Hi Arash, Have you installed PySINGA via the following approaches ( http://singa.apache.org/en/docs/installation.html#install-pysinga)?
# in build/ folder cmake -DUSE_CUDA=ON -DUSE_PYTHON=ON .. make cd python pip install . or using the python wheel On Thu, Sep 22, 2016 at 2:02 PM, Arash Shafiei <[email protected]> wrote: > Dear Wang Wei, > > We are still working on the activity recognition project and have some > difficulties installing SINGA. > > The python2.7 is installed. Trying to do RNN training using "python2.7 > train.py linux_input.txt", we get the following error: > > ImportError: No module name singa. > > It seems that python cannot find the singa module. We tried to set the > path of singa in PYTHONPATH but it didn't help. > > Python is installed in: > > /usr/include/python2.7 > /usr/lib/python2.7 > /usr/bin/python2.7 > > Singa is installed in: > /usr/local/include/singa > /usr/local/bin/singa > > Thanks. > Arash > > > On Fri, Sep 16, 2016 at 3:24 PM, Wang Wei <[email protected]> wrote: > >> Sure. >> You can use the API of RNN (http://singa.apache.org/en/do >> cs/layer.html#singa.layer.RNN) to construct a supervised model and train >> it with labels. >> E.g., a simple model would be input->rnn->dense->softmax->cross-entropy >> loss. >> The loss is computed to reduce the prediction error against the ground >> truth labels. >> >> You may need to read some papers on activity recognition and find proper >> datasets, e.g. http://archive.ics.uci.edu/ml/ >> datasets/Heterogeneity+Activity+Recognition. >> I think one problem is that you need to do some feature engineering for >> the raw data, whose feature has few dimensions. >> We have tried to use the original 3-dimendional feature with a MLP model, >> the performance is not good (about 50% accuracy). But the 500-dimensional >> feature in the above dataset has about 90% accuracy. >> >> Best, >> Wei >> >> >
