core package¶

core.yolov4 module¶

core.yolov4.YOLO(input_layer, NUM_CLASS, model='yolov4')¶

This function calls either YOLOv4 or YOLOv3

Parameters

input_layer (object) – Tensor
NUM_CLASS (int) – How many classes there are (80 for coco)
model (str) – States if yolov3 or yolov4 should be run

Returns

List of conv layers at differnt stages in the CNN

Return type

list[objects]

core.yolov4.YOLOv3(input_layer, NUM_CLASS)¶

The Yolov3 function

Parameters

input_layer (object) – Tensor
NUM_CLASS (int) – How many classes there are (80 for coco)

Returns

List of conv layers at differnt stages in the CNN

Return type

list[objects]

core.yolov4.YOLOv4(input_layer, NUM_CLASS)¶

The Yolov4 function

Parameters

input_layer (object) – Tensor
NUM_CLASS (int) – How many classes there are (80 for coco)

Returns

List of conv layers at differnt stages in the CNN

Return type

list[objects]

core.yolov4.compute_loss(pred, conv, label, bboxes, STRIDES, NUM_CLASS, IOU_LOSS_THRESH, i=0)¶

Computes the losses of the prediction using the iou functions

Parameters

pred (object) – The prediction itself
conv (object) – The conv layer
label (object) – The labels for the corresponding prediction
bboxes (object) – The bounding boxes for use in calculating the iou
STRIDES (ndarray) – How many pixels to move the neural network’s filter
NUM_CLASS (int) – FILLER
IOU_LOSS_THRESH (float) – The IOU threshold amount
i (int) – Keeps track of which stride to be using

Returns

int – giou_loss
float – conf_loss
object – prob_loss

core.yolov4.decode(conv_output, output_size, NUM_CLASS, STRIDES, ANCHORS, i=0, XYSCALE=[1, 1, 1])¶

Function to decode the incoming Tensor

Parameters

conv_output (object) – Tensor in training
output_size (int) – What size the output should be
NUM_CLASS (int) – How many classes there are (80 for coco)
STRIDES (ndarray) – How many pixels to move the neural network’s filter
ANCHORS (List[int]) – List of all the anchors for the training of the NN
i (int) – Keeps track of which anchor, stride and XYSCALE to be using
XYSCALE (list[floats]) – How much to scale the x and y coords of the data.

Returns

object – Prediction of variables
object – Prediction probabilities

core.yolov4.decode_train(conv_output, output_size, NUM_CLASS, STRIDES, ANCHORS, i=0, XYSCALE=[1, 1, 1])¶

Function to decode the incoming Tensor for training new models

Parameters

conv_output (object) – Tensor in training
output_size (int) – What size the output should be
NUM_CLASS (int) – How many classes there are (80 for coco)
STRIDES (int) – How many pixels to move the neural network’s filter
ANCHORS (List[int]) – List of all the anchors for the training of the NN
i (int) – Keeps track of which anchor, stride and XYSCALE to be using
XYSCALE (list) – How much to scale the x and y coords of the data.

Returns

Concatenated list of predictions and confidence scores

Return type

List[objects]

core.yolov4.filter_boxes(box_xywh, scores, score_threshold=0.4, input_shape=tensorflow.constant)¶

Filter out the bounding boxes that are below the score threshold

Parameters

box_xywh (object) – The box coords
scores (object) – The scores for the bounding boxes
score_threshold (float) – The constant for which is the threshold for what is allowed/not
input_shape (list[int]) – list of dimensions of the input shape

Returns

tuple of the boxes along with the confidence level

Return type

tuple

core.backbone module¶

core.backbone.cspdarknet53(input_data)¶

This function represents the backbone CNN for cspdarknet53 for YOLOV4

Parameters

input_data (object) – Tensor

Returns

object – Returns the tensor after it has run through some of the layers
object – Returns the tensor after it has run through most the layers
object – Returns the final tensor after it has run through all the layers

core.backbone.darknet53(input_data)¶

This function represents the backbone CNN for darknet53 for YOLOV3

Parameters

input_data (object) – Tensor

Returns

object – Returns the tensor after it has run through some of the layers
object – Returns the tensor after it has run through most the layers
object – Returns the final tensor after it has run through all the layers

core.common module¶

class core.common.BatchNormalization(*args: Any, **kwargs: Any)¶

Bases: tensorflow.keras.layers.tensorflow.keras.layers.BatchNormalization._name

“Frozen state” and “inference mode” are two separate concepts. layer.trainable = False is to freeze the layer, so the layer will use stored moving var and mean in the “inference mode”, and both gama and beta will not be updated !

