class DataLoader:
""" Reads in pickled activations that have been extracted.
Parameters
----------
activations_dir : str
Directory containing the extracted activations.
corpus : Corpus
Corpus containing the labels for each sentence.
test_activations_dir : str, optional
Directory containing the extracted test activations. If not
provided the train activation set will be split and partially
used as test set.
test_corpus : Corpus, optional
Corpus containing the test labels for each sentence. Must be
provided if `test_activations_dir` is provided.
selection_func : SelectFunc, optional
Selection function that determines whether a corpus item should
be taken into account. If such a function has been used during
extraction, make sure to pass it along here as well.
"""
def __init__(
self,
activations_dir: str,
corpus: Corpus,
test_activations_dir: Optional[str] = None,
test_corpus: Optional[Corpus] = None,
selection_func: SelectFunc = lambda sen_id, pos, example: True,
) -> None:
assert corpus is not None, "`corpus`should be provided!"
self.train_labels = create_labels_from_corpus(
corpus, selection_func=selection_func)
if test_activations_dir is not None:
self.test_activation_reader = ActivationReader(
test_activations_dir)
assert test_corpus is not None, "`test_corpus` should be provided!"
self.test_labels = create_labels_from_corpus(
test_corpus, selection_func=selection_func)
else:
self.test_activation_reader = None
self.test_labels = None
self.activation_reader = ActivationReader(activations_dir)
self.data_len = len(self.activation_reader)
def create_data_split(
self,
activation_name: ActivationName,
data_subset_size: int = -1,
train_test_split: float = 0.9,
) -> DataDict:
""" Creates train/test data split of activations
Parameters
----------
activation_name : ActivationName
(layer, name) tuple indicating the activations to be read in
data_subset_size : int, optional
Subset size of data to train on. Defaults to -1, indicating
the entire data set.
train_test_split : float
Percentage of the train/test split. If separate test
activations are provided this split won't be used.
Defaults to 0.9/0.1.
"""
if data_subset_size != -1:
assert (0 < data_subset_size <= self.data_len
), "Size of subset can't be bigger than the full data set."
train_activations = self.activation_reader.read_activations(
activation_name)
# Shuffle activations
data_size = self.data_len if data_subset_size == -1 else data_subset_size
indices = np.random.choice(range(data_size), data_size, replace=False)
train_activations = train_activations[indices]
train_labels = self.train_labels[indices]
if self.test_activation_reader is not None:
test_activations = self.test_activation_reader.read_activations(
activation_name)
test_labels = self.test_labels
else:
split = int(data_size * train_test_split)
test_activations = train_activations[split:]
test_labels = train_labels[split:]
train_activations = train_activations[:split]
train_labels = train_labels[:split]
return {
"train_x": train_activations,
"train_y": train_labels,
"test_x": test_activations,
"test_y": test_labels,
}
def main(config):
classifier = joblib.load('output/' + config.name +
'/classifiers/hx_l1.joblib')
train_reader = ActivationReader('output/' + config.name +
'/activations/train')
test_reader = ActivationReader('output/' + config.name +
'/activations/test')
if 'raw' in config.name:
representation = 'raw'
if 'neume' in config.name:
representation = 'neume'
if 'syllable' in config.name:
representation = 'syllable'
if 'pitch' in config.name:
notes = 'pitch'
if 'interval' in config.name:
notes = 'interval'
if '_20_' in config.name:
seq_length = 20
else:
seq_length = 30
# train_corpus = import_corpus_from_path('data/inputs/' + notes + '_20_' + representation + '_mode_corpus_train.txt', ['sen', 'labels'])
test_corpus = import_corpus_from_path(
'data/inputs/' + notes + '_' + str(seq_length) + '_' + representation +
'_mode_corpus_test.txt', ['sen', 'labels'])
test_genre_corpus = import_corpus_from_path(
'data/inputs/' + notes + '_' + str(seq_length) + '_' + representation +
'_string_genre_corpus_test.txt', ['sen', 'labels'])
##############################################################
##############################################################
hx_1_test = test_reader.read_activations((1, 'hx'))
test_labels = np.zeros(int(hx_1_test.shape[0] / seq_length))
test_genres = np.zeros(int(hx_1_test.shape[0] / seq_length))
for i in range(len(test_labels)):
test_labels[i] = test_corpus[i].labels[19]
test_genres[i] = test_genre_corpus[i].labels[19]
test_genres = test_genres.astype(int)
test_labels = test_labels.astype(int)
with open(
'data/inputs/' + notes + '_' + str(seq_length) + '_' +
representation + '_genre_vocab.txt', 'rb') as vf:
vocab_lines = vf.readlines()
vocab_lines = [line.decode('utf-8') for line in vocab_lines]
count = Counter(test_genres)
