def consult_skill(self, skill_name):
if skill_name in self.consulted_skills:
return
logging.debug("consulting skill '%s'" % skill_name)
m = self.load_skill(skill_name)
self.consulted_skills.add(skill_name)
skill_dir = self.skill_paths[skill_name]
try:
# print m
# print getattr(m, '__all__', None)
# for name in dir(m):
# print name
for m2 in getattr(m, 'DEPENDS'):
self.consult_skill(m2)
if hasattr(m, 'PL_SOURCES'):
for inputfn in m.PL_SOURCES:
pl_path = "%s/%s" % (skill_dir, inputfn)
pyxsb_command("consult('%s')." % pl_path)
except:
logging.error('failed to load skill "%s"' % skill_name)
logging.error(traceback.format_exc())
sys.exit(1)
return m
def consult_skill (self, skill_name):
if skill_name in self.consulted_skills:
return
logging.debug("consulting skill '%s'" % skill_name)
m = self.load_skill(skill_name)
self.consulted_skills.add(skill_name)
skill_dir = self.skill_paths[skill_name]
try:
# print m
# print getattr(m, '__all__', None)
# for name in dir(m):
# print name
for m2 in getattr (m, 'DEPENDS'):
self.consult_skill(m2)
if hasattr(m, 'PL_SOURCES'):
for inputfn in m.PL_SOURCES:
pl_path = "%s/%s" % (skill_dir, inputfn)
pyxsb_command("consult('%s')."% pl_path)
except:
logging.error('failed to load skill "%s"' % skill_name)
logging.error(traceback.format_exc())
sys.exit(1)
return m
def process_input(self, ctx, inp_raw, run_trace=False, do_eliza=True):
""" process user input, return score, responses, actions, solutions, context """
if run_trace:
pyxsb_command("trace.")
else:
pyxsb_command("notrace.")
tokens_raw = tokenize(inp_raw, ctx.lang)
tokens = []
for t in tokens_raw:
if t == u'nspc':
continue
tokens.append(t)
inp = u" ".join(tokens)
ctx.set_inp(inp)
self.mem_set(ctx.realm, 'action', None)
logging.debug(
'==============================================================================='
)
logging.debug('process_input: %s' % repr(inp))
#
# do we have an exact match in our training data for this input?
#
found_resp = False
for lang, d, md5s, args, src_fn, src_line in self.dte.lookup_data_train(
inp, ctx.lang):
afn, acode = self.dte.lookup_code(md5s)
ecode = '%s\n%s(ctx' % (acode, afn)
if args:
for arg in args:
ecode += ',%s' % repr(arg)
ecode += ')\n'
logging.debug('exact training data match found: %s:%s' %
(src_fn, src_line))
logging.debug(ecode)
# import pdb; pdb.set_trace()
try:
exec(ecode, globals(), locals())
found_resp = True
except:
logging.error('EXCEPTION CAUGHT %s' % traceback.format_exc())
logging.error(ecode)
if not found_resp:
logging.debug('no exact training data match for this input found.')
#
# ask neural net if we did not find an answer
#
resps = ctx.get_resps()
if not resps and self.nlp_model:
from nlp_model import _START, _STOP, _OR
logging.debug('trying neural net on: %s' % repr(inp))
try:
# ok, exact matching has not yielded any results -> use neural network to
# generate response(s)
# import pdb; pdb.set_trace()
predicted_ids = self.nlp_model.predict(inp)
# x = self.nlp_model.compute_x(inp)
# # logging.debug("x: %s -> %s" % (utterance, x))
# source, source_len, dest, dest_len = self.nlp_model._prepare_batch ([[x, []]], offset=0)
# # predicted_ids: GreedyDecoder; [batch_size, max_time_step, 1]
# # BeamSearchDecoder; [batch_size, max_time_step, beam_width]
# predicted_ids = self.tf_model.predict(self.tf_session, encoder_inputs=source,
# encoder_inputs_length=source_len)
# # for seq_batch in predicted_ids:
# # for k in range(5):
# # logging.debug('--------- k: %d ----------' % k)
# # seq = seq_batch[:,k]
# # for p in seq:
# # if p == -1:
# # break
# # decoded = self.inv_output_dict[p]
# # logging.debug (u'%s: %s' %(p, decoded))
# extract best codes, run them all to see which ones yield the highest scoring responses
cmd = []
# for p in predicted_ids[0][:,0]:
for decoded in predicted_ids:
# if p in self.inv_output_dict:
# decoded = self.inv_output_dict[p]
# else:
# decoded = p
if decoded == _STOP or decoded == _OR:
try:
logging.debug('trying cmd: %s' % repr(cmd))
afn, acode = self.dte.lookup_code(cmd[0])
ecode = '%s\n%s(ctx' % (acode, afn)
if len(cmd) > 1:
for arg in cmd[1:]:
ecode += ',%s' % repr(json.loads(arg))
ecode += ')\n'
logging.debug(ecode)
exec(ecode, globals(), locals())
except:
logging.debug('EXCEPTION CAUGHT %s' %
traceback.format_exc())
cmd = []
if decoded == _STOP:
break
else:
cmd.append(decoded)
except:
# probably ok (prolog code generated by neural network might not always work)
logging.error('EXCEPTION CAUGHT %s' % traceback.format_exc())
pyxsb_command("notrace.")