call(x, training=False)¶

core.common.convolutional(input_layer, filters_shape, downsample=False, activate=True, bn=True, activate_type='leaky')¶

Applies convultions to input layers

Parameters

input_layer (object) – A conv layer
filters_shape (tuple) – A tuple of the filter shapes
downsample (bool) – If True then downsamples the layer
activate (bool) – If True then apply an activation function to the layer
bn (bool) – If True then apply BatchNormalization to layer
activate_type (str) – What activation function should be applied

Returns

A convolution

Return type

object

core.common.mish(x)¶

Applies the mish activation function to x

Parameters: x (object) – A conv layer ready to be activated
Returns: Activated conv layer
Return type: object

core.common.residual_block(input_layer, input_channel, filter_num1, filter_num2, activate_type='leaky')¶

Takes input and combines it with a copy of itself that has been applied to the convolutional function twice

Parameters

input_layer (object) – A conv layer
input_channel (int) – How many input channels
filter_num1 (int) – Filter1 for conv2D within convolutional function
filter_num2 (int) – Filter2 for conv2D within convolutional function
activate_type (str) – What activation function should be applied

Returns

Conv layer

Return type

object

core.common.route_group(input_layer, groups, group_id)¶

core.common.upsample(input_layer)¶

core.dataset module¶

class core.dataset.Dataset(FLAGS, is_training: bool, dataset_type: str = 'converted_coco')¶

Bases: object

implement Dataset here

load_annotations()¶

loads the annotations of the dataset

Returns: A list of the annotations within the dataset
Return type: list

parse_annotation(annotation)¶

Performs parsing of the annotations of the dataset

Parameters

annotation (str) – A single annotation

Returns

ndarray – The image itself
ndarray – The bbox of the annotation

Raises

KeyError – Raises a KeyError if the path does not exist

preprocess_true_boxes(bboxes)¶

Processes the ground truth bounding boxes

Parameters: bboxes (ndarray) – The bounding boxes of the images
Returns: The processed bounding box split up into the needed parts
Return type: ndarrays

random_crop(image, bboxes)¶

performs a random horizontal crop of the image for use in training

Parameters

image (ndarray) – The image itself ready to be cropped if needed
bboxes (ndarray) – Array of the bounding boxes

Returns

ndarray – The cropped image
ndarray – The bounding box of the image

random_horizontal_flip(image, bboxes)¶

performs a random horizontal flip of the image for use in training

Parameters

image (ndarray) – The image itself ready to be flipped if needed
bboxes (ndarray) – Array of the bounding boxes

Returns

ndarray – The flipped image
ndarray – The bounding box of the image

random_translate(image, bboxes)¶

performs a random translate of the image for use in training

Parameters

image (ndarray) – The image itself ready to be translated if needed
bboxes (ndarray) – Array of the bounding boxes

Returns

ndarray – The translated image
ndarray – The bounding box of the image

core.utils_y module¶

core.utils_y.bbox_ciou(bboxes1, bboxes2)¶

Complete IoU

Parameters

bboxes1 ((a, b, ..., 4)) –
bboxes2 ((A, B, ..., 4)) – x:X is 1:n or n:n or n:1

Returns

Return type

Test

core.utils_y.bbox_giou(bboxes1, bboxes2)¶

Generalized IoU

Parameters

bboxes1 ((a, b, ..., 4)) –
bboxes2 ((A, B, ..., 4)) – x:X is 1:n or n:n or n:1

Returns

Return type

Test

core.utils_y.bbox_iou(bboxes1, bboxes2)¶

Test

Parameters

bboxes1 ((a, b, ..., 4)) –
bboxes2 ((A, B, ..., 4)) – x:X is 1:n or n:n or n:1

Returns

Return type

Test

core.utils_y.count_objects(data, by_class=False, allowed_classes=None)¶

core.utils_y.draw_bbox(image, bboxes, info=False, counted_classes=None, show_label=True, allowed_classes=None)¶

core.utils_y.format_boxes(bboxes, image_height, image_width)¶

Formats the bounding boxes

Parameters

bboxes (ndarray) – The bounding boxes
image_height (int) – The height of the image
image_width (int) – The width of the image

