src/utils.py

from typing import List
import torch
import numpy as np
import cv2
import random

from pytorch_grad_cam.base_cam import BaseCAM
from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
from pytorch_grad_cam.utils.model_targets import ClassifierOutputTarget


def cells_to_bboxes(predictions, anchors, S, is_preds=True):
    """
    Scales the predictions coming from the model to
    be relative to the entire image such that they for example later
    can be plotted or.
    INPUT:
    predictions: tensor of size (N, 3, S, S, num_classes+5)
    anchors: the anchors used for the predictions
    S: the number of cells the image is divided in on the width (and height)
    is_preds: whether the input is predictions or the true bounding boxes
    OUTPUT:
    converted_bboxes: the converted boxes of sizes (N, num_anchors, S, S, 1+5) with class index,
                      object score, bounding box coordinates
    """
    BATCH_SIZE = predictions.shape[0]
    num_anchors = len(anchors)
    box_predictions = predictions[..., 1:5]
    if is_preds:
        anchors = anchors.reshape(1, len(anchors), 1, 1, 2)
        box_predictions[..., 0:2] = torch.sigmoid(box_predictions[..., 0:2])
        box_predictions[..., 2:] = torch.exp(box_predictions[..., 2:]) * anchors
        scores = torch.sigmoid(predictions[..., 0:1])
        best_class = torch.argmax(predictions[..., 5:], dim=-1).unsqueeze(-1)
    else:
        scores = predictions[..., 0:1]
        best_class = predictions[..., 5:6]

    cell_indices = (
        torch.arange(S)
        .repeat(predictions.shape[0], 3, S, 1)
        .unsqueeze(-1)
        .to(predictions.device)
    )
    x = 1 / S * (box_predictions[..., 0:1] + cell_indices)
    y = 1 / S * (box_predictions[..., 1:2] + cell_indices.permute(0, 1, 3, 2, 4))
    w_h = 1 / S * box_predictions[..., 2:4]
    converted_bboxes = torch.cat((best_class, scores, x, y, w_h), dim=-1).reshape(BATCH_SIZE, num_anchors * S * S, 6)
    return converted_bboxes.tolist()



def intersection_over_union(boxes_preds, boxes_labels, box_format="midpoint"):
    """
    Video explanation of this function:
    https://youtu.be/XXYG5ZWtjj0

    This function calculates intersection over union (iou) given pred boxes
    and target boxes.

    Parameters:
        boxes_preds (tensor): Predictions of Bounding Boxes (BATCH_SIZE, 4)
        boxes_labels (tensor): Correct labels of Bounding Boxes (BATCH_SIZE, 4)
        box_format (str): midpoint/corners, if boxes (x,y,w,h) or (x1,y1,x2,y2)

    Returns:
        tensor: Intersection over union for all examples
    """

    if box_format == "midpoint":
        box1_x1 = boxes_preds[..., 0:1] - boxes_preds[..., 2:3] / 2
        box1_y1 = boxes_preds[..., 1:2] - boxes_preds[..., 3:4] / 2
        box1_x2 = boxes_preds[..., 0:1] + boxes_preds[..., 2:3] / 2
        box1_y2 = boxes_preds[..., 1:2] + boxes_preds[..., 3:4] / 2
        box2_x1 = boxes_labels[..., 0:1] - boxes_labels[..., 2:3] / 2
        box2_y1 = boxes_labels[..., 1:2] - boxes_labels[..., 3:4] / 2
        box2_x2 = boxes_labels[..., 0:1] + boxes_labels[..., 2:3] / 2
        box2_y2 = boxes_labels[..., 1:2] + boxes_labels[..., 3:4] / 2

    if box_format == "corners":
        box1_x1 = boxes_preds[..., 0:1]
        box1_y1 = boxes_preds[..., 1:2]
        box1_x2 = boxes_preds[..., 2:3]
        box1_y2 = boxes_preds[..., 3:4]
        box2_x1 = boxes_labels[..., 0:1]
        box2_y1 = boxes_labels[..., 1:2]
        box2_x2 = boxes_labels[..., 2:3]
        box2_y2 = boxes_labels[..., 3:4]

    x1 = torch.max(box1_x1, box2_x1)
    y1 = torch.max(box1_y1, box2_y1)
    x2 = torch.min(box1_x2, box2_x2)
    y2 = torch.min(box1_y2, box2_y2)

    intersection = (x2 - x1).clamp(0) * (y2 - y1).clamp(0)
    box1_area = abs((box1_x2 - box1_x1) * (box1_y2 - box1_y1))
    box2_area = abs((box2_x2 - box2_x1) * (box2_y2 - box2_y1))

    return intersection / (box1_area + box2_area - intersection + 1e-6)

def non_max_suppression(bboxes, iou_threshold, threshold, box_format="corners"):
    """
    Video explanation of this function:
    https://youtu.be/YDkjWEN8jNA

