def build(self):
"""
Build Retina Net architecture.
"""
# Image size must be dividable by 2 multiple times.
h, w = self.cf.patch_size[:2]
if h / 2 ** 5 != int(h / 2 ** 5) or w / 2 ** 5 != int(w / 2 ** 5):
raise Exception("Image size must be dividable by 2 at least 5 times "
"to avoid fractions when downscaling and upscaling."
"For example, use 256, 320, 384, 448, 512, ... etc. ")
# instanciate abstract multi dimensional conv class and backbone model.
conv = mutils.NDConvGenerator(self.cf.dim)
backbone = utils.import_module('bbone', self.cf.backbone_path)
# build Anchors, FPN, Classifier / Bbox-Regressor -head
self.np_anchors = mutils.generate_pyramid_anchors(self.logger, self.cf)
self.anchors = torch.from_numpy(self.np_anchors).float().cuda()
# self.Fpn = backbone.FPN(self.cf, conv, operate_stride1=self.cf.operate_stride1)
self.num_classes=self.cf.head_classes
# i3d_rgb = backbone.I3D(num_classes=self.num_classes)
######## HEART OF I3D RGB Weight Transfer
# print("Transferring weight from i3d_rgb model...........")
# i3d_rgb = backbone.I3D(num_classes=self.num_classes)
# i3d_rgb.load_state_dict(torch.load('weight/model_rgb.pth'), strict=False)
# print('Success..........')
# self.Fpn = i3d_rgb
self.Fpn = backbone.I3D(num_classes=self.num_classes)
self.Classifier = Classifier(self.cf, conv)
self.BBRegressor = BBRegressor(self.cf, conv)
def build(self):
"""
Build Retina Net architecture.
"""
# Image size must be dividable by 2 multiple times.
h, w = self.cf.patch_size[:2]
if h / 2**5 != int(h / 2**5) or w / 2**5 != int(w / 2**5):
raise Exception(
"Image size must be dividable by 2 at least 5 times "
"to avoid fractions when downscaling and upscaling."
"For example, use 256, 320, 384, 448, 512, ... etc. ")
# instanciate abstract multi dimensional conv class and backbone model.
conv = mutils.NDConvGenerator(self.cf.dim)
backbone = utils.import_module('bbone', self.cf.backbone_path)
# build Anchors, FPN, Classifier / Bbox-Regressor -head
self.np_anchors = mutils.generate_pyramid_anchors(self.logger, self.cf)
self.anchors = torch.from_numpy(self.np_anchors).float().cuda()
self.Fpn = backbone.FPN(self.cf,
conv,
operate_stride1=self.cf.operate_stride1)
self.Classifier = Classifier(self.cf, conv)
self.BBRegressor = BBRegressor(self.cf, conv)
self.final_conv = conv(self.cf.end_filts,
self.cf.num_seg_classes,
ks=1,
pad=0,
norm=self.cf.norm,
relu=None)
def __init__(self, cf, logger):
super(net, self).__init__()
self.cf = cf
self.logger = logger
backbone = utils.import_module('bbone', cf.backbone_path)
conv = mutils.NDConvGenerator(cf.dim)
# set operate_stride1=True to generate a unet-like FPN.)
self.fpn = backbone.FPN(cf, conv, operate_stride1=True).cuda()
self.conv_final = conv(cf.end_filts, cf.num_seg_classes, ks=1, pad=0, norm=cf.norm, relu=None)
if self.cf.weight_init is not None:
logger.info("using pytorch weight init of type {}".format(self.cf.weight_init))
mutils.initialize_weights(self)
else:
logger.info("using default pytorch weight init")
