> targets = fblualib::luaGetTensorChecked<int>(L, 2);
+ const thpp::Tensor<T> input = *fblualib::luaGetTensorChecked<T>(L, 1);
+ const thpp::Tensor<int> targets = *fblualib::luaGetTensorChecked<int>(L, 2);
const bool forwardOnly = lua_toboolean(L, 3);
- thpp::Tensor<T> gradInput = fblualib::luaGetTensorChecked<T>(L, 4);
+ thpp::Tensor<T> gradInput = *fblualib::luaGetTensorChecked<T>(L, 4);
const int nFrame = input.size(0);
const int inputLength = input.size(1);
@@ -156,7 +156,7 @@ int forwardBackward(lua_State* L) {
template <class T>
int naiveDecoding(lua_State* L) {
- const thpp::Tensor<T>& input = fblualib::luaGetTensorChecked<T>(L, 1);
+ const thpp::Tensor<T>& input = *fblualib::luaGetTensorChecked<T>(L, 1);
const int nFrame = input.size(0);
const int inputLength = input.size(1);
And then build, Done!
Cheers.
from crnn.
moraval
commented on July 18, 2024
The problem was not compatibility with TH++ but with the newest fblualib library. fblualib::luaGetTensorChecked() returns newly TensorPtr and not Tensor.
I fixed it by adding *() operator but now I can't run the demo.lua, I get segmentation fault using anything from LogMath.
Does anyone know what could be the problem?
@bgshih
from crnn.
bgshih
commented on July 18, 2024
Thanks for your feedback. I have just updated by fblualib to the latest version. However, the compilation goes well on my machine. Maybe it's due the OS or compiler. Still looking into this issue by viewing the thpp source.
from crnn.
moraval
commented on July 18, 2024
I updated my comment above.
The newest version of fblualib has different return value than is required in the code (ctc.cpp, row 22, const thpp::Tensor input = fblualib::luaGetTensorChecked(L, 1);).
It really doen't create any problems in your compilation?
from fblualib - LuaUtils-inl.h
template
typename thpp::Tensor::Ptr luaGetTensorChecked(lua_State* L, int ud) {
auto p = static_cast<typename thpp::Tensor::THType*>(
luaT_toudata(L, ud, thpp::Tensor::kLuaTypeName));
if (!p) {
luaL_error(L, "Not a valid tensor");
}
return typename thpp::Tensor::Ptr(p);
}
from crnn.
bgshih
commented on July 18, 2024
@moraval Glad to hear that. The widths are not 1-pixel. I have not calculate the precise width of the receptive fields. I guess it's 8 pixels. Note that CNN takes whole image as input, not the receptive fields.
from crnn.
CatWang
commented on July 18, 2024
@bgshih I have encountered the same problem and I have tried your solution of downgrading. But it seems that it doesn't work now.
Maybe this solution can work well at that time but not for now.
I wonder if you can provide some other solutions.
Thanks in advance.
from crnn.
bgshih
commented on July 18, 2024
@CatWang I tried again using Torch and THPP of the latest version. However, I still could not reproduce any compiling errors. Can you describe your environment, e.g. operating system?
from crnn.
liuwenran
commented on July 18, 2024
I meet the same problem and solve it like @moraval. Thanks
from crnn.
liuwenran
commented on July 18, 2024
but when I run demo, I got this:
Loading model...
/home/bbnc/torch/install/bin/luajit: ./utilities.lua:253: assertion failed!
stack traceback:
[C]: in function 'assert'
./utilities.lua:253: in function 'loadModelState'
demo.lua:32: in main chunk
[C]: in function 'dofile'
...bbnc/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:131: in main chunk
[C]: at 0x00406670
how to solve this? @moraval Thanks
from crnn.
