Development and training of a CNN component using EMADL2CPP
Prerequisites
- Linux. Ubuntu Linux 16.04 and 18.04 were used during testing.
- Deep learning backend:
- MXNet
- training - generated is Python code. Required is Python 2.7 or higher, Python packages
h5py
,mxnet
(for training on CPU) or e.g.mxnet-cu75
for CUDA 7.5 (for training on GPU with CUDA, concrete package should be selected according to CUDA version). Follow official instructions on MXNet site - prediction - generated code is C++. Install MXNet using official instructions on MXNet site for C++.
- training - generated is Python code. Required is Python 2.7 or higher, Python packages
- Caffe2
- Install Caffe2 using provided instructions from this link.
- training - generated is Python code. Required is Python 2.7
- prediction - generated code is C++.
- MXNet
HowTo
- Define a EMADL component containing architecture of a neural network and save it in a
.emadl
file. For more information on architecture language please refer to CNNArchLang project. An example of NN architecture:
component VGG16{
ports in Z(0:255)^{3, 224, 224} image,
out Q(0:1)^{1000} predictions;
implementation CNN {
def conv(filter, channels){
Convolution(kernel=(filter,filter), channels=channels) ->
Relu()
}
def fc(){
FullyConnected(units=4096) ->
Relu() ->
Dropout(p=0.5)
}
image ->
conv(filter=3, channels=64, ->=2) ->
Pooling(pool_type="max", kernel=(2,2), stride=(2,2)) ->
conv(filter=3, channels=128, ->=2) ->
Pooling(pool_type="max", kernel=(2,2), stride=(2,2)) ->
conv(filter=3, channels=256, ->=3) ->
Pooling(pool_type="max", kernel=(2,2), stride=(2,2)) ->
conv(filter=3, channels=512, ->=3) ->
Pooling(pool_type="max", kernel=(2,2), stride=(2,2)) ->
conv(filter=3, channels=512, ->=3) ->
Pooling(pool_type="max", kernel=(2,2), stride=(2,2)) ->
fc() ->
fc() ->
FullyConnected(units=1000) ->
Softmax() ->
predictions
}
}
- Define a training configuration for this network and store it in a
.cnnt file
, the name of the file should be the same as that of the corresponding architecture (e.g.VGG16.emadl
andVGG16.cnnt
). For more information on architecture language please refer to CNNTrainLang project. An example of a training configuration:
configuration VGG16{
num_epoch:10
batch_size:64
normalize:true
load_checkpoint:false
optimizer:adam{
learning_rate:0.01
learning_rate_decay:0.8
step_size:1000
}
}
-
Generate GPL code for the specified deep learning backend using the jar package of a EMADL2CPP generator. The generator receives the following command line parameters:
-
-m
path to directory with EMADL models -
-r
name of the root model -
-o
output path -
-b
backend
Assume both the architecture definition
VGG16.emadl
and the corresponding training configurationVGG16.cnnt
are located in a foldermodels
and the target code should be generated intotarget
folder usingMXNet
backend. An example of a command is then:
java -jar embedded-montiarc-emadl-generator-0.2.4-SNAPSHOT-jar-with-dependencies.jar -m models -r VGG16 -o target -b MXNET
You can find the EMADL2CPP jar here
-
-
When the target code is generated, the corresponding trainer file (e.g.
CNNTrainer_<root_model_name>.py
in case of MXNet) can be executed.
Building and running an application for TORCS
Prerequisites
-
Linux. Ubuntu Linux 16.04 and 18.04 were used during testing.
-
Armadillo (at least armadillo version 6.600 must be used) Official instructions at Armadillo Website.
