def trainIO(train_index, test_index):
# Prepare data
train_input, train_output = data.getSplit(train_index)
test_input, test_output = data.getSplit(test_index)
datahelper = DataHelper(train_input, train_output, test_input, test_output,
config, data)
train_input = np.empty((len(train_input), 0))
test_input = np.empty((len(test_input), 0))
if config.use_target_text:
if config.use_bert:
train_input = np.concatenate(
[train_input,
datahelper.getTargetBertFeatures(mode='train')],
axis=1)
test_input = np.concatenate(
[test_input,
datahelper.getTargetBertFeatures(mode='test')],
axis=1)
else:
train_input = np.concatenate([
train_input,
np.array([
datahelper.pool_text(utt)
for utt in datahelper.vectorizeUtterance(mode='train')
])
],
axis=1)
test_input = np.concatenate([
test_input,
np.array([
datahelper.pool_text(utt)
for utt in datahelper.vectorizeUtterance(mode='test')
])
],
axis=1)
if config.use_target_video:
train_input = np.concatenate(
[train_input,
datahelper.getTargetVideoPool(mode='train')], axis=1)
test_input = np.concatenate(
[test_input,
datahelper.getTargetVideoPool(mode='test')], axis=1)
if config.use_target_audio:
train_input = np.concatenate(
[train_input,
datahelper.getTargetAudioPool(mode='train')], axis=1)
test_input = np.concatenate(
[test_input,
datahelper.getTargetAudioPool(mode='test')], axis=1)
if train_input.shape[1] == 0:
print("Invalid modalities")
exit(1)
# Aux input
if config.use_author:
train_input_author = datahelper.getAuthor(mode="train")
test_input_author = datahelper.getAuthor(mode="test")
train_input = np.concatenate([train_input, train_input_author], axis=1)
test_input = np.concatenate([test_input, test_input_author], axis=1)
if config.use_context:
if config.use_bert:
train_input_context = datahelper.getContextBertFeatures(
mode="train")
test_input_context = datahelper.getContextBertFeatures(mode="test")
else:
train_input_context = datahelper.getContextPool(mode="train")
test_input_context = datahelper.getContextPool(mode="test")
train_input = np.concatenate([train_input, train_input_context],
axis=1)
test_input = np.concatenate([test_input, test_input_context], axis=1)
train_output = datahelper.oneHotOutput(mode="train",
size=config.num_classes)
test_output = datahelper.oneHotOutput(mode="test", size=config.num_classes)
return train_input, train_output, test_input, test_output
def train_io(
config: Config, data: DataLoader, train_index: Iterable[int],
test_index: Iterable[int]
) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
train_input, train_output = data.get_split(train_index)
test_input, test_output = data.get_split(test_index)
datahelper = DataHelper(train_input, train_output, test_input, test_output,
config, data)
train_input = np.empty((len(train_input), 0))
test_input = np.empty((len(test_input), 0))
if config.use_target_text:
if config.use_bert:
train_input = np.concatenate([
train_input,
datahelper.get_target_bert_feature(mode="train")
],
axis=1)
test_input = np.concatenate(
[test_input,
datahelper.get_target_bert_feature(mode="test")],
axis=1)
else:
train_input = np.concatenate([
train_input,
np.array([
datahelper.pool_text(utt)
for utt in datahelper.vectorize_utterance(mode="train")
])
],
axis=1)
test_input = np.concatenate([
test_input,
np.array([
datahelper.pool_text(utt)
for utt in datahelper.vectorize_utterance(mode="test")
])
],
axis=1)
if config.use_target_video:
train_input = np.concatenate(
[train_input,
datahelper.get_target_video_pool(mode="train")],
axis=1)
test_input = np.concatenate(
[test_input,
datahelper.get_target_video_pool(mode="test")],
axis=1)
if config.use_target_audio:
train_input = np.concatenate(
[train_input,
datahelper.get_target_audio_pool(mode="train")],
axis=1)
test_input = np.concatenate(
[test_input,
datahelper.get_target_audio_pool(mode="test")],
axis=1)
if train_input.shape[1] == 0:
raise ValueError("Invalid modalities")
# Aux input
if config.use_author:
train_input_author = datahelper.get_author(mode="train")
test_input_author = datahelper.get_author(mode="test")
train_input = np.concatenate([train_input, train_input_author], axis=1)
test_input = np.concatenate([test_input, test_input_author], axis=1)
if config.use_context:
if config.use_bert:
train_input_context = datahelper.get_context_bert_features(
mode="train")
test_input_context = datahelper.get_context_bert_features(
mode="test")
else:
train_input_context = datahelper.get_context_pool(mode="train")
test_input_context = datahelper.get_context_pool(mode="test")
train_input = np.concatenate([train_input, train_input_context],
axis=1)
test_input = np.concatenate([test_input, test_input_context], axis=1)
train_output = datahelper.one_hot_output(mode="train",
size=config.num_classes)
test_output = datahelper.one_hot_output(mode="test",
size=config.num_classes)
return train_input, train_output, test_input, test_output