jmrichardson
commented on July 18, 2024
I know this is an old thread, but having the same issue when compiling:
[ 66%] Building CXX object CMakeFiles/crnn.dir/ctc.cpp.o
/home/john/Downloads/crnn/src/cpp/ctc.cpp: In instantiation of ‘int {anonymous}::forwardBackward(lua_State*) [with T = float; lua_State = lua_State]’:
/home/john/Downloads/crnn/src/cpp/ctc.cpp:194:16: required from ‘const luaL_Reg {anonymous}::Registerer<float>::functions_ [3]’
/home/john/Downloads/crnn/src/cpp/ctc.cpp:203:24: required from ‘static void {anonymous}::Registerer<T>::registerFunctions(lua_State*) [with T = float; lua_State = lua_State]’
/home/john/Downloads/crnn/src/cpp/ctc.cpp:210:24: required from here
/home/john/Downloads/crnn/src/cpp/ctc.cpp:22:76: error: conversion from ‘thpp::TensorBase<float, thpp::Storage<float>, thpp::Tensor<float> >::Ptr {aka thpp::TensorPtr<thpp::Tensor<float> >}’ to non-scalar type ‘const thpp::Tensor<float>’ requested
const thpp::Tensor<T> input = fblualib::luaGetTensorChecked<T>(L, 1);
I have tried downgrading as suggested by @moraval , but unable to get them to compile due to other dependencies. Can you please help with a recommendation or updating the code? I am actually trying to use this code inside TextBoxes++ but also having the same issue with it's modified crnn package. Thanks in advance.
By the way, I have all the fblualib dependencies compiled, torch, etc. I am on Ubuntu 16.04
Thank you
from crnn.
umdreamer
commented on July 18, 2024
@jmrichardson Same problem, have you solved the problem yet?
[ 25%] Building CXX object CMakeFiles/crnn.dir/ctc.cpp.o
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp: In instantiation of ‘int {anonymous}::forwardBackward(lua_State*) [with T = float; lua_State = lua_State]’:
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:200:16: required from ‘const luaL_Reg {anonymous}::Registerer<float>::functions_ [3]’
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:209:24: required from ‘static void {anonymous}::Registerer<T>::registerFunctions(lua_State*) [with T = float; lua_State = lua_State]’
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:216:24: required from here
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:22:76: error: conversion from ‘thpp::TensorBase<float, thpp::Storage<float>, thpp::Tensor<float> >::Ptr {aka thpp::TensorPtr<thpp::Tensor<float> >}’ to non-scalar type ‘const thpp::Tensor<float>’ requested
const thpp::Tensor<T> input = fblualib::luaGetTensorChecked<T>(L, 1);
^
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:23:78: error: conversion from ‘thpp::TensorBase<int, thpp::Storage<int>, thpp::Tensor<int> >::Ptr {aka thpp::TensorPtr<thpp::Tensor<int> >}’ to non-scalar type ‘const thpp::Tensor<int>’ requested
const thpp::Tensor<int> targets = fblualib::luaGetTensorChecked<int>(L, 2);
^
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:25:76: error: conversion from ‘thpp::TensorBase<float, thpp::Storage<float>, thpp::Tensor<float> >::Ptr {aka thpp::TensorPtr<thpp::Tensor<float> >}’ to non-scalar type ‘thpp::Tensor<float>’ requested
thpp::Tensor<T> gradInput = fblualib::luaGetTensorChecked<T>(L, 4);
^
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp: In instantiation of ‘int {anonymous}::naiveDecoding(lua_State*) [with T = float; lua_State = lua_State]’:
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:200:16: required from ‘const luaL_Reg {anonymous}::Registerer<float>::functions_ [3]’
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:209:24: required from ‘static void {anonymous}::Registerer<T>::registerFunctions(lua_State*) [with T = float; lua_State = lua_State]’
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:216:24: required from here
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:159:73: error: invalid initialization of reference of type ‘const thpp::Tensor<float>&’ from expression of type ‘thpp::TensorBase<float, thpp::Storage<float>, thpp::Tensor<float> >::Ptr {aka thpp::TensorPtr<thpp::Tensor<float> >}’
const thpp::Tensor<T>& input = fblualib::luaGetTensorChecked<T>(L, 1);
^
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp: In instantiation of ‘int {anonymous}::forwardBackward(lua_State*) [with T = double; lua_State = lua_State]’:
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:200:16: required from ‘const luaL_Reg {anonymous}::Registerer<double>::functions_ [3]’
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:209:24: required from ‘static void {anonymous}::Registerer<T>::registerFunctions(lua_State*) [with T = double; lua_State = lua_State]’
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:217:25: required from here