#
# extract highest-scoring responses
#
resps = ctx.get_resps()
if not resps and do_eliza:
logging.debug('producing ELIZA-style response for input %s' % inp)
from psychology import psychology
psychology.do_eliza(ctx)
resps = ctx.get_resps()
#
# pick random response
#
if len(resps) > 0:
i = random.randrange(0, len(resps))
out, score, action, action_arg = resps[i]
ctx.commit_resp(i)
logging.debug(u'picked resp #%d (score: %f): %s' % (i, score, out))
logging.debug(u'MEM: %s' % ctx.realm)
memd = self.mem_dump(ctx.realm)
for k, v, score in memd:
logging.debug(u'MEM: %-20s: %s (%f)' % (k, v, score))
logging.debug(u'MEM: %s' % ctx.user)
memd = self.mem_dump(ctx.user)
for k, v, score in memd:
logging.debug(u'MEM: %-20s: %s (%f)' % (k, v, score))
else:
out = u''
score = 0.0
logging.debug(u'No response found.')
action = self.mem_get(ctx.realm, 'action')
return out, score, action
def test_skill(self, skill_name, run_trace=False, test_name=None):
if run_trace:
pyxsb_command("trace.")
else:
pyxsb_command("notrace.")
m = self.skills[skill_name]
logging.info('running tests of skill %s ...' % (skill_name))
num_tests = 0
num_fails = 0
for tc in self.dte.lookup_tests(skill_name):
t_name, self.lang, prep_code, prep_fn, rounds, src_fn, self.src_line = tc
if test_name:
if t_name != test_name:
logging.info('skipping test %s' % t_name)
continue
ctx = self.create_context(user=TEST_USER,
realm=TEST_REALM,
test_mode=True)
round_num = 0
num_tests += 1
self.mem_clear(TEST_REALM)
self.mem_clear(TEST_USER)
# prep
if prep_code:
pcode = '%s\n%s(ctx)\n' % (prep_code, prep_fn)
try:
exec(pcode, globals(), locals())
except:
logging.error('EXCEPTION CAUGHT %s' %
traceback.format_exc())
for test_inp, test_out, test_action, test_action_arg in rounds:
logging.info("test_skill: %s round %d test_inp : %s" %
(t_name, round_num, repr(test_inp)))
logging.info("test_skill: %s round %d test_out : %s" %
(t_name, round_num, repr(test_out)))
# look up code in data engine
matching_resp = False
acode = None
ctx.set_inp(test_inp)
self.mem_set(ctx.realm, 'action', None)
for lang, d, md5s, args, src_fn, src_line in self.dte.lookup_data_train(
test_inp, self.lang):
afn, acode = self.dte.lookup_code(md5s)
ecode = '%s\n%s(ctx' % (acode, afn)
if args:
for arg in args:
ecode += ',%s' % repr(arg)
ecode += ')\n'
# import pdb; pdb.set_trace()
try:
exec(ecode, globals(), locals())
except:
logging.error(
'test_skill: %s round %d EXCEPTION CAUGHT %s' %
(t_name, round_num, traceback.format_exc()))
logging.error(ecode)
if acode is None:
logging.error(
u'Error: %s: no training data for test_in "%s" found in DB!'
% (t_name, test_inp))
num_fails += 1
break
resps = ctx.get_resps()
for i, resp in enumerate(resps):
actual_out, score, actual_action, actual_action_arg = resp
# logging.info("test_skill: %s round %d %s" % (clause.location, round_num, repr(abuf)) )
if len(test_out) > 0:
if len(actual_out) > 0:
actual_out = u' '.join(
tokenize(actual_out, self.lang))
logging.info(
"test_skill: %s round %d actual_out : %s (score: %f)"
% (t_name, round_num, actual_out, score))
if actual_out != test_out:
logging.info(
"test_skill: %s round %d UTTERANCE MISMATCH." %
(t_name, round_num))
continue # no match
logging.info("test_skill: %s round %d UTTERANCE MATCHED!" %
(t_name, round_num))
matching_resp = True
ctx.commit_resp(i)
# check action
if test_action:
afn, acode = self.dte.lookup_code(test_action)
ecode = '%s\n%s(ctx' % (acode, afn)
if test_action_arg:
ecode += ',%s' % repr(test_action_arg)
ecode += ')\n'
exec(ecode, globals(), locals())
break
if not matching_resp:
logging.error(
u'test_skill: %s round %d no matching response found.'
