def __init__(self, base, linthresh, linscale):
Transform.__init__(self)
self.base = base
self.linthresh = linthresh
self.linscale = linscale
self._linscale_adj = (linscale / (1.0 - self.base ** -1))
self._log_base = np.log(base)
def __init__(self, base, linthresh):
Transform.__init__(self)
self.base = base
self.linthresh = linthresh
log_base = np.log(base)
logb_linthresh = np.log(linthresh) / log_base
self._linadjust = 1.0 - logb_linthresh
def prepTransformation(self):
"""
Setup of the view and model matrices are done
"""
BufferHelper.sendUniformToShaders('projection', self.projection_mat,
'm4')
# setup view matrix
view_mat = np.array(
Transform.lookat(Global.EYE, Global.LOOKAT, Global.UP))
BufferHelper.sendUniformToShaders('view', view_mat, 'm4')
# setup model matrix
model_mat = np.matrix(np.identity(4, dtype=np.float32))
# note that because GLSL uses row major, need to do SRT (instead of TRS)
# scale
model_mat *= Transform.scale(self.scale, self.scale, self.scale)
# rotations
model_mat *= self.rotationMatrix
# transform
model_mat *= self.translationMatrix
BufferHelper.sendUniformToShaders('model', model_mat, 'm4')
def __init__(self, base, linthresh):
Transform.__init__(self)
self.base = base
self.linthresh = linthresh
self._log_base = np.log(base)
self._log_linthresh = np.log(linthresh) / self._log_base
self._linadjust = linthresh / (np.log(linthresh) / self._log_base)
def __init__(self, base, linthresh):
Transform.__init__(self)
self.base = base
self.linthresh = linthresh
self._log_base = np.log(base)
self._log_linthresh = np.log(linthresh) / self._log_base
self._linadjust = linthresh / (np.log(linthresh) / self._log_base)
def __init__(self, base, nonpos):
Transform.__init__(self)
self.base = base
if nonpos == 'mask':
self._handle_nonpos = _mask_non_positives
else:
self._handle_nonpos = _clip_non_positives
def __init__(self, base, linthresh):
Transform.__init__(self)
self.base = base
self.linthresh = linthresh
log_base = np.log(base)
logb_linthresh = np.log(linthresh) / log_base
self._linadjust = 1.0 - logb_linthresh
def __init__(self, base, nonpos):
Transform.__init__(self)
self.base = base
if nonpos == 'mask':
self._handle_nonpos = _mask_non_positives
else:
self._handle_nonpos = _clip_non_positives
def __init__(self, base, linthresh, linscale):
Transform.__init__(self)
symlog = SymmetricalLogScale.SymmetricalLogTransform(base, linthresh, linscale)
self.base = base
self.linthresh = linthresh
self.invlinthresh = symlog.transform(linthresh)
self.linscale = linscale
self._linscale_adj = (linscale / (1.0 - self.base ** -1))
def __init__(self, base, linthresh):
Transform.__init__(self)
self.base = base
self.linthresh = abs(linthresh)
self._log_base = np.log(base)
logb_linthresh = np.log(linthresh) / self._log_base
self._linadjust = 1.0 - logb_linthresh
self._linscale = 1.0 / linthresh
def __init__(self, base, linthresh):
Transform.__init__(self)
self.base = base
self.linthresh = abs(linthresh)
self._log_base = np.log(base)
logb_linthresh = np.log(linthresh) / self._log_base
self._linadjust = 1.0 - logb_linthresh
self._linscale = 1.0 / linthresh
def __init__(self, base, linthresh):
Transform.__init__(self)
self.base = base
self.linthresh = abs(linthresh)
self._log_base = np.log(base)
self._logb_linthresh = np.log(linthresh) / self._log_base
self._logb_minlog = np.floor(self._logb_linthresh - 1e-10)
