def __init__(self, word_features_vocab_sizes: List[int],
vec_features_size: int, hidden_size: int,
num_layers: int) -> None:
super().__init__()
self._word_features_encoder = maybe_cuda(
WordFeaturesEncoder(word_features_vocab_sizes, hidden_size, 1,
hidden_size))
self._features_classifier = maybe_cuda(
DNNScorer(hidden_size + vec_features_size, hidden_size,
num_layers))
def __init__(self, vec_features_size: int, word_feature_sizes: List[int],
hidden_size: int, num_layers: int) -> None:
super().__init__()
# Consider making the word embedding the same for all
# token-type inputs, also for tactic-type inputs
self._word_features_encoder = maybe_cuda(
WordFeaturesEncoder(word_feature_sizes, hidden_size, 1,
hidden_size))
self._features_classifier = maybe_cuda(
DNNScorer(hidden_size + vec_features_size, hidden_size,
num_layers))
def __init__(self,
wordf_sizes : List[int],
vecf_size : int,
hidden_size : int,
num_layers : int,
stem_vocab_size : int)\
-> None:
super().__init__()
self._word_features_encoder = maybe_cuda(
WordFeaturesEncoder(wordf_sizes, hidden_size, 1, hidden_size))
self._features_classifier = maybe_cuda(
DNNClassifier(hidden_size + vecf_size, hidden_size,
stem_vocab_size, num_layers))
self._softmax = maybe_cuda(nn.LogSoftmax(dim=1))
pass
def __init__(self, vec_features_size : int,
word_feature_vocab_sizes : List[int],
term_token_vocab_size : int,
hidden_size : int, num_layers : int,
tactic_vocab_size : int) -> None:
super().__init__()
self._goal_encoder = EncoderRNN(term_token_vocab_size, hidden_size, hidden_size)
self._hyp_encoder = EncoderRNN(term_token_vocab_size, hidden_size, hidden_size)
self._word_features_encoder = WordFeaturesEncoder(word_feature_vocab_sizes,
hidden_size, num_layers-1,
hidden_size)
self._layer = nn.Linear(hidden_size * 3 + vec_features_size, hidden_size)
self._out_layer = nn.Linear(hidden_size, tactic_vocab_size)
self._softmax = maybe_cuda(nn.LogSoftmax(dim=1))
pass
def _get_model(self, arg_values : Namespace,
tactic_vocab_size : int,
goal_vocab_size : int) \
-> CopyArgModel:
assert self._word_feature_functions
assert self._vec_feature_functions
feature_vec_size = sum([feature.feature_size()
for feature in self._vec_feature_functions])
word_feature_vocab_sizes = [feature.vocab_size()
for feature in self._word_feature_functions]
return CopyArgModel(FindArgModel(tactic_vocab_size,
goal_vocab_size, arg_values.max_length,
arg_values.hidden_size),
WordFeaturesEncoder(word_feature_vocab_sizes,
arg_values.hidden_size, 1,
arg_values.hidden_size),
DNNClassifier(arg_values.hidden_size + feature_vec_size,
arg_values.hidden_size, tactic_vocab_size,
3))