% (t_name, round_num))
num_fails += 1
break
round_num += 1
return num_tests, num_fails
def __init__(self,
db_url=DEFAULT_DB_URL,
xsb_arch_dir=DEFAULT_XSB_ARCH_DIR,
toplevel=DEFAULT_TOPLEVEL,
skill_paths=DEFAULT_SKILL_PATHS,
lang=DEFAULT_LANG,
nlp_model_args=DEFAULT_NLP_MODEL_ARGS,
skill_args=DEFAULT_SKILL_ARGS,
uttclass_model_args=DEFAULT_UTTCLASS_MODEL_ARGS):
self.lang = lang
self.nlp_model_args = nlp_model_args
self.skill_args = skill_args
self.uttclass_model_args = uttclass_model_args
#
# database connection
#
self.engine = model.data_engine_setup(db_url, echo=False)
self.Session = sessionmaker(bind=self.engine)
self.session = self.Session()
#
# TensorFlow (deferred, as tf can take quite a bit of time to set up)
#
self.tf_session = None
self.nlp_model = None
self.uttclass_model = None
#
# skill management, setup
#
self.skills = {} # skill_name -> module obj
self.skill_paths = {} # skill_name -> pathname
self.consulted_skills = set()
self.toplevel = toplevel
self.all_skills = []
# import pdb; pdb.set_trace()
if skill_paths:
for sp in skill_paths[::-1]:
sys.path.insert(0, sp)
else:
# auto-config
# __file__ -> ../skills
mp = os.path.dirname(os.path.abspath(__file__)) + '/skills'
sys.path.insert(0, mp)
# ./skills
cwd = os.getcwd()
sys.path.insert(0, cwd + '/skills')
# .
sys.path.insert(0, cwd)
for mp in sys.path:
logging.debug("Module search path: %s" % mp)
self.load_skill(toplevel)
#
# Prolog engine, data engine
#
pyxsb_start_session(xsb_arch_dir)
self.dte = DataEngine(self.session)
pyxsb_command('import default_sys_error_handler/1 from error_handler.')
pyxsb_command(
'assertz((default_user_error_handler(Ball):-default_sys_error_handler(Ball))).'
)
#
# restore memory
#
q = u''
for m in self.session.query(model.Mem):
v = json_to_xsb(m.v)
if q:
q += ', '
q += u"assertz(memory('%s', '%s', %s, %f))" % (m.realm, m.k,
unicode(v), m.score)
if not q:
q = u'assertz(memory(self, self, self, 1.0))'
q += u'.'
pyxsb_command(q)
def process_input (self, ctx, inp_raw, run_trace=False, do_eliza=True):
""" process user input, return score, responses, actions, solutions, context """
if run_trace:
pyxsb_command("trace.")
else:
pyxsb_command("notrace.")
tokens_raw = tokenize(inp_raw, ctx.lang)
tokens = []
for t in tokens_raw:
if t == u'nspc':
continue
tokens.append(t)
inp = u" ".join(tokens)
ctx.set_inp(inp)
self.mem_set (ctx.realm, 'action', None)
logging.debug('===============================================================================')
logging.debug('process_input: %s' % repr(inp))
#
# do we have an exact match in our training data for this input?
#
found_resp = False
for lang, d, md5s, args, src_fn, src_line in self.dte.lookup_data_train (inp, ctx.lang):
afn, acode = self.dte.lookup_code(md5s)
ecode = '%s\n%s(ctx' % (acode, afn)
if args:
for arg in args:
ecode += ',%s' % repr(arg)
ecode += ')\n'
logging.debug ('exact training data match found: %s:%s' % (src_fn, src_line))
logging.debug (ecode)
# import pdb; pdb.set_trace()
try:
exec (ecode, globals(), locals())
found_resp = True
except:
logging.error('EXCEPTION CAUGHT %s' % traceback.format_exc())
logging.error(ecode)
if not found_resp:
logging.debug('no exact training data match for this input found.')