-
ROS, Java runtime environment, GCC/Clang and armadillo - install using your linux distribution tools, e.g. apt in Ubuntu:
apt-get install ros-base-dev clang openjdk-8-jre
-
MXNet - install using official instructions at MXNet Website for C++
-
TORCS (see below)
TORCS Installation
-
Download customized TORCS distribution from the DeepDriving site
-
Unpack downloaded archive, navigate to the
DeepDriving/torcs-1.3.6
directory -
Compile and install by running
./configure --prefix=/opt/torcs && make -j && make install && make datainstall
-
Remove original TORCS tracks and copy customized tracks:
&& cp -rf ../modified_tracks/* /opt/torcs/share/games/torcs/tracks/```
-
Start TORCS by running
/opt/torcs/bin/torcs
Further installation help can be found in the Readme file provided with the DeepDriving distribution.
TORCS Setup
-
Run TORCS
-
Configure race
- Select Race -> Quick Race -> Configure Race
- Select one of the maps with the chenyi- prefix and click Accept
- Remove all drivers from the Selected section on the left by selecting every driver and clicking (De)Select
- Select driver chenyi on the right side and add it by clicking (De)Select
- Add other drivers with the chenyi- prefix if needed
- Click Accept -> Accept -> New Race
Example of a drivers configuration screen:
-
Use keys
1-9
andM
to hide all the widgets such as the speedometer, map, etc. from the TORCS screen -
Use
F2
key to switch between camera modes to select the mode when the car or it's parts are not visible -
Use
PgUp/PgDown
keys to switch between cars and selectchenyi
- the car that does not drive on its own
Code generation and running the project
- Download and unpack the archive that contains all EMA and EMADL component for an application
- Run
generate.sh
script. It will generate the code to thetarget
folder, copy the handwritten part of the project (communication with TORCS via shared memory) as well as the weights of the trained CNN and finally build the project - Start TORCS and configure race as described above. Select mode where host car is not visible
- Go to the
target
folder and startrun.sh
script. It will open two three terminals: one for the ROS core, one for the TORCSCOmponent (application part responsible for communication with TORCS) and one for the Mastercomponent (application part generated from the models at step 2 which is repsondible for application logic)
Troubleshooting Help
ERROR: CNNPredictor_dp_mastercomponent_dpnet.h:4:33: fatal error: mxnet/c_predict_api.h: No such file or directory.
FIX: Copy compiled mxnet lib and include files to usr/lib and usr/include respectively. Replace YOUR_MXNET_REPOSITORY with your corresponding information:
cd YOUR_MXNET_REPOSITORY/incubator-mxnet/lib
sudo cp * /usr/lib
cd YOUR_MXNET_REPOSITORY/incubator-mxnet/include
sudo cp -r * /usr/include
ERROR: HelperA.h:79:28: error: ‘sqrtmat’ was not declared in this scope.
FIX: Copy compiled armadillo lib and include files to usr/lib and usr/include respectively. Replace YOUR_ARMADILLO_REPOSITORY and VERSION (e.g. 8.500.1) with your corresponding information:
cd YOUR_ARMADILLO_REPOSITORY/armadillo-VERSION
sudo cp libarmadillo* /usr/lib
cd YOUR_ARMADILLO_REPOSITORY/armadillo-VERSION/include
sudo cp -r * /usr/include
ERROR: Coordinator_dp_mastercomponent.cpp.o: undefined reference to symbol 'dsyrk_' usr/lib/libopenblas.so.0: error adding symbols: DSO missing from command line (after executing Run generate2ros.sh).
FIX: Once generate2ros.sh was executed, modify the file YOUR_TORCSDL_REPOSITORY/torcs_dl/doc/deep_driving_project/target/Mastercomponent/dp_mastercomponent/coordinator/CMakeLists.txt to include the blas and openblas libraries, i.e.:
target_link_libraries(Coordinator_dp_mastercomponent RosAdapter_dp_mastercomponent dp_mastercomponent Threads::Threads -lblas -lopenblas)
Then navigate to YOUR_TORCSDL_REPOSITORY/torcs_dl/doc/deep_driving_project/target and execute build_all.sh. Make sure you delete the build folders to remove the existed compilation configurations for both components:
cd YOUR_TORCSDL_REPOSITORY/torcs_dl/doc/deep_driving_project/target
bash build_all.sh
Finally, the deep driving project will be compiled successfully.