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:22:76: error: conversion from ‘thpp::TensorBase<double, thpp::Storage<double>, thpp::Tensor<double> >::Ptr {aka thpp::TensorPtr<thpp::Tensor<double> >}’ to non-scalar type ‘const thpp::Tensor<double>’ requested
const thpp::Tensor<T> input = fblualib::luaGetTensorChecked<T>(L, 1);
^
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:23:78: error: conversion from ‘thpp::TensorBase<int, thpp::Storage<int>, thpp::Tensor<int> >::Ptr {aka thpp::TensorPtr<thpp::Tensor<int> >}’ to non-scalar type ‘const thpp::Tensor<int>’ requested
const thpp::Tensor<int> targets = fblualib::luaGetTensorChecked<int>(L, 2);
^
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:25:76: error: conversion from ‘thpp::TensorBase<double, thpp::Storage<double>, thpp::Tensor<double> >::Ptr {aka thpp::TensorPtr<thpp::Tensor<double> >}’ to non-scalar type ‘thpp::Tensor<double>’ requested
thpp::Tensor<T> gradInput = fblualib::luaGetTensorChecked<T>(L, 4);
^
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp: In instantiation of ‘int {anonymous}::naiveDecoding(lua_State*) [with T = double; lua_State = lua_State]’:
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:200:16: required from ‘const luaL_Reg {anonymous}::Registerer<double>::functions_ [3]’
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:209:24: required from ‘static void {anonymous}::Registerer<T>::registerFunctions(lua_State*) [with T = double; lua_State = lua_State]’
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:217:25: required from here
/home/data1/github-clone/TextBoxes_plusplus/crnn/src/cpp/ctc.cpp:159:73: error: invalid initialization of reference of type ‘const thpp::Tensor<double>&’ from expression of type ‘thpp::TensorBase<double, thpp::Storage<double>, thpp::Tensor<double> >::Ptr {aka thpp::TensorPtr<thpp::Tensor<double> >}’
const thpp::Tensor<T>& input = fblualib::luaGetTensorChecked<T>(L, 1);
^
CMakeFiles/crnn.dir/build.make:86: recipe for target 'CMakeFiles/crnn.dir/ctc.cpp.o' failed
make[2]: *** [CMakeFiles/crnn.dir/ctc.cpp.o] Error 1
CMakeFiles/Makefile2:67: recipe for target 'CMakeFiles/crnn.dir/all' failed
make[1]: *** [CMakeFiles/crnn.dir/all] Error 2
Makefile:83: recipe for target 'all' failed
make: *** [all] Error 2
I cloned the Textboxes_plusplus at the HEAD, and installed the dependencies,
- fblualib: has to modify the CMakeLists.txt on c++14,
- fbthrift: c++14 option
- folly: HEAD at v2018.04.02
- mstch: latest
- thpp: CMakeLists.txt
- torch: latest
- wangle: latest
- zstd: latest
There needs some modification for these libriaries to make them work. As I have to use the C++14 compilation option. My system is Ubuntu 16.04 x86_64, with CUDA 8.0 and CUDNN 5.1.
Now everything looks OK, but the crnn can not still compiled. The problem seems the fblualib has changed its API. Need to know which version the CRNN used to build.
from crnn.
AmirmasoudGhasemi
commented on July 18, 2024
I did what @umdreamer suggested. The code was built without error. but when I am running the demo.lua in Textbox_plusplus, the code stopped at the line after the changes(line 157):
const int nFrame = input.size(0);
It is the first time that input is called in the code. Do I need to change anything else?
from crnn.
umdreamer
commented on July 18, 2024
@mpech, Thank you for your modification. I get code compiled, and can run the demo sucessfully. As I didn't notice that the code fblualib::luaGetTensorChecked(L,1) actually is to create a new object. So in the runtime, the Core will be dumped to indicate error. Now it works using your code.
Cheers.
from crnn.
Related Issues (20)
- compile fbthrift in ubuntu 16.04
- Could you give me some suggestions for training dataset? HOT 1
- About BK-tree you mentioned in the paper HOT 2
- The recognition results are reversed
- 我是cpu环境咋办
- create_dataset occur something wrong
- index-pack failed
- how to generate rgb lmdb by your tools
- Estimating receptive field to feature sequence vector mapping
- lmdb.Error:../data: Function not implemented
- Nan will appear with GPU training, but not with CPU
- lmdb.Error HOT 4
- how to run this command in google colab th main_train.lua ../models/foo_model/
- Update README as some links is invalid
- Data Requirement
- which dataset I can use for crnn training ? HOT 2
- Outdated docker
- How to convert Synth900k to lmdb? HOT 1
- manifest unknown -> manifest for kaixhin/cuda-torch:latest not found: manifest unknown: manifest unknown
- Creating lmdb dataset in google colab HOT 1
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from crnn.