#
# ask neural net if we did not find an answer
#
resps = ctx.get_resps()
if not resps and self.nlp_model:
from nlp_model import _START, _STOP, _OR
logging.debug('trying neural net on: %s' % repr(inp))
try:
# ok, exact matching has not yielded any results -> use neural network to
# generate response(s)
# import pdb; pdb.set_trace()
predicted_ids = self.nlp_model.predict(inp)
# x = self.nlp_model.compute_x(inp)
# # logging.debug("x: %s -> %s" % (utterance, x))
# source, source_len, dest, dest_len = self.nlp_model._prepare_batch ([[x, []]], offset=0)
# # predicted_ids: GreedyDecoder; [batch_size, max_time_step, 1]
# # BeamSearchDecoder; [batch_size, max_time_step, beam_width]
# predicted_ids = self.tf_model.predict(self.tf_session, encoder_inputs=source,
# encoder_inputs_length=source_len)
# # for seq_batch in predicted_ids:
# # for k in range(5):
# # logging.debug('--------- k: %d ----------' % k)
# # seq = seq_batch[:,k]
# # for p in seq:
# # if p == -1:
# # break
# # decoded = self.inv_output_dict[p]
# # logging.debug (u'%s: %s' %(p, decoded))
# extract best codes, run them all to see which ones yield the highest scoring responses
cmd = []
# for p in predicted_ids[0][:,0]:
for decoded in predicted_ids:
# if p in self.inv_output_dict:
# decoded = self.inv_output_dict[p]
# else:
# decoded = p
if decoded == _STOP or decoded == _OR:
try:
logging.debug('trying cmd: %s' % repr(cmd))
afn, acode = self.dte.lookup_code(cmd[0])
ecode = '%s\n%s(ctx' % (acode, afn)
if len(cmd)>1:
for arg in cmd[1:]:
ecode += ',%s' % repr(json.loads(arg))
ecode += ')\n'
logging.debug(ecode)
exec (ecode, globals(), locals())
except:
logging.debug('EXCEPTION CAUGHT %s' % traceback.format_exc())
cmd = []
if decoded == _STOP:
break
else:
cmd.append(decoded)
except:
# probably ok (prolog code generated by neural network might not always work)
logging.error('EXCEPTION CAUGHT %s' % traceback.format_exc())
pyxsb_command("notrace.")
#
# extract highest-scoring responses
#
resps = ctx.get_resps()
if not resps and do_eliza:
logging.debug ('producing ELIZA-style response for input %s' % inp)
from psychology import psychology
psychology.do_eliza(ctx)
resps = ctx.get_resps()
#
# pick random response
#
if len(resps)>0:
i = random.randrange(0, len(resps))
out, score, action, action_arg = resps[i]
ctx.commit_resp(i)
logging.debug(u'picked resp #%d (score: %f): %s' % (i, score, out))
logging.debug(u'MEM: %s' % ctx.realm)
memd = self.mem_dump(ctx.realm)
for k, v, score in memd:
logging.debug(u'MEM: %-20s: %s (%f)' % (k, v, score))
logging.debug(u'MEM: %s' % ctx.user)
memd = self.mem_dump(ctx.user)
for k, v, score in memd:
logging.debug(u'MEM: %-20s: %s (%f)' % (k, v, score))
else:
out = u''
score = 0.0
logging.debug(u'No response found.')
action = self.mem_get (ctx.realm, 'action')
return out, score, action
def test_skill (self, skill_name, run_trace=False, test_name=None):
if run_trace:
pyxsb_command("trace.")
else:
pyxsb_command("notrace.")