    Does Non Max Suppression given bboxes

    Parameters:
        bboxes (list): list of lists containing all bboxes with each bboxes
        specified as [class_pred, prob_score, x1, y1, x2, y2]
        iou_threshold (float): threshold where predicted bboxes is correct
        threshold (float): threshold to remove predicted bboxes (independent of IoU)
        box_format (str): "midpoint" or "corners" used to specify bboxes

    Returns:
        list: bboxes after performing NMS given a specific IoU threshold
    """

    assert type(bboxes) == list

    bboxes = [box for box in bboxes if box[1] > threshold]
    bboxes = sorted(bboxes, key=lambda x: x[1], reverse=True)
    bboxes_after_nms = []

    while bboxes:
        chosen_box = bboxes.pop(0)

        bboxes = [
            box
            for box in bboxes
            if box[0] != chosen_box[0]
            or intersection_over_union(
                torch.tensor(chosen_box[2:]),
                torch.tensor(box[2:]),
                box_format=box_format,
            )
            < iou_threshold
        ]

        bboxes_after_nms.append(chosen_box)

    return bboxes_after_nms




def draw_predictions(image: np.ndarray, boxes: List[List], class_labels: List[str]) -> np.ndarray:
    """Plots predicted bounding boxes on the image"""

    colors = [[random.randint(0, 255) for _ in range(3)] for name in class_labels]

    im = np.array(image)
    height, width, _ = im.shape
    bbox_thick = int(0.6 * (height + width) / 600)

    # Create a Rectangle patch
    for box in boxes:
        assert len(box) == 6, "box should contain class pred, confidence, x, y, width, height"
        class_pred = box[0]
        conf = box[1]
        box = box[2:]
        upper_left_x = box[0] - box[2] / 2
        upper_left_y = box[1] - box[3] / 2
        
        x1  = int(upper_left_x * width)
        y1 = int(upper_left_y * height)
        
        x2 = x1 + int(box[2] * width)
        y2 = y1 + int(box[3] * height)
        
        cv2.rectangle(
            image,
            (x1, y1), (x2, y2),
            color=colors[int(class_pred)],
            thickness=bbox_thick
        )
        text = f"{class_labels[int(class_pred)]}: {conf:.2f}"
        t_size = cv2.getTextSize(text, 0, 0.7, thickness=bbox_thick // 2)[0]
        c3 = (x1 + t_size[0], y1 - t_size[1] - 3)

        cv2.rectangle(image, (x1, y1), c3, colors[int(class_pred)], -1)
        cv2.putText(
            image,
            text,
            (x1, y1 - 2),
            cv2.FONT_HERSHEY_SIMPLEX,
            0.7,
            (0, 0, 0),
            bbox_thick // 2,
            lineType=cv2.LINE_AA,
        )

    return image


class YoloCAM(BaseCAM):
    def __init__(self, model, target_layers, use_cuda=False,
                 reshape_transform=None):
        super(YoloCAM, self).__init__(model,
                                       target_layers,
                                       use_cuda,
                                       reshape_transform,
                                       uses_gradients=False)

    def forward(self,
                input_tensor: torch.Tensor,
                scaled_anchors: torch.Tensor,
                targets: List[torch.nn.Module],
                eigen_smooth: bool = False) -> np.ndarray:

        if self.cuda:
            input_tensor = input_tensor.cuda()

        if self.compute_input_gradient:
            input_tensor = torch.autograd.Variable(input_tensor,
                                                   requires_grad=True)

        outputs = self.activations_and_grads(input_tensor)
        if targets is None:
            bboxes = [[] for _ in range(1)]
            for i in range(3):
                batch_size, A, S, _, _ = outputs[i].shape
                anchor = scaled_anchors[i]
                boxes_scale_i = cells_to_bboxes(
                    outputs[i], anchor, S=S, is_preds=True
                )
                for idx, (box) in enumerate(boxes_scale_i):
                    bboxes[idx] += box
            
            nms_boxes = non_max_suppression(
                bboxes[0], iou_threshold=0.5, threshold=0.4, box_format="midpoint",
            )
            # target_categories = np.argmax(outputs.cpu().data.numpy(), axis=-1)
            target_categories = [box[0] for box in nms_boxes]
            targets = [ClassifierOutputTarget(
                category) for category in target_categories]

        if self.uses_gradients:
            self.model.zero_grad()
            loss = sum([target(output)
                       for target, output in zip(targets, outputs)])
            loss.backward(retain_graph=True)

        # In most of the saliency attribution papers, the saliency is
        # computed with a single target layer.
        # Commonly it is the last convolutional layer.
        # Here we support passing a list with multiple target layers.
        # It will compute the saliency image for every image,
        # and then aggregate them (with a default mean aggregation).
        # This gives you more flexibility in case you just want to
        # use all conv layers for example, all Batchnorm layers,
        # or something else.
        cam_per_layer = self.compute_cam_per_layer(input_tensor,
                                                   targets,
                                                   eigen_smooth)
        return self.aggregate_multi_layers(cam_per_layer)
    
    def get_cam_image(self,
                      input_tensor,
                      target_layer,
                      target_category,
                      activations,
                      grads,
                      eigen_smooth):
        return get_2d_projection(activations)