def download_im_with_text(search_query,
start,
end,
dataset_dir='data',
subdir='photos'):
"""Download images using the urls in the dataframe specified by the search query.
Parameters:
search_query: A string giving the sentiment to load the corresponding dataframe.
start: A start index for the loaded dataframe.
end: An end index for the loaded dataframe. -1 corresponds to the last row.
dataset_dir: A directory where the dataframes are stored.
subdir: A subdirectory to store the photos.
Returns:
Images downloaded in the directory dataset_dir/subdir/search_query, having
the posts ids as names.
"""
# Load data
emb_name = 'glove'
text_dir = 'text_model'
emb_dir = 'embedding_weights'
filename = 'glove.6B.50d.txt'
if emb_name == 'word2vec':
vocabulary, embedding = _load_embedding_weights_word2vec(
text_dir, emb_dir, filename)
else:
vocabulary, embedding = _load_embedding_weights_glove(
text_dir, emb_dir, filename)
df = preprocess_one_df(vocabulary, embedding, search_query, _POST_SIZE)
if end == -1:
end = df.shape[0]
links = df['photo']
# Create subdir if it doesn't exist
if not tf.gfile.Exists(os.path.join(dataset_dir, subdir)):
tf.gfile.MakeDirs(os.path.join(dataset_dir, subdir))
# Create search_query folder if it doesn't exist
photos_dir = os.path.join(dataset_dir, subdir, search_query)
if not tf.gfile.Exists(photos_dir):
tf.gfile.MakeDirs(photos_dir)
for i in range(start, end):
# Check for NaNs
if links[i] == links[i]:
# Open url and convert to JPEG image
try:
f = urlopen(links[i])
except Exception:
continue
image_file = io.BytesIO(f.read())
im = Image.open(image_file)
# The filename is the index of the image in the dataframe
filename = str(i) + '.jpg'
im.convert('RGB').save(os.path.join(photos_dir, filename), 'JPEG')
def convert_images_with_text(dataset_dir, num_valid, photos_subdir='photos', tfrecords_subdir='tfrecords'):
"""Downloads the photos and convert them to TFRecords.
Parameters:
dataset_dir: The data directory.
photos_subdir: The subdirectory where the photos are stored.
tfrecords_subdir: The subdirectory to store the TFRecords files.
"""
# Create the tfrecords_subdir if it doesn't exist
if not tf.gfile.Exists(os.path.join(dataset_dir, tfrecords_subdir)):
tf.gfile.MakeDirs(os.path.join(dataset_dir, tfrecords_subdir))
if _dataset_exists(dataset_dir, photos_subdir):
print('Dataset files already exist. Exiting without re-creating them.')
return
photo_filenames, class_names = _get_filenames_and_classes(
dataset_dir, photos_subdir, full_list='shuffled_test.csv')
class_names_to_ids = dict(zip(class_names, range(len(class_names))))
# Divide into train and test:
random.seed(_RANDOM_SEED)
random.shuffle(photo_filenames)
training_filenames = photo_filenames[num_valid:]
validation_filenames = photo_filenames[:num_valid]
# Load dataframes
df_dict = dict()
emb_name = 'glove'
text_dir = 'text_model'
emb_dir = 'embedding_weights'
filename = 'glove.6B.50d.txt'
if emb_name == 'word2vec':
vocabulary, embedding = _load_embedding_weights_word2vec(
text_dir, emb_dir, filename)
else:
vocabulary, embedding = _load_embedding_weights_glove(
text_dir, emb_dir, filename)
for emotion in class_names:
df_dict[emotion] = preprocess_one_df(
vocabulary, embedding, emotion, _POST_SIZE)
# First, convert the training and validation sets.
_convert_dataset_with_text('train', training_filenames, class_names_to_ids,
dataset_dir, df_dict, tfrecords_subdir)
_convert_dataset_with_text('validation', validation_filenames, class_names_to_ids,
dataset_dir, df_dict, tfrecords_subdir)
# Write the train/validation split size
train_valid_split = dict(
zip(['train', 'validation'], [len(photo_filenames) - num_valid, num_valid]))
train_valid_filename = os.path.join(
dataset_dir, photos_subdir, _TRAIN_VALID_FILENAME)
with tf.gfile.Open(train_valid_filename, 'w') as f:
for split_name in train_valid_split:
size = train_valid_split[split_name]
f.write('%s:%d\n' % (split_name, size))
# Finally, write the labels file:
labels_to_class_names = dict(zip(range(len(class_names)), class_names))
dataset_utils.write_label_file(
labels_to_class_names, dataset_dir, photos_subdir)
#_clean_up_temporary_files(dataset_dir)
print('\nFinished converting the dataset!')
'happy', 'sad', 'scared', 'angry', 'surprised', 'disgusted', 'annoyed',
'bored', 'love', 'calm', 'amazed', 'optimistic', 'pensive', 'ashamed',
'excited'
] #removed interested
df_dict = dict()
text_dir = 'text_model'
emb_dir = 'embedding_weights'
filename = 'glove.6B.50d.txt'
vocabulary, embedding = _load_embedding_weights_glove(text_dir, emb_dir,
filename)
_POST_SIZE = 50
for emotion in emotions:
df = preprocess_one_df(vocabulary, embedding, emotion, _POST_SIZE)
df_dict[emotion] = df
columns = [
u'id', u'post_url', u'type', u'timestamp', u'date', u'tags', u'liked',
u'note_count', u'photo', u'text', u'search_query', u'text_list',
u'text_len'
]
df_all = pd.DataFrame(columns=[])
for emotion in emotions:
df_all = pd.concat([df_all, df_dict[emotion]]).reset_index(drop=True)
# Get top 1000 most occurring words
vocabulary_count = dict(
zip(vocabulary, np.zeros(len(vocabulary), dtype=np.int32)))