m = self.skills[skill_name]
logging.info('running tests of skill %s ...' % (skill_name))
num_tests = 0
num_fails = 0
for tc in self.dte.lookup_tests(skill_name):
t_name, self.lang, prep_code, prep_fn, rounds, src_fn, self.src_line = tc
if test_name:
if t_name != test_name:
logging.info ('skipping test %s' % t_name)
continue
ctx = self.create_context(user=TEST_USER, realm=TEST_REALM, test_mode=True)
round_num = 0
num_tests += 1
self.mem_clear(TEST_REALM)
self.mem_clear(TEST_USER)
# prep
if prep_code:
pcode = '%s\n%s(ctx)\n' % (prep_code, prep_fn)
try:
exec (pcode, globals(), locals())
except:
logging.error('EXCEPTION CAUGHT %s' % traceback.format_exc())
for test_inp, test_out, test_action, test_action_arg in rounds:
logging.info("test_skill: %s round %d test_inp : %s" % (t_name, round_num, repr(test_inp)) )
logging.info("test_skill: %s round %d test_out : %s" % (t_name, round_num, repr(test_out)) )
# look up code in data engine
matching_resp = False
acode = None
ctx.set_inp(test_inp)
self.mem_set (ctx.realm, 'action', None)
for lang, d, md5s, args, src_fn, src_line in self.dte.lookup_data_train (test_inp, self.lang):
afn, acode = self.dte.lookup_code(md5s)
ecode = '%s\n%s(ctx' % (acode, afn)
if args:
for arg in args:
ecode += ',%s' % repr(arg)
ecode += ')\n'
# import pdb; pdb.set_trace()
try:
exec (ecode, globals(), locals())
except:
logging.error('test_skill: %s round %d EXCEPTION CAUGHT %s' % (t_name, round_num, traceback.format_exc()))
logging.error(ecode)
if acode is None:
logging.error (u'Error: %s: no training data for test_in "%s" found in DB!' % (t_name, test_inp))
num_fails += 1
break
resps = ctx.get_resps()
for i, resp in enumerate(resps):
actual_out, score, actual_action, actual_action_arg = resp
# logging.info("test_skill: %s round %d %s" % (clause.location, round_num, repr(abuf)) )
if len(test_out) > 0:
if len(actual_out)>0:
actual_out = u' '.join(tokenize(actual_out, self.lang))
logging.info("test_skill: %s round %d actual_out : %s (score: %f)" % (t_name, round_num, actual_out, score) )
if actual_out != test_out:
logging.info("test_skill: %s round %d UTTERANCE MISMATCH." % (t_name, round_num))
continue # no match
logging.info("test_skill: %s round %d UTTERANCE MATCHED!" % (t_name, round_num))
matching_resp = True
ctx.commit_resp(i)
# check action
if test_action:
afn, acode = self.dte.lookup_code(test_action)
ecode = '%s\n%s(ctx' % (acode, afn)
if test_action_arg:
ecode += ',%s' % repr(test_action_arg)
ecode += ')\n'
exec (ecode, globals(), locals())
break
if not matching_resp:
logging.error (u'test_skill: %s round %d no matching response found.' % (t_name, round_num))
num_fails += 1
break
round_num += 1
return num_tests, num_fails
def __init__(self,
db_url = DEFAULT_DB_URL,
xsb_arch_dir = DEFAULT_XSB_ARCH_DIR,
toplevel = DEFAULT_TOPLEVEL,
skill_paths = DEFAULT_SKILL_PATHS,
lang = DEFAULT_LANG,
nlp_model_args = DEFAULT_NLP_MODEL_ARGS,
skill_args = DEFAULT_SKILL_ARGS,
uttclass_model_args = DEFAULT_UTTCLASS_MODEL_ARGS):
self.lang = lang
self.nlp_model_args = nlp_model_args
self.skill_args = skill_args
self.uttclass_model_args = uttclass_model_args
#
# database connection
#
self.engine = model.data_engine_setup(db_url, echo=False)
self.Session = sessionmaker(bind=self.engine)
self.session = self.Session()
#
# TensorFlow (deferred, as tf can take quite a bit of time to set up)
#
self.tf_session = None
self.nlp_model = None
self.uttclass_model = None
#
# skill management, setup
#
self.skills = {} # skill_name -> module obj
self.skill_paths = {} # skill_name -> pathname
self.consulted_skills = set()
self.toplevel = toplevel
self.all_skills = []
# import pdb; pdb.set_trace()
if skill_paths:
for sp in skill_paths[::-1]:
sys.path.insert(0,sp)
else:
# auto-config
# __file__ -> ../skills
mp = os.path.dirname(os.path.abspath(__file__)) + '/skills'
sys.path.insert(0, mp)
# ./skills
cwd = os.getcwd()
sys.path.insert(0, cwd + '/skills')
# .
sys.path.insert(0, cwd)
for mp in sys.path:
logging.debug ("Module search path: %s" % mp)
self.load_skill(toplevel)
#
# Prolog engine, data engine
#
pyxsb_start_session(xsb_arch_dir)
self.dte = DataEngine(self.session)
pyxsb_command('import default_sys_error_handler/1 from error_handler.')
pyxsb_command('assertz((default_user_error_handler(Ball):-default_sys_error_handler(Ball))).')
#
# restore memory
#
q = u''
for m in self.session.query(model.Mem):
v = json_to_xsb(m.v)
if q:
q += ', '
q += u"assertz(memory('%s', '%s', %s, %f))" % (m.realm, m.k, unicode(v), m.score)
if not q:
q = u'assertz(memory(self, self, self, 1.0))'
q += u'.'
pyxsb_command(q)