def __init__(self, image_dir, label_dir, batch_size, width, height,
no_use_skip_connections, base_filter_size, no_use_batch_norm):
# test data reader
iter_items = data.img_xys_iterator(
image_dir=image_dir,
label_dir=label_dir,
batch_size=batch_size,
patch_width_height=None, # i.e. no patchs
distort_rgb=False,
flip_left_right=False,
random_rotation=False,
repeat=False,
width=width,
height=height,
one_shot=False)
self.iter_init_op, (self.test_imgs, self.test_xys_bitmaps) = iter_items
# build the model
self.model = model.Model(
self.test_imgs,
is_training=False,
use_skip_connections=not no_use_skip_connections,
base_filter_size=base_filter_size,
use_batch_norm=not no_use_batch_norm)
# define loss ops for calculating xent
self.model.calculate_losses_wrt(labels=self.test_xys_bitmaps)
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir)
with open("%s/opts.json" % ckpt_dir, "w") as f:
f.write(json.dumps(vars(opts)))
#from tensorflow.python import debug as tf_debug
#tf.keras.backend.set_session(tf_debug.LocalCLIDebugWrapperSession(tf.Session()))
# Build readers / model for training
# training can be either patch based, or full resolution
train_imgs_xys_bitmaps = data.img_xys_iterator(
image_dir=opts.train_image_dir,
label_dir=opts.label_dir,
batch_size=opts.batch_size,
patch_width_height=opts.patch_width_height,
distort_rgb=True,
flip_left_right=opts.flip_left_right,
random_rotation=opts.random_rotate,
repeat=True,
width=None if opts.patch_width_height else opts.width,
height=None if opts.patch_width_height else opts.height)
# TODO: could we do all these calcs in test.pr_stats (rather than iterating twice) ??
# test images are always full res
test_imgs_xys_bitmaps = data.img_xys_iterator(image_dir=opts.test_image_dir,
label_dir=opts.label_dir,
batch_size=opts.batch_size,
patch_width_height=None,
distort_rgb=False,
flip_left_right=False,
random_rotation=False,
default=None,
help='If set, max number of seconds to run.')
opts = parser.parse_args()
print("opts %s" % opts, file=sys.stderr)
np.set_printoptions(precision=2,
threshold=10000,
suppress=True,
linewidth=10000)
# Build readers for train and test data.
train_imgs, train_xys_bitmaps = data.img_xys_iterator(
image_dir=opts.train_image_dir,
label_dir=opts.label_dir,
batch_size=opts.batch_size,
patch_fraction=opts.patch_fraction,
distort_rgb=True,
flip_left_right=opts.flip_left_right,
random_rotation=opts.random_rotate,
repeat=True)
test_imgs, test_xys_bitmaps = data.img_xys_iterator(
image_dir=opts.test_image_dir,
label_dir=opts.label_dir,
batch_size=1,
patch_fraction=1,
distort_rgb=False,
flip_left_right=False,
random_rotation=False,
repeat=True)
print(test_imgs.get_shape())
print(test_xys_bitmaps.get_shape())
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--image-dir', type=str, required=True)
parser.add_argument('--label-dir', type=str, required=True)
parser.add_argument('--run', type=str, required=True, help='model')
parser.add_argument('--batch-size', type=int, default=1)
parser.add_argument('--no-use-skip-connections', action='store_true')
parser.add_argument('--no-use-batch-norm', action='store_true')
parser.add_argument('--base-filter-size', type=int, default=16)
opts = parser.parse_args()
print(opts)
# test data reader
test_imgs, test_xys_bitmaps = data.img_xys_iterator(image_dir=opts.image_dir,
label_dir=opts.label_dir,
batch_size=opts.batch_size,
patch_fraction=1,
distort_rgb=False,
flip_left_right=False,
random_rotation=False,
repeat=False)
with tf.variable_scope("train_test_model") as scope: # clumsy :/
model = model.Model(test_imgs,
is_training=False,
use_skip_connections=not opts.no_use_skip_connections,
base_filter_size=opts.base_filter_size,
use_batch_norm=not opts.no_use_batch_norm)
model.calculate_losses_wrt(labels=test_xys_bitmaps,
batch_size=opts.batch_size)
sess = tf.Session()
model.restore(sess, "ckpts/%s" % opts.run)
parser.add_argument('--base-filter-size', type=int, default=8)
parser.add_argument('--width', type=int, default=768, help='input image width')
parser.add_argument('--height',
type=int,
default=1024,
help='input image height')
opts = parser.parse_args()
print(opts)
# test data reader
test_imgs, test_xys_bitmaps = data.img_xys_iterator(
image_dir=opts.image_dir,
label_dir=opts.label_dir,
batch_size=opts.batch_size,
patch_width_height=None, # i.e. no patchs
distort_rgb=False,
flip_left_right=False,
random_rotation=False,
repeat=False,
width=opts.width,
height=opts.height)
model = model.Model(test_imgs,
is_training=False,
use_skip_connections=not opts.no_use_skip_connections,
base_filter_size=opts.base_filter_size,
use_batch_norm=not opts.no_use_batch_norm)
model.calculate_losses_wrt(labels=test_xys_bitmaps)
sess = tf.Session()
model.restore(sess, "ckpts/%s" % opts.run)