Currently the only module which could
deal with the CNN directly is IPCV, but it is just for
inferencing, means we load the pre-trained model into Scilab
thought IPCV and use it for forward pass (prediction). Currently
the module could import models from caffe, tensorflow, and onnx
(some bugs detected on important ONNX, will fix it).
The link you sent was our initial
approach to try using the caffe-python binding and use the
scilab-python (PIMS) module to import the model for inferencing as
well. While PIMS is good for importing the python modules and use
in Scilab, however, the syntax would be very similar to the
Python, and some datatype are not compatible between 2. In this
case, might as well we just use the python for training?
We also used caffe framework for CNN
training before (using the binary) and using the Scilab to control
the setting, since their binary is reading some external files for
hyperparameters, and the google prototxt to design the CNN, which
make the design much easier. This went quite well but the I could
not see the future of caffe and seems that the attention was
shifted to caffe2 which is now merge into pytorch.
This, infact is a good news, as pytorch
does have a C++ interface, libtorch, which could be much easier to
be integrated into Scilab, and we have done some testing on the
integration and it seems working well, but to make this happened,
I would call upon more developers for Scilab to contribute ideas
and their experiences especially in C++ so the project could move
In a few weeks time, I will create a
git project, if any Scilab + C++ experts are interested, please
continue this conversation, I believe we could make something up.
On 23/8/2019 4:24 PM, P M wrote:
I am investigating if Scilab is capable to build a CNN or
So far I understood the ANN-Toolbox is not able to build
these neural network types.
Also the Neural-Network-Module does not seem to support
CNN's or FCNN's.
However it seems to be possible by using caffee together
with scilab's python toolbox.