vicariousinc/science_rcn

How to train using Yahoo or reCaptcha?

Opened this issue · 7 comments

I tried to using the program to train using Yahoo's data provided in your blog, but failed to print the following error message:

INFO:main:Training on 124 images...
INFO:main:Testing on 0 images...
Traceback (most recent call last):
File "science_rcn/run.py", line 230, in
parallel=options.parallel)
File "science_rcn/run.py", line 95, in run_experiment
print "Total test accuracy = {}".format(float(correct) / len(test_results))
ZeroDivisionError: float division by zero

We used the following commond:
python science_rcn/run.py --train_size 20 --test_size 20 --parallel

Thanks

Hi Yeguixin,

  • The problem occurs at:

INFO:main:Testing on 0 images...

This means run.py cannot access testing samples. I can reproduce your error by removing folders 0/1/2.. from /testing of the provided MNIST dataset. Therefore, it may be the folder structure, as line 133 def _load_data assumes...

    training/
        0/
        1/
        2/
        ...
    testing/
        0/
        ...

.. inside data/MNIST folder.

This git was made to work with MNIST at first so you have to modify the file(s) yourself if you want to run other datasets. A simple command won't do it in this case.

I know, you mean you only published the code to recognize MNIST? But I want to konw how to train and test the Yahoo or reCaptcha using the dataset you provided in your paper. Thanks.

Hi Yeguixin,

Sorry, but, just to clarify, it wasn't I who published the code/research. I just meant to help by what I could see in the code directly (run.py, line 2):

Reference implementation of a two-level RCN model for MNIST classification experiments.

I'm afraid I can't provide a better answer than that so you could wait for an official reply from the devs. Thanks.

I see. Thanks

@yeguixin Have you successfully applied it to the traning and testing in the reCaptcha dataset? if so, could you share you codes with me?

anyone succeed ?

did anyone got it ??