Utilities¶

The dh_segment.utils module contains the parameters for config with sacred package, image label vizualization functions and miscelleanous helpers.

Parameters¶

`ModelParams`(**kwargs)	Parameters related to the model
`TrainingParams`(**kwargs)	Parameters to configure training process

Label image helpers¶

`label_image_to_class`(label_image, classes_file)	rtype tensorflow.Tensor
`class_to_label_image`(class_label, classes_file)	rtype tensorflow.Tensor
`multilabel_image_to_class`(label_image, …)	Combines image annotations with classes info of the txt file to create the input label for the training.
`multiclass_to_label_image`(…)	rtype tensorflow.Tensor
`get_classes_color_from_file`(classes_file)	rtype `ndarray`
`get_n_classes_from_file`(classes_file)	rtype `int`
`get_classes_color_from_file_multilabel`(…)	Get classes and code labels from txt file.
`get_n_classes_from_file_multilabel`(classes_file)	rtype `int`

Evaluation utils¶

`Metrics`()
`intersection_over_union`(cnt1, cnt2, shape_mask)

Miscellaneous helpers¶

`parse_json`(filename)
`dump_json`(filename, dict)
`load_pickle`(filename)
`dump_pickle`(filename, obj)
`hash_dict`(params)

class dh_segment.utils.PredictionType¶

Variables

CLASSIFICATION –
REGRESSION –
MULTILABEL –

CLASSIFICATION = 'CLASSIFICATION'¶

MULTILABEL = 'MULTILABEL'¶

REGRESSION = 'REGRESSION'¶

classmethod parse(prediction_type)¶

class dh_segment.utils.VGG16ModelParams¶

CORRECTED_VERSION = None¶

INTERMEDIATE_CONV = [[(256, 3)]]¶

PRETRAINED_MODEL_FILE = 'pretrained_models/vgg_16.ckpt'¶

SELECTED_LAYERS_UPSCALING = [True, True, True, True, False, False]¶

UPSCALE_PARAMS = [[(32, 3)], [(64, 3)], [(128, 3)], [(256, 3)], [(512, 3)], [(512, 3)]]¶

class dh_segment.utils.ResNetModelParams¶

CORRECT_VERSION = False¶

INTERMEDIATE_CONV = None¶

PRETRAINED_MODEL_FILE = 'pretrained_models/resnet_v1_50.ckpt'¶

SELECTED_LAYERS_UPSCALING = [True, True, True, True, True]¶

UPSCALE_PARAMS = [(32, 0), (64, 0), (128, 0), (256, 0), (512, 0)]¶

class dh_segment.utils.UNetModelParams¶

CORRECT_VERSION = False¶

INTERMEDIATE_CONV = None¶

PRETRAINED_MODEL_FILE = None¶

SELECTED_LAYERS_UPSCALING = None¶

UPSCALE_PARAMS = None¶

class dh_segment.utils.ModelParams(**kwargs)¶

Parameters related to the model

check_params()¶

class dh_segment.utils.TrainingParams(**kwargs)¶

Parameters to configure training process

Variables

n_epochs (int) – number of epoch for training
evaluate_every_epoch (int) – the model will be evaluated every n epochs
learning_rate (float) – the starting learning rate value
exponential_learning (bool) – option to use exponential learning rate
batch_size (int) – size of batch
data_augmentation (bool) – option to use data augmentation (by default is set to False)
data_augmentation_flip_lr (bool) – option to use image flipping in right-left direction
data_augmentation_flip_ud (bool) – option to use image flipping in up down direction
data_augmentation_color (bool) – option to use data augmentation with color
data_augmentation_max_rotation (float) – maximum angle of rotation (in radians) for data augmentation
data_augmentation_max_scaling (float) – maximum scale of zooming during data augmentation (range: [0,1])
make_patches (bool) – option to crop image into patches. This will cut the entire image in several patches
patch_shape (tuple) – shape of the patches
input_resized_size (int) – size (in pixel) of the image after resizing. The original ratio is kept. If no resizing is wanted, set it to -1
weights_labels (list) – weight given to each label. Should be a list of length = number of classes
training_margin (int) – size of the margin to add to the images. This is particularly useful when training with patches
local_entropy_ratio (float) –
local_entropy_sigma (float) –
focal_loss_gamma (float) – value of gamma for the focal loss. See paper : https://arxiv.org/abs/1708.02002

check_params()¶

Checks if there is no parameter inconsistency

Return type: None

dh_segment.utils.label_image_to_class(label_image, classes_file)¶

Return type: tensorflow.Tensor

dh_segment.utils.class_to_label_image(class_label, classes_file)¶

Return type: tensorflow.Tensor

dh_segment.utils.multilabel_image_to_class(label_image, classes_file)¶

Combines image annotations with classes info of the txt file to create the input label for the training.

Parameters

label_image (tensorflow.Tensor) – annotated image [H,W,Ch] or [B,H,W,Ch] (Ch = color channels)
classes_file (str) – the filename of the txt file containing the class info

Return type

tensorflow.Tensor

Returns

[H,W,Cl] or [B,H,W,Cl] (Cl = number of classes)

dh_segment.utils.multiclass_to_label_image(class_label_tensor, classes_file)¶

Return type: tensorflow.Tensor

dh_segment.utils.get_classes_color_from_file(classes_file)¶

Return type: ndarray

dh_segment.utils.get_n_classes_from_file(classes_file)¶

Return type: int

dh_segment.utils.get_classes_color_from_file_multilabel(classes_file)¶

Get classes and code labels from txt file. This function deals with the case of elements with multiple labels.

Parameters: classes_file (str) – file containing the classes (usually named classes.txt)
Return type: Tuple[ndarray, <built-in function array>]
Returns: for each class the RGB color (array size [N, 3]); and the label’s code (array size [N, C]), with N the number of combinations and C the number of classes

dh_segment.utils.get_n_classes_from_file_multilabel(classes_file)¶

Return type: int

dh_segment.utils.parse_json(filename)¶

dh_segment.utils.dump_json(filename, dict)¶

dh_segment.utils.load_pickle(filename)¶

dh_segment.utils.dump_pickle(filename, obj)¶

dh_segment.utils.hash_dict(params)¶

class dh_segment.utils.Metrics¶

compute_accuracy()¶

compute_iu()¶

compute_miou()¶

compute_mse()¶

compute_prf(beta=1)¶

compute_psnr()¶

save_to_json(json_filename)¶

Return type: None

dh_segment.utils.intersection_over_union(cnt1, cnt2, shape_mask)¶