def __init__ (self, min_size=1):
super().__init__()
self.gt_matcher = cpp.GTMatcher(FLAGS.match_th, FLAGS.max_masks, min_size)
self.priors = []
if os.path.exists(FLAGS.priors):
with open(FLAGS.priors, 'r') as f:
for l in f:
if l[0] == '#':
continue
s, r = l.strip().split(' ')
s, r = float(s), float(r)
# w * h = s * s
# w / h = r
w = math.sqrt(s * s * r)
h = math.sqrt(s * s / r)
self.priors.append([w, h])
pass
pass
pass
aardvark.print_red("PRIORS %s" % str(self.priors))
# TODO: need a better way to generalize this to multiple priors and 0 priors
self.n_priors = len(self.priors)
if self.n_priors == 0:
self.n_priors = 1
pass
def __init__ (self, min_size=1):
super().__init__()
#self.gt_matcher = cpp.GTMatcher(FLAGS.match_th, FLAGS.max_masks, min_size)
priors = []
if os.path.exists(FLAGS.priors):
priors = load_priors(FLAGS.priors)
if len(priors) == 0:
priors.append([1,1])
aardvark.print_red("PRIORS %s" % str(priors))
# TODO: need a better way to generalize this to multiple priors and 0 priors
self.n_priors = len(priors)
self.priors = priors
pass
def extra_stream_config (self, is_training):
if len(self.priors) > 0:
aardvark.print_red('priors %s' % str(self.priors))
augments = aardvark.load_augments(is_training)
shift = 0
if is_training:
shift = FLAGS.clip_shift
return {
"annotate": [1],
"transforms":
[{"type": "resize", "max_size": FLAGS.max_size}] +
augments + [
#{"type": "clip", "round": FLAGS.backbone_stride},
{"type": "clip", "shift": shift, "width": FLAGS.fix_width, "height": FLAGS.fix_height, "round": FLAGS.clip_stride},
{"type": "anchors.dense.box", 'downsize': FLAGS.anchor_stride, 'lower_th': FLAGS.lower_th, 'upper_th': FLAGS.upper_th, 'weighted': False, 'priors': self.priors, 'params_default': 1.0},
{"type": "box_feature"},
{"type": "rasterize"},
]
}
def __init__(self):
priors = []
# read in priors
# what RPN estimates is the delta between priors and the real
# regression target.
if os.path.exists(FLAGS.rpn_priors):
with open(FLAGS.rpn_priors, 'r') as f:
for l in f:
if l[0] == '#':
continue
vs = [float(v) for v in l.strip().split(' ')]
assert len(vs) == FLAGS.rpn_params
priors.append(vs)
pass
pass
pass
if len(priors) == 0:
priors.append([1.0] * FLAGS.rpn_params)
pass
aardvark.print_red("PRIORS %s" % str(priors))
self.priors = np.array(priors, dtype=np.float32)
pass
def inference (self, images, classes, is_training):
assert FLAGS.clip_stride % FLAGS.backbone_stride == 0
backbone = aardvark.create_stock_slim_network(FLAGS.backbone, images, is_training, global_pool=False, stride=FLAGS.backbone_stride)
if FLAGS.finetune:
backbone = tf.stop_gradient(backbone)
with slim.arg_scope(aardvark.default_argscope(self.is_training)):
if FLAGS.multistep > 0:
if FLAGS.multistep == 1:
aardvark.print_red("multistep = 1 doesn't converge well")
net = slim_multistep_upscale(backbone, FLAGS.backbone_stride, FLAGS.reduction, FLAGS.multistep)
logits = slim.conv2d(net, classes, 3, 1, activation_fn=None, padding='SAME')
else:
logits = slim.conv2d_transpose(backbone, classes, FLAGS.backbone_stride * 2, FLAGS.backbone_stride, activation_fn=None, padding='SAME')
if FLAGS.finetune:
assert FLAGS.colorspace == 'RGB'
def is_trainable (x):
return not x.startswith(FLAGS.backbone)
self.init_session, self.variables_to_train = aardvark.setup_finetune(FLAGS.finetune, is_trainable)
return logits
def __init__(self):
super().__init__()
if FLAGS.classes > 1:
aardvark.print_red("Classes should be number of point classes,")
aardvark.print_red("not counting background. Usually 1.")
pass