self._linadjust = np.abs((np.log(linthresh) - self._logb_minlog) /
linthresh)
def __init__(self, base, linthresh):
Transform.__init__(self)
self.base = base
self.linthresh = abs(linthresh)
self._log_base = np.log(base)
self._logb_linthresh = np.log(linthresh) / self._log_base
self._logb_minlog = np.floor(self._logb_linthresh - 1e-10)
self._linadjust = np.abs((np.log(linthresh) - self._logb_minlog) /
linthresh)
def glutMouseMoveEvent(self, x, y):
if hasattr(self, 'ctrlDown') and self.ctrlDown:
self.translationMatrix *= Transform.translate(
(self.xorigpos - x) * -0.01, (self.yorigpos - y) * 0.01, 0)
else:
self.rotationMatrix *= Transform.yrotate(
(self.xorigpos - x) * -0.01)
self.rotationMatrix *= Transform.xrotate(
(self.yorigpos - y) * -0.01)
self.xorigpos, self.yorigpos = x, y
def __init__(self, *args, **kwargs):
"""Create a Vega Line Chart"""
super(Line, self).__init__(*args, **kwargs)
#Line Updates
self.scales['x'].type = 'linear'
self.scales['y'].type = 'linear'
if self._is_datetime:
self.scales['x'].type = 'time'
self.scales['color'] = Scale(name='color', type='ordinal',
domain=DataRef(data='table', field='data.col'),
range='category20')
del self.marks[0]
transform = MarkRef(data='table',
transform=[Transform(type='facet', keys=['data.col'])])
enter_props = PropertySet(x=ValueRef(scale='x', field="data.idx"),
y=ValueRef(scale='y', field="data.val"),
stroke=ValueRef(scale="color", field='data.col'),
stroke_width=ValueRef(value=2))
new_mark = Mark(type='group', from_=transform,
marks=[Mark(type='line',
properties=MarkProperties(enter=enter_props))])
self.marks.append(new_mark)
def __init__(self, *args, **kwargs):
"""Create a Vega Stacked Area Chart"""
super(StackedArea, self).__init__(*args, **kwargs)
facets = Transform(type='facet', keys=['data.idx'])
stats = Transform(type='stats', value='data.val')
stat_dat = Data(name='stats', source='table', transform=[facets, stats])
self.data['stats'] = stat_dat
self.scales['x'].zero = False
self.scales['y'].domain = DataRef(field='sum', data='stats')
stackit = Transform(type='stack', point='data.idx', height='data.val')
self.marks[0].from_.transform.append(stackit)
self.marks[0].marks[0].properties.enter.y.scale = 'y'
self.marks[0].marks[0].properties.enter.y.field = 'y'
del self.marks[0].marks[0].properties.enter.y2.value
self.marks[0].marks[0].properties.enter.y2.field = 'y2'
def __getitem__(self, item):
# type: (int) -> Tuple(Tensor, Tensor)
image = Image.open(self.images_list[item]).convert('RGB')
mask = np.load(self.masks_list[item])
mask = Image.fromarray(mask)
#Apply data augmentation in TRAIN mode
augment = self.mode == DSMode.TRAIN
image, mask = Transform(flip_prob=0.5,
degrees=10,
minscale=0.6,
augment=augment)(image, mask)
return image, mask
def reshape(self, width, height):
"""
GLUT reshape function with the addition of defining the projection matrix
Parameters
----------
width : int
height : int
Note
----
You may need to call this class if is not automatically called by the GLUT function:
If it is not called, the projection frustum is not created.
"""
gl.glViewport(0, 0, width, height)
self.projection_mat = Transform.perspective(Global.FOVY,
width / height,
Global.ZNEAR, Global.ZFAR)
def transform_file(filename, output="out.f90",
dim=1, interleave=False, compress=True, row_major=False):
"""Apply tfortran source code transformations.
