Bases: NetDataLogger
NetDataLogger for semantic segmentation
initialize SemSegNetDataLogger parameters
Source code in niceml/data/netdataloggers/semsegnetdatalogger.py
| def __init__(self, max_log: int = 10, scale: bool = True):
"""initialize SemSegNetDataLogger parameters"""
super().__init__()
self.scale: bool = scale # If true, the masks are scaled to the image size.
# If false the images are scaled to the mask size.
self.max_log: int = max_log
self.log_count: int = 0
self.mask_colors: List[Tuple[int]] = []
|
Functions
initialize
initialize(data_description, exp_context, set_name)
initialize SemSegNetDataLogger parameters before training
Source code in niceml/data/netdataloggers/semsegnetdatalogger.py
| def initialize(
self,
data_description: OutputImageDataDescription,
exp_context: ExperimentContext,
set_name: str,
):
"""initialize SemSegNetDataLogger parameters before training"""
super().initialize(
data_description=data_description,
exp_context=exp_context,
set_name=set_name,
)
mask_colors = get_color_array(
list(range(self.data_description.get_output_channel_count()))
)
self.mask_colors = [
[int(value * 255) for value in color] for color in mask_colors
]
|
log_data
log_data(net_inputs, net_targets, data_info_list)
Saves as many images with corresponding masks as defined in self.max_log.
The images are saved into self.output_path. For each input image,
the associated masks are added to the image.
Parameters:
-
net_inputs
(ndarray)
–
Input images as np.ndarray
-
net_targets
(ndarray)
–
Target masks as np.ndarray scaled by OUTPUT_IMAGE_SIZE_DIVISOR
-
data_info_list
(List[SemSegDataInfo])
–
Associated data information of input and
Returns:
Source code in niceml/data/netdataloggers/semsegnetdatalogger.py
| def log_data(
self,
net_inputs: np.ndarray,
net_targets: np.ndarray,
data_info_list: List[SemSegDataInfo],
):
"""
Saves as many images with corresponding masks as defined in `self.max_log`.
The images are saved into `self.output_path`. For each input image,
the associated masks are added to the image.
Args:
net_inputs: Input images as `np.ndarray`
net_targets: Target masks as `np.ndarray` scaled by OUTPUT_IMAGE_SIZE_DIVISOR
data_info_list: Associated data information of input and
destination with extended information
Returns:
None
"""
if self.log_count >= self.max_log:
return
for net_input, net_target, data_info in zip(
net_inputs, net_targets, data_info_list
):
instance_labels = [
SemSegInstanceLabel(
class_name=self.data_description.get_output_channel_names()[
class_idx
],
class_index=class_idx,
color=tuple(self.mask_colors[class_idx]),
active=True,
mask=net_target[:, :, class_idx] * 255,
# `draw_error_mask_on_image` doesn't work with binary masks.
# RGB values are required. * 255 converts mask to RGB
)
for class_idx in range(self.data_description.get_output_channel_count())
if net_target[:, :, class_idx].max() > 0
]
if self.scale:
factor = ImageSize(
net_input.shape[1], net_input.shape[0]
).get_division_factor(self.data_description.get_output_image_size())
instance_labels = [
label.scale_label(scale_factor=factor) for label in instance_labels
]
self._draw_image(
image=net_input,
instance_labels=instance_labels,
data_info=data_info,
)
self.log_count += 1
if self.log_count >= self.max_log:
break
|