shall we provide a python interface?

Question

shall we provide a python interface?

llcao opened this issue 8 years ago · comments

Liangliang Cao commented 8 years ago

py interface may be slower, but easier to use....

Wei Tan · Answer 1 · Wed Apr 27 2016 08:43:29 GMT+0800 (China Standard Time)

yes. let us work on python script to invoke cuMF, and also preprocess the input data.

Wei Tan · Answer 2 · Thu May 05 2016 22:16:46 GMT+0800 (China Standard Time)

I would like to go even a step further, i.e., to wrap cuMF cuda kernels as tensorflow Ops and use TF to "re-write" the glue code. By this means cuMF can run on TF.

Wei Tan · Answer 3 · Fri May 13 2016 04:05:01 GMT+0800 (China Standard Time)

A Tensorflow interface is added at:
https://github.com/wei-tan/CuMF/blob/master/tensorflow/cumf_as_tensorflow_ops_test.ipynb

Liangliang Cao · Answer 4 · Fri May 13 2016 08:28:52 GMT+0800 (China Standard Time)

very cool!
forgive me if I am picky, it seems easier to use if you separate the whole function into (1) training and (2) testing

Wei Tan · Answer 5 · Fri May 13 2016 08:35:35 GMT+0800 (China Standard Time)

In this case, do we still split training set into training and validation?

Liangliang Cao · Answer 6 · Fri May 13 2016 09:06:27 GMT+0800 (China Standard Time)

not really, users may want to determine their training and testing set.
What I want is simply as:
[thetaT,XT] = als_module.train_als(csrRow, csrCol, csrVal, cscRow, cscCol, cscVal, cooRow)
cooValPred = als_module.pred_als(cooRowTest, cooColTest)
rmse = evaluate_mse(cooValTest, cooValPred)

Alain Richardt · Answer 7 · Sun Aug 28 2016 21:40:33 GMT+0800 (China Standard Time)

Hi, I am attempting to run the TensorFlow example - I run the build script, which builds OK and creates als_tf.so and libALS.so in the current directory, but when I execute "cumf_as_tensorflow_ops_test.py" it says

tensorflow.python.framework.errors.NotFoundError: libALS.so: cannot open shared object file: No such file or directory

So I move the libraries to /usr/lib :
cp *.so /usr/lib

and then running it generates the following error:

I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcublas.so locally I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcudnn.so locally I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcufft.so locally I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcuda.so.1 locally I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcurand.so locally Traceback (most recent call last): File "cumf_as_tensorflow_ops_test.py", line 10, in <module> als_module = tf.load_op_library(lib_path) File "/root/CuMF/env3/lib/python3.4/site-packages/tensorflow/python/framework/load_library.py", line 75, in load_op_library raise errors._make_specific_exception(None, None, error_msg, error_code) tensorflow.python.framework.errors.NotFoundError: /usr/lib/libALS.so: undefined symbol: _Z17updateXWithCGHostPfS_S_iif

Can I get some guidance on what went wrong please?

Wei Tan · Answer 8 · Sun Sep 11 2016 10:23:07 GMT+0800 (China Standard Time)

Hi sorry we updated the ALS code with a new .o file (the CG solver), but we have not updated it the build libALS.so file. I will update it soon and let you know when it is done.

Wei Tan · Answer 9 · Sun Sep 11 2016 23:18:35 GMT+0800 (China Standard Time)

@alainrichardt : I just pushed a update to the code fixing the issue you saw. Here is the result I got from python notebook:
Hello, TensorFlow!
[[ 0.91662455]]