def create_dataset (data_dir, num_parallel_calls = 4, patch_size = [64, 64, 64],
overlap = 32, image_shape = [236 , 320, 260], start = [0, 0, 50],
num_imgaes_loaded = 32, batch_size = 64, prefetched_buffer_size = 8000):
NUM_CLASSES = 3
# choose & fetch all required data / discard subjects missing crucial data
list_images = parse_tf.fetch_paths(data_dir, '_F_')
print(list_images)
list_labels = [parse_label(x) for x in list_images]
# print(list_labels)
num_samples = len(list_images)
print('num_samples ', num_samples)
# define the patch you want to crop
patch_size = patch_size # size of the patches per axis
overlap = overlap
image_shape = image_shape
start = start
# number of patches to extract from each image
num_patches = len(patches.compute_patch_indices(image_shape=image_shape,
patch_size=patch_size,
overlap=overlap,
start=start))
print('number of patches cropped per image: ', num_patches)
buffer_size = num_imgaes_loaded * num_patches # shuffle patches from 34 different big images
# This function splits a whole image into many splits and return
get_patches_fn = lambda image: patches.get_patches(image,
num_patches=num_patches,
image_shape=image_shape,
patch_size=patch_size,
overlap=overlap,
start=start)
# generate placeholders that receive paths of type str
images = tf.placeholder(tf.string, shape=[None])
labels = tf.placeholder(tf.int32, shape=[None])
# create dataset for your needs
dataset_image = tf.data.TFRecordDataset(images)
dataset_labels = tf.data.Dataset.from_tensor_slices(labels)
dataset = tf.data.Dataset.zip((dataset_image, dataset_labels))
# Note: this could be rewritten as one map call
# map parse function to each zipped element
dataset = dataset.map(
map_func=lambda a, b: (convert_tf.parse_function(a),b),
num_parallel_calls=num_parallel_calls)
# This is for single channel, extand the last axis as the channel axis
dataset = dataset.map(
map_func=lambda a, b: (tf.expand_dims(a, -1),
tf.one_hot(b, NUM_CLASSES),),
num_parallel_calls=num_parallel_calls)
dataset = dataset.map(map_func=lambda a, b: (get_patches_fn(a),
[b for _ in range(num_patches)]),
num_parallel_calls=num_parallel_calls)
dataset = dataset.apply(tf.contrib.data.unbatch())
# here one could use shuffle, repeat, prefetch, ...
dataset = dataset.shuffle(buffer_size=buffer_size)
dataset_batched = dataset.batch(batch_size=batch_size)
dataset_batched = dataset_batched.prefetch(buffer_size = prefetched_buffer_size)
dataset_batched = dataset_batched.repeat()
iterator =dataset_batched.make_initializable_iterator()
next_element = iterator.get_next()
# dummy "model"
result = next_element
with tf.Session() as sess:
sess.run(iterator.initializer,
feed_dict={images: list_images, labels: list_labels})
while True:
try:
yield sess.run(result)
except tf.errors.OutOfRangeError:
print('finished')
break
elif path_name.find('fb') != -1:
return 1
if __name__ == '__main__':
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
num_parallel_calls = 4
# define some example values
data_dir = '/home/d1274/no_backup/d1274/data'
b_viewer = True
b_verbose = True
# choose & fetch all required data / discard subjects missing crucial data
list_images = parse_tf.fetch_paths(data_dir, '_F_')
print(list_images)
list_labels = [parse_label(x) for x in list_images]
num_samples = len(list_images)
print('num_samples ',num_samples)
# define the patch you want to crop
patch_size = [128, 128, 128] # size of the patches per axis
overlap = 32
image_shape = [236, 320 , 260]
start = [0 , 0, 50]
# number of patches to extract from each image
num_patches = len(patches.compute_patch_indices(image_shape = image_shape,
patch_size = patch_size,