Returns

The now formatted bounding boxes

Return type

ndarray

core.utils_y.freeze_all(model, frozen=True)¶

core.utils_y.get_anchors(anchors)¶

Fetches and convers the anchors list into a formatted numpy array

Parameters: anchors (list[int]) – The anchors ready to be reshaped and converted
Returns: The reshaped anchors
Return type: ndarray

core.utils_y.image_preprocess(image, target_size, gt_boxes=None)¶

core.utils_y.load_config(FLAGS)¶

Loads the config file

Parameters

FLAGS (flag) – The flag that determine which model is being used

Returns

ndarray – Array[int] of the strides
ndarray – Array[int] of the anchors
int – The amount of classes to be classified (80 for coco)
ndarray – Array[float] of the XYscales

core.utils_y.load_freeze_layer(model='yolov4')¶

This function loads which freeze layout to use

Parameters: model (str) – which model is being used yolo v3/v4
Returns: A list which id’s need to be used
Return type: list[str]

core.utils_y.load_weights(model, weights_file, model_name='yolov4')¶

This void function loads the weights

Parameters

model (object) – The model being used
weights_file (weights) – Weights for pretrained model
model_name (str) – which model is being used yolo v3/v4

Returns

Return type

None

core.utils_y.nms(bboxes, iou_threshold, sigma=0.3, method='nms')¶

[summary]

Parameters

bboxes ([type]) – (xmin, ymin, xmax, ymax, score, class)
iou_threshold ([type]) – [description]
sigma (float, optional) – [description], by default 0.3
method (str, optional) – [description], by default ‘nms’

Returns

[description]

Return type

[type]

Note

soft-nms, https://arxiv.org/pdf/1704.04503.pdf: https://github.com/bharatsingh430/soft-nms

core.utils_y.read_class_names(class_file_name)¶

Reads in the classname file and converts it into a dict

Parameters: class_file_name (names) – The .names file of all the class names, separated by a new line
Returns: The IDs and names of the class names
Return type: dict

core.utils_y.unfreeze_all(model, frozen=False)¶

core.utils module¶

core.utils.bbox_ciou(bboxes1, bboxes2)¶

Complete IoU

Parameters

bboxes1 ((a, b, ..., 4)) –
bboxes2 ((A, B, ..., 4)) – x:X is 1:n or n:n or n:1

Returns

Return type

Test

core.utils.bbox_giou(bboxes1, bboxes2)¶

Generalized IoU

Parameters

bboxes1 ((a, b, ..., 4)) –
bboxes2 ((A, B, ..., 4)) – x:X is 1:n or n:n or n:1

Returns

Return type

Test

core.utils.bbox_iou(bboxes1, bboxes2)¶

TEST

Parameters

bboxes1 ((a, b, ..., 4)) –
bboxes2 ((A, B, ..., 4)) – x:X is 1:n or n:n or n:1

Returns

Return type

Test

core.utils.draw_bbox(image, bboxes, info=False, show_label=True, classes=None)¶

core.utils.format_boxes(bboxes, image_height, image_width)¶

Formats the bounding boxes

Parameters

bboxes (ndarray) – The bounding boxes
image_height (int) – The height of the image
image_width (int) – The width of the image

Returns

The now formatted bounding boxes

Return type

ndarray

core.utils.freeze_all(model, frozen=True)¶

core.utils.get_anchors(anchors)¶

Fetches and convers the anchors list into a formatted numpy array

Parameters: anchors (list[int]) – The anchors ready to be reshaped and converted
Returns: The reshaped anchors
Return type: ndarray

core.utils.image_preprocess(image, target_size, gt_boxes=None)¶

core.utils.load_config(FLAGS)¶

Loads the config file

Parameters

FLAGS (flag) – The flag that determine which model is being used

Returns

ndarray – Array[int] of the strides
ndarray – Array[int] of the anchors
int – The amount of classes to be classified (80 for coco)
ndarray – Array[float] of the XYscales

core.utils.load_freeze_layer(model='yolov4')¶

This function loads which freeze layout to use

Parameters: model (str) – which model is being used yolo v3/v4
Returns: A list which id’s need to be used
Return type: list[str]

core.utils.load_weights(model, weights_file, model_name='yolov4')¶

This void function loads the weights

Parameters

model (object) – The model being used
weights_file (weights) – Weights for pretrained model
model_name (str) – which model is being used yolo v3/v4

Returns

Return type

None

core.utils.nms(bboxes, iou_threshold, sigma=0.3, method='nms')¶

[summary]

Parameters

bboxes ([type]) – (xmin, ymin, xmax, ymax, score, class)
iou_threshold ([type]) – [description]
sigma (float, optional) – [description], by default 0.3
method (str, optional) – [description], by default ‘nms’

Returns

[description]

Return type

[type]

Note

soft-nms, https://arxiv.org/pdf/1704.04503.pdf: https://github.com/bharatsingh430/soft-nms

core.utils.read_class_names(class_file_name)¶

Reads in the classname file and converts it into a dict

Parameters: class_file_name (names) – The .names file of all the class names, separated by a new line
Returns: The IDs and names of the class names
Return type: dict

core.utils.unfreeze_all(model, frozen=False)¶