Since rgf_python hasn't any special requirements (for compiler, environment, etc.), I think it good idea to have wheels on PyPI site (and the sources in .tar.gz, of course). I believe providing successfully compiled binaries will prevent many strange errors like recent ones.

We need wheels for two platforms: first for macOS and Linux and second for Windows.

The final result should be similar to this one:

But each wheel for each platform should have 32bit and 64bit version.

Binaries we could get from Travis and Appveyor as artifacts (I can do this). The one problem I see now is that Travis hasn't 32bit machines, but I believe we'll overcome this problem 😃 .

@fukatani When you'll have time, please search how to appropriate name wheels according to target platforms and how to post them at PyPI. Or I can do it more later.

How to deal with this error:

Ran 0 examples: 0 success, 0 failure, 0 error

None Ran 0 examples: 0 success, 0 failure, 0 error

None Ran 0 examples: 0 success, 0 failure, 0 error

None Traceback (most recent call last): File "/Users/k.den/Desktop/For_Submission/1_source_code/test.py", line 25, in pred = rgf_model.predict_proba(X_eval)[:, 1] File "/usr/local/lib/python3.6/site-packages/rgf/sklearn.py", line 652, in predict_proba class_proba = clf.predict_proba(X) File "/usr/local/lib/python3.6/site-packages/rgf/sklearn.py", line 798, in predict_proba 'This is rgf_python error.'.format(_TEMP_PATH)) Exception: Model learning result is not found in /tmp/rgf. This is rgf_python error.

Process finished with exit code 1

For bugs and unexpected issues, please provide the following information, so that we could reproduce them on our system.

Environment Info

Operating System: MacOS Sierra 10.12 | Ubuntu 16.04.3 LTS

Python version: 3.6.1

rgf_python version: HEAD (pulled from github)

Whether test.py is passed or not: FAILED (errors=24)

Error Message

ModuleNotFoundError: No module named 'rgf.sklearn'; 'rgf' is not a package

Reproducible Example

from rgf.sklearn import RGFClassifier

This issue is related with a previous one. A month ago I wrapped rgf_python using the reticulate package in R. It can be installed on Linux, and somehow cumbersome on Macintosh and Windows (on Windows currently it works only from the command prompt). I opened the issue as suggested by @fukatani

Hello,

i have read the previous thread on the same post, but it does not seem to solve my problem, because the previous case had string included in dataset and all i have got are all numbers. Could you please let me know what could be the problem??

Much appreciated !

skf = StratifiedKFold(n_splits = kfold, random_state=1)
for i, (train_index, test_index) in enumerate(skf.split(X, y)):
    X_train, X_eval = X[train_index], X[test_index]
    y_train, y_eval = y[train_index], y[test_index]
   
    rgf_model = RGFClassifier(max_leaf=400,
                    algorithm="RGF_Sib",
                    test_interval=100,
                    verbose=True).fit( X_train, y_train)
    pred = rgf_model.predict_proba(X_eval)[:,1]
    print( "Gini = ", eval_gini(y_eval, pred) )

and

---------------------------------------------------------------------------
Exception                                 Traceback (most recent call last)
<ipython-input-17-b27ba3506d06> in <module>()
     12                     test_interval=100,
     13                     verbose=True).fit( X_train, y_train)
---> 14     pred = rgf_model.predict_proba(X_eval)[:,1]
     15     print( "Gini = ", eval_gini(y_eval, pred) )

C:\Anaconda3\lib\site-packages\rgf\sklearn.py in predict_proba(self, X)
    644                              % (self._n_features, n_features))
    645         if self._n_classes == 2:
--> 646             y = self._estimators[0].predict_proba(X)
    647             y = _sigmoid(y)
    648             y = np.c_[y, 1 - y]

C:\Anaconda3\lib\site-packages\rgf\sklearn.py in predict_proba(self, X)
    796         if not model_files:
    797             raise Exception('Model learning result is not found in {0}. '
--> 798                             'This is rgf_python error.'.format(_TEMP_PATH))
    799         latest_model_loc = sorted(model_files, reverse=True)[0]
    800 

Exception: Model learning result is not found in C:\Users\hp\temp\rgf. This is rgf_python error.

Fixed #122. Appveyor suggests only 1 parallel job at free tier, GitHub Actions - 20.

Should be considered as a continuation of #328. Same changes as for *nix OSes: latest R version; stop producing 32bit artifacts.

