Source Code Generation with Generative Adversarial Networks (SeqGAN)
- Tensorflow r0.11
- Cuda 7.5 or higher (for GPU)
- nltk python package
Character Recurrent Neural Network
def media ( self ) :
choices = s
def cgets_to_reating_request ( _default ) :
charset = _errors
def field_with__ in get_language ( ) :
if func . __iter__
import unicode_litible
self . encode ( self ) :
if isinstance ( value , items ) :
if not os . path . lower for dotage_nurn :
pass
except XTERTAD_MI_NUN_FITCL
@ DEFILL
self . _funacod_location . copy ( i )
return s
It writes some texts that look like a program. But clearly Char-RNN is a bad programmer. (it can implement functions).
Word Recurrent Neural Network
class number_format ( _html_parser . signals . alias ) :
def __init__ ( self , commit = False ) :
widgets = value
content = [ ]
s = key
self . kwargs = current_app
self . _clean_form ( )
def required ( self , name ) :
value = self . add_prefix ( name )
def nud ( self , filter_expr , subdir ) :
value = force_text ( value )
else :
return self . _headers . contents tempdir . set_app ( cls )
if timezone . to_locale ( value ) :
return formats . zone msgs
As you can see, Word-RNN can hold on to the context longer than Char-RNN, thus it writes a longer program. (it can implement classes).
CodeGAN - Reinforce
Debugging...
CodeGAN - Polcy Gradient
(().__doc__=func.mime.<month>\d{1,2})-(?
contents.rel.creation.utils.models.tzinfo.creation.template.management.image.utils.utils.utils..value,
self.dirname(writer.month,d._meta.core.data.isdir(os.match(bmatch.value,
self.introspection.write(self.six.=self._archive.name=nameself.quote_name(self.dicts.tzinfo.utils.core.attname]=.')
parser.%new_path,self.__args,**self.pk.add(opts.model.get(',','._out.compress(pickle.children.checks.lru_cacheinstead.
_out.style.nodelist_true.xml.management.path.dayifself.empty_form.utils.!
template..errnoore.field_name).options._current[field.%s.mail.paginator.iteritems(self.serializers.day,
dt.match(bmatch._default_manager.fk.utils.__class__.data.xml.rawdataself.write(self.append(self._dir)or
utils._meta.__dict__.validators.abspath(os.platform.')parser.'._tzname=self.!
utils.forms.path.date.match(register_to.value,self._func_pathself.path.*ritem.loader.utils.attname]=.endswith('.data.file)
path.widget.style.stream._out=outself.xml.widget.path..content).compile(r(^[!
environ.')parser.stdout.datetime.core.andnote.utils.moves.moves.size=sizeself.http._errors[field].core.stdout.*?
__name__,bits.))\.://,re.rel.#$%&'*+/=?
db.:\.isdir(os.utils.write(self.utils.core.path.path.=y._css.:\.
Obviously, something went wrong. I will investigate this after the finals. :)
The model used for the code generation is called Sequence Generative Adversarial Nets (with Policy Gradient).
The illustration of SeqGAN. Left: D is trained over the real data and the generated data by G. Right: G is trained by policy gradient where the final reward signal is provided by D and is passed back to the intermediate action value via Monte Carlo search.
The research paper SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient has been accepted at the Thirty-First AAAI Conference on Artificial Intelligence (AAAI-17). The final version of the paper will be updated soon.
Move to codegan-pg folder and run
python pretrain_experiment.py
will start maximum likelihood training with default parameters. In the same folder, run
python sequence_gan.py
will start SeqGAN training.
This is one of many exiting projects going on in the DeepCoding Project. Stay tuned for more awesome stuff.
Note: I built it on top of the original implementation of SeqGAN which is based on the previous work by ofirnachum. Many thanks to ofirnachum and LantaoYu.
After running the experiments, the learning curve should be like this:
