def get_test_data(img_frontal_path, img_lateral_path, config):
image_frontal = np.zeros([1, config.image_size, config.image_size, 3])
image_frontal[0] = image_utils.getImages(img_frontal_path,
config.image_size)
image_lateral = np.zeros([1, config.image_size, config.image_size, 3])
image_lateral[0] = image_utils.getImages(img_lateral_path,
config.image_size)
sentence = np.zeros(
[1, config.max_sentence_num * config.max_sentence_length])
mask = np.zeros([1, config.max_sentence_num * config.max_sentence_length])
return image_frontal, image_lateral, sentence, mask
def get_inputs(imgs_path,
captions_path,
split_list_path,
vocabulary_path,
image_size=224,
max_caption_length=20):
with open(vocabulary_path, 'r') as file:
vocabulary = json.load(file)
word2id = {}
for i in range(vocabulary.__len__()):
word2id[vocabulary[i]] = i
with open(captions_path, 'r') as file:
caption_dict = json.load(file)
with open(split_list_path, 'r') as file:
split_id_list = json.load(file)
image_list, sentence_list, mask_list, image_id_list = [], [], [], []
for image_id in split_id_list:
filename = str('%012d' % image_id) + '.jpg'
img_file_path = os.path.join(imgs_path, filename)
image = image_utils.getImages(img_file_path, image_size)
image_list.append(image)
caption = caption_dict[str(image_id)][0]
raw_word_list = nltk.word_tokenize(caption)
word_list = []
for word in raw_word_list: # filt out the word not contains in vocabulary
if vocabulary.__contains__(word):
word_list.append(word)
if word_list.__len__() < max_caption_length:
word_list.append('</S>')
for _ in range(max_caption_length - word_list.__len__()):
word_list.append('<EOS>')
else:
word_list = word_list[:max_caption_length]
sentence = np.zeros(shape=[max_caption_length])
mask = np.ones(shape=[max_caption_length])
for i in range(max_caption_length):
sentence[i] = word2id[word_list[i]]
if word_list[i] == '<EOS>':
mask[i] = 0
sentence_list.append(sentence)
mask_list.append(mask)
image_id_list.append([image_id])
return image_list, sentence_list, mask_list, image_id_list
def get_train_tfrecord(imgs_path, data_entry_path, split_list_path, vocabulary_path, tfrecord_path, config):
with open(vocabulary_path, 'r') as f:
vocabulary = json.load(f)
word2id = {}
for i in range(vocabulary.__len__()):
word2id[vocabulary[i]] = i
filenames = os.listdir(imgs_path)
with open(data_entry_path, 'r') as f:
data_dict = json.load(f)
with open(split_list_path, 'r') as f:
split_id_list = json.load(f)
writer = tf.python_io.TFRecordWriter(tfrecord_path)
for id in split_id_list:
two_name = []
for filename in filenames:
if id == filename.split('_')[0]:
two_name.append(filename)
frontal_image_name, lateral_image_name = two_name[0], two_name[1]
if two_name[0] > two_name[1]:
frontal_image_name, lateral_image_name = two_name[1], two_name[0]
image_frontal = image_utils.getImages(os.path.join(imgs_path, frontal_image_name), config.image_size)
image_frontal = image_frontal.reshape([config.image_size*config.image_size*3])
image_lateral = image_utils.getImages(os.path.join(imgs_path, lateral_image_name), config.image_size)
image_lateral = image_lateral.reshape([config.image_size*config.image_size*3])
sent_list = data_dict[id]
if sent_list.__len__() > config.max_sentence_num:
sent_list = sent_list[:config.max_sentence_num]
word_list = []
for sent in sent_list:
words = nltk.word_tokenize(sent)
if words.__len__() >= config.max_sentence_length:
for i in range(config.max_sentence_length - 1):
word_list.append(words[i])
word_list.append('</S>')
else:
for i in range(words.__len__()):
word_list.append(words[i])
word_list.append('</S>')
for _ in range(config.max_sentence_length - words.__len__() - 1):
word_list.append('<EOS>')
for _ in range(config.max_sentence_num - sent_list.__len__()):
word_list.append('</S>')
for _ in range(config.max_sentence_length-1):
word_list.append('<EOS>')
# print(word_list.__len__())
sentence = np.zeros(shape=[config.max_sentence_num * config.max_sentence_length], dtype=np.int64)
mask = np.ones(shape=[config.max_sentence_num * config.max_sentence_length], dtype=np.int64)
for i in range(config.max_sentence_num*config.max_sentence_length):
sentence[i] = word2id[word_list[i]]
if word_list[i] == '<EOS>':
mask[i] = 0
image_id = int(id[3:])
example = tf.train.Example(
features=tf.train.Features(
feature={
'image_frontal_pixels': tf.train.Feature(float_list=tf.train.FloatList(value=image_frontal)),
'image_lateral_pixels': tf.train.Feature(float_list=tf.train.FloatList(value=image_lateral)),
'sentence': tf.train.Feature(int64_list=tf.train.Int64List(value=sentence)),
'mask': tf.train.Feature(int64_list=tf.train.Int64List(value=mask)),
'image_id': tf.train.Feature(int64_list=tf.train.Int64List(value=[image_id]))
}
)
)
serialized = example.SerializeToString()
writer.write(serialized)
print('%s write to tfrecord success!' % tfrecord_path)
def get_train_tfrecord(imgs_path,
data_entry_path,
data_label_path,
