def encode(self, state: Dict[Text, float]) -> np.ndarray:
"""Returns a binary vector indicating which features are active.
Given a dictionary of states (e.g. 'intent_greet',
'prev_action_listen',...) return a binary vector indicating which
features of `self.input_features` are in the bag. NB it's a
regular double precision float array type.
For example with two active features out of five possible features
this would return a vector like `[0 0 1 0 1]`
If intent features are given with a probability, for example
with two active features and two uncertain intents out
of five possible features this would return a vector
like `[0.3, 0.7, 1.0, 0, 1.0]`.
If this is just a padding vector we set all values to `-1`.
padding vectors are specified by a `None` or `[None]`
value for states.
"""
if not self.num_features:
raise Exception(
"BinarySingleStateFeaturizer was not prepared before encoding."
)
if state is None or None in state:
return np.ones(self.num_features, dtype=np.int32) * -1
# we are going to use floats and convert to int later if possible
used_features = np.zeros(self.num_features, dtype=np.float)
using_only_ints = True
for state_name, prob in state.items():
if state_name in self.input_state_map:
idx = self.input_state_map[state_name]
used_features[idx] = prob
using_only_ints = using_only_ints and utils.is_int(prob)
else:
logger.debug(
"Feature '{}' (value: '{}') could not be found in "
"feature map. Make sure you added all intents and "
"entities to the domain".format(state_name, prob))
if using_only_ints:
# this is an optimization - saves us a bit of memory
return used_features.astype(np.int32)
else:
return used_features
def encode(self, state: Dict[Text, float]) -> np.ndarray:
"""Returns a binary vector indicating which tokens are present."""
if not self.num_features:
raise Exception("LabelTokenizerSingleStateFeaturizer "
"was not prepared before encoding.")
if state is None or None in state:
return np.ones(self.num_features, dtype=np.int32) * -1
# we are going to use floats and convert to int later if possible
used_features = np.zeros(self.num_features, dtype=np.float)
using_only_ints = True
for state_name, prob in state.items():
using_only_ints = using_only_ints and utils.is_int(prob)
if state_name in self.user_labels:
if PREV_PREFIX + ACTION_LISTEN_NAME in state:
# else we predict next action from bot action and memory
for t in state_name.split(self.split_symbol):
used_features[self.user_vocab[t]] += prob
elif state_name in self.slot_labels:
offset = len(self.user_vocab)
idx = self.slot_labels.index(state_name)
used_features[offset + idx] += prob
elif state_name[len(PREV_PREFIX):] in self.bot_labels:
action_name = state_name[len(PREV_PREFIX):]
for t in action_name.split(self.split_symbol):
offset = len(self.user_vocab) + len(self.slot_labels)
idx = self.bot_vocab[t]
used_features[offset + idx] += prob
else:
logger.warning(
f"Feature '{state_name}' could not be found in feature map."
)
if using_only_ints:
# this is an optimization - saves us a bit of memory
return used_features.astype(np.int32)
else:
return used_features
def test_is_int():
assert utils.is_int(1)
assert utils.is_int(1.0)
assert not utils.is_int(None)
assert not utils.is_int(1.2)
assert not utils.is_int("test")