I suppose it's time to release a new version with the support of warm start.

@fukatani Please release new Python version, and then @mlampros please upload to CRAN new R version.

@fukatani Please attach Linux i686 executable file to GitHub release - I've just tested replacing files into wheels and it works locally, so should work on Travis too! :-)

Refer to https://github.com/fukatani/rgf_python/issues/81#issuecomment-348662123.

Hi @fukatani ! Can you add more platforms (Windows, MacOS) to Travis? I don't know how, but it's possible 😄 : [Screenshot from xgboost repo] Maybe it can help: https://github.com/dmlc/xgboost/blob/master/.travis.yml

If there is a limitation to number of tests, maybe it's better to split Python version tests between platforms: Windows + 2.7, Linux + 3.4, MacOS + 3.5 (I think you understand me).

I am having the above error. I have made rgf1.2 and have tested using rgf1.2's own perl test script. This works. I have installed rgf_python and run the python setup as specified. I have changed the two folder locations to rgf1.2..\rgf executable and a temp folder that exist.

In python when I try to import I get the error Cannot import name 'RGFClassifier'. I tried to run the exact code in the test.py script provided in with rgf_python and this same error occurs.

Strangely, I have /usr/local/lib/python3.5/dist-packages/rgf_sklearn-0.0.0-py3.5.egg/rgf in my path when I do run

import sys
sys.path

in python. I also in /usr/local/lib/python3.5/dist-packages I only have the rfg-sklearn-0.0.0-py3.5.egg and no rgf-sklearn as I would expect as the following appeared towards the end of the setup.py,

Extracting rgf_sklearn-0.0.0-py3.5.egg to /usr/local/lib/python3.5/dist-packages
Adding rgf-sklearn 0.0.0 to easy-install.pth file

Fixed #184.

This PR adds the support of warm-start in RGF estimators, save_model() method which is needed to obtain binary model file and for further passing in init_model argument.

Also, this PR adds tests with analysis of exception message (as I promised in https://github.com/RGF-team/rgf/pull/258#issuecomment-439685042).

For bugs and unexpected issues, please provide the following information, so that we could reproduce them on our system.

Environment Info

Operating System: Windows 10

RGF/FastRGF/rgf_python version: 3.5.0-9

Python version (for rgf_python errors): 3.5.0-9

Error Message

Reproducible Example

Error when running RGF from R console as shown in the pic. Installation of RGF should be working fine as shown in the pic. RGF was installed via devtools.

Hi there

I'm trying to install rgf/fastrgf and use the python wrapper to launch the executables.

I've installed using pip install rgf_python

However when i import the rgf module i get a user warning

UserWarning: Cannot find FastRGF executable files. FastRGF estimators will be unavailable for usage.
  warnings.warn("Cannot find FastRGF executable files. FastRGF estimators will be unavailable for usage.")

To fix this issue i've compiled the rgf and fastrgf binaries* and added them to my $PATH variable (confirmed in bash that they are in the PATH) however i still get the same error. I've looked a bit into the rgf/utils get_paths and is_fastrgf_executable functions however i'm not completely sure why it fails?

*binaries: i was not sure which binaries are needed so i've added the following rgf, forest_predict, forest_train, discretized_trainer, discretized_gendata, auc

System Python: conda 3.6.1 OS: ubuntu 16.04

Initial support for dumping the RGF model is already implemented in #161. At present it's possible to print the model to the console. But it's good idea to bring the possibility of dumping the model to the file (e.g. JSON).

@StrikerRUS:

Really like new features introduced in this PR. But please think about "real dump" of a model. I suppose it'll be more useful than just printing to the console.

@fukatani:

For example dump in JSON format like lightGBM. It's convenient and we may support it in the future, but we should do it with another PR.

I found not documented parameter f_ratio in RGF. This corresponding to LightGBM feature_fraction and XGB colsample_bytree.

I tried these parameter with boston regression example. In small max_leaf(300), f_ratio=0.9 improves score to 11.0 from 11.8, but in many max_leaf(5000), f_ratio=0.95 degrared score to 10.34 from 10.19810.

After all, is there no value to use f_ratio < 1.0?

#Now, sklearn integration tests for FastRGFClassifier and FastRGFClassifier.

FastRGF doesn't work well for small samples, that is reason for test failed. I doubt inside Fast RGF executable inside. I inspect Fast RGF by debugger, discretization boundaries are invalid.

At least we should raise understandable error from RGF python if discretization failed.