split_list_path,
image_size=224,
label_num=15,
mode='train',
D=40):
with open(data_entry_path, 'r') as file:
data_dict = json.load(file)
with open(data_label_path, 'r') as file:
data_label_dict = json.load(file)
finding_labels = data_label_dict.keys()
with open(split_list_path, 'r') as file:
lines = file.readlines()
split_id_list = [line.strip() for line in lines]
Q = 112120 # all image num
Qn = 60361 # all no disease num
for i in range(D):
subsets_num = split_id_list.__len__() / D + 1
sub_split_id_list = split_id_list[i * subsets_num:(i + 1) *
subsets_num]
tfrecord_name = 'data/tfrecord/' + mode + '-%02d.tfrecord' % i
writer = tf.python_io.TFRecordWriter(tfrecord_name)
for image_index in sub_split_id_list:
img_file_path = os.path.join(imgs_path, image_index)
image = image_utils.getImages(img_file_path, image_size)
image_pixels = image.reshape([image_size * image_size * 3])
label_one_hot = np.zeros(shape=[label_num], dtype=np.int64)
labels = data_dict[image_index]
mask_beta = np.zeros(shape=[label_num], dtype=np.float32)
mask_lambda = np.zeros(shape=[label_num], dtype=np.float32)
for label in labels:
for j in range(label_num):
mask_lambda[j] = (0.0 + Q -
data_label_dict[finding_labels[j]]) / Q
if label == finding_labels[j]:
label_one_hot[j] = 1
mask_beta[j] = (0.0 + Qn) / Q
else:
mask_beta[j] = (0.0 + Q - Qn) / Q
example = tf.train.Example(features=tf.train.Features(
feature={
'image_pixels':
tf.train.Feature(float_list=tf.train.FloatList(
value=image_pixels)),
'label_one_hot':
tf.train.Feature(int64_list=tf.train.Int64List(
value=label_one_hot)),
'mask_beta':
tf.train.Feature(float_list=tf.train.FloatList(
value=mask_beta)),
'mask_lambda':
tf.train.Feature(float_list=tf.train.FloatList(
value=mask_lambda))
}))
serialized = example.SerializeToString()
writer.write(serialized)
print('%s write to tfrecord success!' % tfrecord_name)
image_path = 'data/CXR3_IM-1384-1001.png'
FLAGS = tf.app.flags.FLAGS
tf.flags.DEFINE_string('img', 'data/examples/CXR3_IM-1384-1001.png', 'The test image path')
model_path = 'data/model/my-test-68000'
data_label_path = 'data/data_label.json'
md = Model(is_training=False, batch_size=1)
with open(data_label_path, 'r') as file:
labels = json.load(file).keys()
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver.restore(sess, model_path)
images = np.zeros([md.batch_size, md.image_size, md.image_size, 3])
images[0] = image_utils.getImages(FLAGS.img, md.image_size)
paddings = np.zeros([md.batch_size, md.label_num])
feed_dict = {md.images: images,
md.labels: paddings,
md.mask_beta: paddings,
md.mask_lambda: paddings}
predictions = sess.run(md.predictions, feed_dict=feed_dict)
for j in range(md.label_num):
print('disease \'%s\', prob = %s' % (labels[j], round(predictions[0][j], 5)))
def get_train_tfrecord(imgs_path,
captions_path,
split_list_path,
vocabulary_path,
image_size=224,
max_caption_length=20,
mode='train',
D=40):
with open(vocabulary_path, 'r') as file:
vocabulary = json.load(file)
word2id = {}
for i in range(vocabulary.__len__()):
word2id[vocabulary[i]] = i
with open(captions_path, 'r') as file:
caption_dict = json.load(file)
with open(split_list_path, 'r') as file:
split_id_list = json.load(file)
for i in range(D):
subsets_num = split_id_list.__len__() / D + 1
sub_split_id_list = split_id_list[i * subsets_num:(i + 1) *
subsets_num]
tfrecord_name = 'data/tfrecord/' + mode + '-%02d.tfrecord' % i
writer = tf.python_io.TFRecordWriter(tfrecord_name)
for image_id in sub_split_id_list:
filename = str('%012d' % image_id) + '.jpg'
img_file_path = os.path.join(imgs_path, filename)
image = image_utils.getImages(img_file_path, image_size)
image = image.reshape([image_size * image_size * 3])
caption = caption_dict[str(image_id)][0]
raw_word_list = nltk.word_tokenize(caption)
word_list = []
for word in raw_word_list: # filt out the word not contains in vocabulary
if vocabulary.__contains__(word):
word_list.append(word)
if word_list.__len__() < max_caption_length:
word_list.append('</S>')
for _ in range(max_caption_length - word_list.__len__()):
word_list.append('<EOS>')
else:
word_list = word_list[:max_caption_length]
sentence = np.zeros(shape=[max_caption_length], dtype=np.int64)
mask = np.ones(shape=[max_caption_length], dtype=np.int64)
for i in range(max_caption_length):
sentence[i] = word2id[word_list[i]]
if word_list[i] == '<EOS>':
mask[i] = 0
# image_raw = image.tostring()
example = tf.train.Example(features=tf.train.Features(
feature={
'image_pixels':
tf.train.Feature(float_list=tf.train.FloatList(
value=image)),
'sentence':
tf.train.Feature(int64_list=tf.train.Int64List(
value=sentence)),
'mask':
tf.train.Feature(int64_list=tf.train.Int64List(
value=mask)),
'image_id':
tf.train.Feature(int64_list=tf.train.Int64List(
value=[image_id]))
}))
serialized = example.SerializeToString()
writer.write(serialized)
# print('%s' % filename)
print('%s write to tfrecord success!' % tfrecord_name)