:param filename: input file name
:param dim: dimension
:param output: output file name
"""
template = None
with open(filename, 'r') as f:
template = f.read()
transform = Transform()
transform.dim = dim
transform.compress = compress
transform.interleave = interleave
transform.row_major = row_major
transformed = transform.transform(template)
with open(output, 'w') as f:
f.write(transformed)
def __init__(self, base):
Transform.__init__(self)
self.base = base
def __init__(self, net, cfg):
self.cfg = cfg
self.net = net
self.anchors = generate_anchors(cfg)
if torch.cuda.is_available():
self.net.cuda()
self.anchors = self.anchors.cuda()
# Dataset transform
transform = [
Transform(context_amount=cfg.TRAIN.CROP_CONTEXT_AMOUNT_Z, size=cfg.MODEL.Z_SIZE),
Transform(context_amount=cfg.TRAIN.CROP_CONTEXT_AMOUNT_X, size=cfg.MODEL.X_SIZE,
random_translate=True, random_resize=True, motion_blur=True,
random_translate_range=cfg.TRAIN.DATA_AUG_TRANSLATE_RANGE,
random_resize_scale_min=cfg.TRAIN.DATA_AUG_RESIZE_SCALE_MIN,
random_resize_scale_max=cfg.TRAIN.DATA_AUG_RESIZE_SCALE_MAX
)
]
# Training dataset
trackingnet = TrackingNet(cfg.PATH.TRACKINGNET, subset="train", debug_seq=cfg.TRAIN.DEBUG_SEQ)
imagenet = ImageNetVID(cfg.PATH.ILSVRC, subset="train")
sampler = PairSampler([trackingnet, imagenet], cfg=cfg, transform=transform, pairs_per_video=cfg.TRAIN.PAIRS_PER_VIDEO,
frame_range=cfg.TRAIN.FRAME_RANGE)
# Distractor dataset
coco = CocoDetection(cfg.PATH.COCO, cfg.PATH.COCO_ANN_FILE)
# coco_distractor = COCODistractor(coco, 4000)
coco_positive = COCOPositivePair(coco, 4000, cfg=cfg, transform=transform)
coco_negative = COCONegativePair(coco, 12000, cfg=cfg, transform=transform)
dataset = ConcatDataset([sampler, coco_positive, coco_negative])
self.dataloader = DataLoader(dataset, batch_size=cfg.TRAIN.BATCH_SIZE, num_workers=4, shuffle=True,
pin_memory=True, drop_last=True)
# Validation dataset
val_trackingnet = TrackingNet(cfg.PATH.TRACKINGNET, subset="val")
val_imagenet = ImageNetVID(cfg.PATH.ILSVRC, subset="val")
validation_sampler = PairSampler([val_trackingnet, val_imagenet], cfg=cfg, transform=transform,
pairs_per_video=1, frame_range=cfg.TRAIN.FRAME_RANGE)
val_coco_positive = COCOPositivePair(coco, 100, cfg=cfg, transform=transform)
val_dataset = ConcatDataset([validation_sampler, val_coco_positive])
if cfg.TRAIN.DEBUG_SEQ >= 0: # When debugging on a single sequence, the validation is performed on the same one
val_dataset = PairSampler([trackingnet], cfg=cfg, transform=transform, pairs_per_video=200)
self.validation_dataloader = DataLoader(val_dataset, batch_size=min(cfg.TRAIN.BATCH_SIZE, 20), num_workers=4,
shuffle=True, pin_memory=True, drop_last=False)
# Loss
self.criterion = MultiBoxLoss(self.anchors, cfg)
self.optimizer = optim.Adam(self.net.parameters(), lr=cfg.TRAIN.LR, weight_decay=cfg.TRAIN.WEIGHT_DECAY)
self.scheduler = optim.lr_scheduler.StepLR(self.optimizer, step_size=cfg.TRAIN.SCHEDULER_STEP_SIZE,
gamma=cfg.TRAIN.SCHEDULER_GAMMA)
# Summary Writer
self.run_id = datetime.now().strftime('%b%d_%H-%M-%S')
if not cfg.DEBUG:
self.save_config()
self.save_code()
self.writer = SummaryWriter(log_dir=os.path.join(cfg.PATH.DATA_DIR, "runs", self.run_id))
self.start_epoch = 0
if cfg.TRAIN.RESUME_CHECKPOINT:
self.start_epoch = utils.load_checkpoint(cfg.TRAIN.RESUME_CHECKPOINT, self.net, self.optimizer)
if torch.cuda.is_available():
self.net = nn.DataParallel(self.net)
self.best_IOU = 0.
def __init__(self, base):
Transform.__init__(self)
self.base = base
