def make_vocab_for_embedding():
embedding_path = os.path.join(
root_path, "embedding/BioWordVec_PubMed_MIMICIII_d200.vec.bin")
MED = KeyedVectors.load_word2vec_format(embedding_path, binary=True)
# print(type(MED))
word_dict = MED.wv.vocab
# print(type(words))
# print(len(words))
raw_words = list(word_dict.keys())
print(len(raw_words))
# remove the word not appear in the dataset (training + val + test)
# load all word
all_word_in_data = set()
train_val_path = os.path.join(root_path,
"data/train_val/All_QA_Pairs_train_val.txt")
test_path = os.path.join(root_path,
"data/test/VQAMed2019_Test_Questions.txt")
with open(train_val_path, "r") as f:
print("process train_val file")
for row in f:
i_q_a = row.rstrip().split("|")
q_words = VQADataProvider.text_to_list(i_q_a[1])
q_words = q_words[:-1]
for w in q_words:
all_word_in_data.add(w)
with open(test_path, "r") as f:
print("process test file")
for row in f:
i_q = row.rstrip().split("|")
q_words = VQADataProvider.text_to_list(i_q[1])
q_words = q_words[:-1]
for w in q_words:
all_word_in_data.add(w)
# filter
words = []
print("filter")
for w in all_word_in_data:
if w in word_dict:
words.append(w)
print(len(words))
# add padding
words.insert(0, "<PAD>")
# # add UNK
# words.append("<UNK>")
# save words list to file for mapping questions to a list of indices
save_path = os.path.join(root_path, "embedding/embed_mapping.pkl")
with open(save_path, "wb") as f:
pickle.dump(words, f)
# save the matrix of embedding to file
save_path_2 = os.path.join(root_path, "embedding/embedding_matrix.npy")
embedding_matrix = np.zeros((len(words), 200))
for i in range(1, len(words)):
embedding_matrix[i] = MED[words[i]]
np.save(save_path_2, embedding_matrix)
return words, embedding_matrix
def gen_txt():
with open(q_voc_path, "r") as f:
q_dic = json.load(f)
word_list = []
exc_list = ["<break>", "<END>", "<START>", "<UNKNOWN>", "<UNK>"]
for k, _ in q_dic.items():
# exclude <break>, <END>, <START>, <UNKNOWN>
if (k not in exc_list):
word_list.append(k)
sent_list = []
_, raw_ques, _ = VQADataProvider.load_raw_iqa(q_a_i_path)
for ques in raw_ques:
sent_list.append(VQADataProvider.text_to_list(ques))
# with open(q_a_i_path, "r") as csvfile:
# # QA = csv.reader(csvfile, delimiter="\t", quotechar='\n')
# for row in QA:
# sent_list.append(data_provider.VQADataProvider.seq_to_list(row[2]))
sent_idx_list = []
for sent in sent_list:
sent_idx_list.append(
[word_list.index(x) for x in sent if x not in exc_list])
with open(wordlist_path, "w") as f:
for item in word_list:
f.write("%s\n" % item)
with open(doc_path, "w") as f:
for sent in sent_idx_list:
f.write(" ".join([str(i) for i in sent]) + "\n")
def make_vocab(sentence_ls, vocab_size=-1):
word_fre_dic = {}
for sent in sentence_ls:
word_ls = VQADataProvider.text_to_list(sent)
for word in word_ls:
if (word in word_fre_dic):
word_fre_dic[word] += 1
else:
word_fre_dic[word] = 1
# sort
vocab_ls = [
k for (k, v) in sorted(
word_fre_dic.items(), key=lambda x: x[1], reverse=True)
]
if (vocab_size != -1 and vocab_size <= len(vocab_ls)):
vocab_ls = vocab_ls[:vocab_size]
vocab_ls.reverse()
# create dict with index
vocab_dict = {}
for i in range(len(vocab_ls)):
vocab_dict[vocab_ls[i]] = i
return vocab_dict
def check_len(filename):
q_len_num = {}
a_len_num = {}
exclude_ls = ["<break>", "<START>", "<END>", "<UNKNOWN>"]
