def load_data(f):
input = open(labelsfile, "rb")
label_names = input.read().split("\n")[:36]
# print("\n".join(label_names))
mosi = Dataloader('http://sorena.multicomp.cs.cmu.edu/downloads/MOSI')
covarep = mosi.covarep()
sentiments = mosi.sentiments(
) # sentiment labels, real-valued. for this tutorial we'll binarize them
train_ids = mosi.train() # set of video ids in the training set
# sort through all the video ID, segment ID pairs
train_set_ids = []
for vid in train_ids:
for sid in covarep['covarep'][vid].keys():
train_set_ids.append((vid, sid))
train_set_audio = np.array([
norm(f, covarep['covarep'][vid][sid]) for (vid, sid) in train_set_ids
if covarep['covarep'][vid][sid]
])
train_set_audio[train_set_audio != train_set_audio] = 0
# y_data = np.array([sentiments[vid][sid] for (vid, sid) in train_set_ids])
# y_train_mc = multiclass(np.array([sentiments[vid][sid] for (vid, sid) in train_set_ids]))
y_train_bin = np.array(
[sentiments[vid][sid] for (vid, sid) in train_set_ids]) > 0
# normalize covarep and facet features, remove possible NaN values
# audio_max = np.max(np.max(np.abs(train_set_audio), axis=0), axis=0)
# train_set_audio = train_set_audio / audio_max
x_data = pandas.DataFrame(data=train_set_audio, columns=label_names)
return label_names, train_set_audio, x_data, y_train_bin
# recall that data at each time step is a tuple (start_time, end_time, feature_vector), we only take the vector
data = np.array([feature[2] for feature in data])
n_rows = data.shape[0]
dim = data.shape[1]
if max_len >= n_rows:
diff = max_len - n_rows
padding = np.zeros((diff, dim))
padded = np.concatenate((padding, data))
return padded
else:
return data[-max_len:]
if __name__ == "__main__":
# Download the data if not present
mosi = Dataloader('http://sorena.multicomp.cs.cmu.edu/downloads/MOSI')
embeddings = mosi.embeddings()
facet = mosi.facet()
covarep = mosi.covarep()
sentiments = mosi.sentiments(
) # sentiment labels, real-valued. for this tutorial we'll binarize them
train_ids = mosi.train() # set of video ids in the training set
valid_ids = mosi.valid() # set of video ids in the valid set
test_ids = mosi.test() # set of video ids in the test set
# Merge different features and do word level feature alignment (align according to timestamps of embeddings)
bimodal = Dataset.merge(embeddings, facet)
trimodal = Dataset.merge(bimodal, covarep)
dataset = trimodal.align('embeddings')
# sort through all the video ID, segment ID pairs