def construct_samples(self):
corpus_path = os.path.join(self.corpus_dir,
self.testing_label_filename)
df = pd.read_json(corpus_path)
with open('val_ids.txt', 'w') as f:
f.write(corpus_path)
f.writelines(df['id'])
for caption_list, vid in zip(
df['caption'],
df['id']): # train_label is a list of caption-vid dicts
path = os.path.join(self.feat_dir, '{}.npy'.format(vid))
feat = np.load(path)
self.feat_dict[vid] = feat
for caption in caption_list: # a list of strings
if not all(ord(c) < 128 for c in caption):
continue # abandon captions with unusual chars
token_list = self.captionToTokenList(caption)
self.data_obj_list.append(
DataObject(myid=vid, caption_list=token_list))
self.data_obj_list = np.array(self.data_obj_list)
self.batch_max_size = len(self.data_obj_list)
print('[Validation] total data size: ' + str(self.batch_max_size))
def build_val_data_obj_list(self):
corpus_path = self.corpus_dir + self.json_filename
data_file = pd.read_json(corpus_path)
max_size = 0
for i in range(0, len(data_file['caption'])):
myid = data_file['id'][i]
path = self.feat_dir + myid + '.npy'
mydat = np.load(path)
str_list = data_file['caption'][i]
tmp_list = []
cap_len_list = []
for j in range(0, len(str_list)):
seq = text_to_word_sequence(str_list[j],
filters=filters,
lower=True,
split=" ")
tmp_list.append(seq)
cap_len_list.append(len(seq) + 1) # added <EOS>
obj = DataObject(path, myid, tmp_list, cap_len_list)
self.dat_dict[myid] = mydat
max_size += 1
self.data_obj_list.append(obj)
self.data_obj_list = np.array(self.data_obj_list)
self.batch_max_size = max_size
def build_train_data_obj_list(self):
corpus_path = self.corpus_dir + self.json_filename
data_file = pd.read_json(corpus_path)
max_size = 0
for i in range(0, len(data_file['caption'])):
myid = data_file['id'][i]
path = self.feat_dir + myid + '.npy'
mydat = np.load(path)
str_list = data_file['caption'][i]
self.dat_dict[myid] = mydat
for j in range(0, len(str_list)):
tmp_list = []
cap_len_list = []
seq = text_to_word_sequence(str_list[j],
filters=filters,
lower=True,
split=" ")
join = " ".join(seq)
tmp_list.append(seq)
cap_len_list.append(len(seq) + 1) # added <EOS> !!
obj = DataObject(path, myid, tmp_list, cap_len_list)
max_size += 1
self.data_obj_list.append(obj)
self.data_obj_list = np.array(self.data_obj_list)
self.batch_max_size = max_size
self.perm = np.arange(self.batch_max_size, dtype=np.int)
self.shuffle_perm()
print('[Training] total data size: ' + str(max_size))
def build_test_data_obj_list(self):
txt = open(self.test_dir + self.id_txt, 'r')
print('load txt: ' + self.test_dir + self.id_txt)
max_size = 0
for line in txt.readlines():
myid = line.split('\n')[0]
path = self.feat_dir + myid + '.npy'
mydat = np.load(path)
obj = DataObject(path, myid)
self.dat_dict[myid] = mydat
max_size += 1
self.data_obj_list.append(obj)
self.data_obj_list = np.array(self.data_obj_list)
self.batch_max_size = max_size
def construct_samples(self):
"""
collect captions and construct a list of samples
:return:
"""
vid_captions_path = os.path.join(self.corpus_dir,
self.training_label_filename)
df = pd.read_json(
vid_captions_path
) # r'D:\video_captioning\data\MLDS_hw2_data\training_label.json')
with open('train_ids.txt', 'w') as f:
f.writelines(df['id'])
train_caption_list = [] # a list of caption-list for each video
for caption_list, vid in zip(
df['caption'],
df['id']): # train_label is a list of caption-vid dicts
train_caption_list.append([])
path = os.path.join(self.feat_dir, '{}.npy'.format(vid))
feat = np.load(path)
self.feat_dict[vid] = feat
for caption in caption_list: # a list of strings
if not all(ord(c) < 128 for c in caption):
continue # abandon captions with unusual chars
token_list = self.captionToTokenList(caption)
train_caption_list[-1].append(
token_list) # for the last video's caption-list
self.data_obj_list.append(
DataObject(myid=vid, caption_list=token_list))
self.data_obj_list = np.array(self.data_obj_list)
self.batch_max_size = len(self.data_obj_list)
self.perm = np.arange(self.batch_max_size, dtype=np.int)
self.shuffle_perm()
print('[Training] total data size: ' + str(self.batch_max_size))
# construct vocab
self.word_freq_dict = defaultdict(int)
# total_word_count = 0.0
for caption_list in train_caption_list:
for caption in caption_list:
for token in caption:
self.word_freq_dict[token] += 1
# total_word_count += 1.0
# also save the testing's vocab
df = pd.read_json(
os.path.join(self.corpus_dir, self.testing_label_filename))
for caption_list in df['caption']:
for caption in caption_list:
token_list = self.captionToTokenList(caption)
for token in token_list:
self.word_freq_dict[token] += 1
# for word in self.word_freq_dict:
# self.word_freq_dict[word] /= np.sum(self.word_freq_dict.values())
# return a new list of k-v tuples, sorted by the freq of word (the value), in the ascending order (reverse)
# word_freq_list = sorted(iter(self.word_freq_dict.items()), key=lambda k_v: k_v[1], reverse=True)
self.idx_to_word = self.marker + list(self.word_freq_dict.keys())
# self.word_index_dict = dict([(self.vocabulary[i], i) for i in range(len(self.vocabulary))])
self.vocab_indices = {
word: idx
for idx, word in enumerate(self.idx_to_word)
}
# store in pickle
with open('vocab_indices.pkl', 'wb') as handle:
pickle.dump(self.vocab_indices, handle)
with open('idx_to_word.pkl', 'wb') as handle:
pickle.dump(self.idx_to_word, handle)
with open('word_freq.pkl', 'wb') as handle:
pickle.dump(self.word_freq_dict, handle)
return len(self.vocab_indices)