# exclude_ls = ["<START>", "<END>"]
with open(filename, "r") as f:
for row in f:
qa = row.rstrip().split("|")
# words_q = qa[1].split()
words_q = VQADataProvider.text_to_list(qa[1])
words_q = [word for word in words_q if word not in exclude_ls]
if (len(words_q) in q_len_num):
q_len_num[len(words_q)] += 1
else:
q_len_num[len(words_q)] = 1
# words_a = qa[2].split()
words_a = VQADataProvider.text_to_list(qa[2])
words_a = [word for word in words_a if word not in exclude_ls]
if (len(words_a) in a_len_num):
a_len_num[len(words_a)] += 1
else:
a_len_num[len(words_a)] = 1
return q_len_num, a_len_num
def check_ans_vocab(filename):
_, _, sentence_ls = VQADataProvider.load_raw_iqa(filename)
word_fre_dict = {}
for sent in sentence_ls:
word_ls = VQADataProvider.text_to_list(sent)
for word in word_ls:
if (word in word_fre_dict):
word_fre_dict[word] += 1
else:
word_fre_dict[word] = 1
# sort
word_fre_dict = sorted(word_fre_dict.items(),
key=lambda kv: kv[1],
reverse=True)
return word_fre_dict
def make_vocab_ans(sentence_ls, vocab_size=-1):
word_fre_dic = {}
for sent in sentence_ls:
word_ls = VQADataProvider.text_to_list(sent)
for word in word_ls:
if (word in word_fre_dic):
word_fre_dic[word] += 1
else:
word_fre_dic[word] = 1
# sort
vocab_ls = [
k for (k, v) in sorted(
word_fre_dic.items(), key=lambda x: x[1], reverse=True)
]
if (vocab_size != -1 and vocab_size <= len(vocab_ls)):
vocab_ls = vocab_ls[:vocab_size - 1]
# add <unknown>
vocab_ls.append("<UNK>")
# # add <ZERO> for padding
# vocab_ls.insert(0, "<ZERO>")
return vocab_ls
def label_img_with_ques_etm():
opt = config.parse_opt()
q_i_a_path = os.path.join(root_path, "data/train/All_QA_Pairs_train.txt")
img_ques_dict = {}
with open(q_i_a_path, "r") as f:
for row in f:
q_i_a = row.strip().split("|")
img = q_i_a[0]
ques = q_i_a[1]
if (img in img_ques_dict):
img_ques_dict[img].append(ques)
else:
img_ques_dict[img] = [ques]
img_topic_dict = {}
for img, qs in img_ques_dict.items():
img_topic_vector = np.zeros(opt.ETM_TOP_NUM)
for q in qs:
words = VQADataProvider.text_to_list(q)
q_t_v = etm_topic_distrib(words)
img_topic_vector = np.add(img_topic_vector, q_t_v)
img_topic_dict[img] = (np.argmax(img_topic_vector)).item()
return img_topic_dict
for row in f:
val_img_ids.append(row.strip())
print("train+val images number: ", len(train_img_ids) + len(val_img_ids))
comb_ids = train_img_ids + val_img_ids
comb_ids = set(comb_ids)
print("unique total images number: ", len(comb_ids))
if __name__ == "__main__":
filename = os.path.join(
root_path,
"data/train_val/QAPairsByCategory/C4_Abnormality_train_val.txt")
_, _, ans_set = load_all2set(filename)
unq_words = set()
for ans in ans_set:
words = VQADataProvider.text_to_list(ans)
for w in words:
unq_words.add(w)
print(len(unq_words))
print(list(unq_words)[:5])
print("<END>" in unq_words)
# filename = "/Users/leishi/Desktop/Internship/vqa2019/ImageClef-2019-VQA-Med-Training/QAPairsByCategory/C4_Abnormality_train.txt"
# all_ans, all_img, all_ques, all_q_i_pairs, all_a_a_pairs = check_raw_data(filename)
# print("\nunique answer length: ", len(all_ans))
# print("\nunique img length: ", len(all_img))
# print("\nunique question length: ", len(all_ques))
# # print("\nall answer")
# # print(all_ans)
# print("\nunique question-image pairs: ", len(all_q_i_pairs))
# print("\nunique question-answer pairs: ", len(all_a_a_pairs))