most_common_genres = count.most_common(3)
genre_vocab = [w.strip() for w in vocab_lines]
resp_verse = genre_vocab.index('Responsory verse')
antiphon = genre_vocab.index('Antiphon')
genre_ind_dict = {}
for genre in most_common_genres:
print(genre_vocab[genre[0]], genre)
genre_ind_dict[genre[0]] = np.where(np.array(test_genres) == genre[0])
ct = 1
embs = {}
for i in [0, 13]:
activation_test = hx_1_test[i::20]
activation_test = activation_test[genre_ind_dict[resp_verse]]
x_emb = TSNE(n_components=2, verbose=2).fit_transform(activation_test)
show_tsne_plot(x_emb, test_labels[genre_ind_dict[resp_verse]] + 1)
def main(config):
classifier = joblib.load('output/' + config.name +
'/classifiers/hx_l1.joblib')
train_reader = ActivationReader('output/' + config.name +
'/activations/train')
test_reader = ActivationReader('output/' + config.name +
'/activations/test')
if 'raw' in config.name:
representation = 'raw'
if 'neume' in config.name:
representation = 'neume'
if 'syllable' in config.name:
representation = 'syllable'
if 'pitch' in config.name:
notes = 'pitch'
if 'interval' in config.name:
notes = 'interval'
if '_20_' in config.name:
seq_length = 20
else:
seq_length = 30
if 'embedding' in config.name:
train_corpus = import_corpus_from_path(
'data/inputs/' + notes + '_' + str(seq_length) + '_' +
representation + '_string_mode_corpus_train.txt',
['sen', 'labels'])
test_corpus = import_corpus_from_path(
'data/inputs/' + notes + '_' + str(seq_length) + '_' +
representation + '_string_mode_corpus_test.txt', ['sen', 'labels'])
test_genre_corpus = import_corpus_from_path(
'data/inputs/' + notes + '_' + str(seq_length) + '_' +
representation + '_string_genre_corpus_test.txt',
['sen', 'labels'])
else:
train_corpus = import_corpus_from_path(
'data/inputs/' + notes + '_' + str(seq_length) + '_' +
representation + '_mode_corpus_train.txt', ['sen', 'labels'])
test_corpus = import_corpus_from_path(
'data/inputs/' + notes + '_' + str(seq_length) + '_' +
representation + '_mode_corpus_test.txt', ['sen', 'labels'])
# hx_1_train = train_reader.read_activations((1,'hx'))
# train_labels = np.zeros(int(hx_1_train.shape[0]/seq_length))
# for i in range(len(train_labels)):
# train_labels[i] = train_corpus[i].labels[19]
# train_scores = np.zeros(seq_length)
# train_stds = np.zeros(seq_length)
# for i in range(seq_length):
# activation_train = hx_1_train[i::seq_length]
# train_pred = classifier.predict(activation_train)
# train_scores[i] = (train_pred == train_labels).mean()
# plt.plot(range(1,seq_length+1), train_scores, c='Red')
##############################################################
##############################################################
hx_1_test = test_reader.read_activations((1, 'hx'))
test_labels = np.zeros(int(hx_1_test.shape[0] / seq_length))
test_genres = np.zeros(int(hx_1_test.shape[0] / seq_length))
for i in range(len(test_labels)):
test_labels[i] = test_corpus[i].labels[19]
test_genres[i] = test_genre_corpus[i].labels[19]
test_genres = test_genres.astype(int)
test_labels = test_labels.astype(int)
with open(
'data/inputs/' + notes + '_' + str(seq_length) + '_' +
representation + '_genre_vocab.txt', 'rb') as vf:
vocab_lines = vf.readlines()
vocab_lines = [line.decode('utf-8') for line in vocab_lines]
count = Counter(test_genres)
most_common_genres = count.most_common(25)
genre_vocab = [w.strip() for w in vocab_lines]
resp_verse = genre_vocab.index('Responsory verse')
antiphon = genre_vocab.index('Antiphon')
genre_ind_dict = {}
for genre in most_common_genres:
print(genre_vocab[genre[0]], genre)
genre_ind_dict[genre[0]] = np.where(np.array(test_genres) == genre[0])
test_scores = np.zeros(seq_length)
test_stds = np.zeros(seq_length)
test_scores_genre = {}
for genre in genre_ind_dict:
test_scores_genre[genre] = np.zeros(seq_length)
preds = []
for i in range(seq_length):
activation_test = hx_1_test[i::seq_length]
test_pred = classifier.predict(activation_test)
preds.append(test_pred)
for genre in genre_ind_dict:
tp = test_pred[genre_ind_dict[genre]]
tl = test_labels[genre_ind_dict[genre]]
test_scores_genre[genre][i] = (tp == tl).mean()
# test_scores[i] = (test_pred == test_labels).mean()
d = {
'antiphon': test_scores_genre[antiphon].tolist(),
'resp_verse': test_scores_genre[resp_verse].tolist()
}
with open('output/' + config.name + '/scores.json', 'w') as f:
json.dump(d, f)
# plt.plot(range(1,seq_length+1), test_scores, label='All genres')
for genre in genre_ind_dict:
plt.plot(range(1, seq_length + 1),
test_scores_genre[genre],
label=genre_vocab[genre])
# plt.plot(range(1,21), test_scores+test_stds, c='Blue')
# plt.plot(range(1,21), test_scores-test_stds, c='Blue')
plt.legend()
plt.xticks(list(range(1, seq_length + 1)))
plt.ylim(0, 1)
fig = plt.gcf()
plt.show()
ans = input('Do you want to save this figure? y/n\n')
if ans == 'y':
fig.savefig('output/' + config.name + '/eval_class.jpg')
print('done')