register loss to total_loss and loss_dict

Input: - boxes: FloatTensor of shape (O, 4) giving boxes in the [x0, y0, x1, y1] format in the [0, 1] coordinate space - H, W: Scal

under construction!!!

Input: - imgs: FloatTensor of shape (N, C, H, W) giving preprocessed images Output: - imgs_de: ByteTensor of shape (N, C, H, W) givi

:param img_list: list of [numpy in shape of (H, W, C)] in BGR :return: list of FloatTensor [(C, H, W)] in RGB

:param logit: (O, 1, H, W) :return: softmax(O, 1, H, W)

:param bbox: (O, 4) in [0, 1] :param feat: (O, D, h, w) :return: (O, D, H, W)

:param img: FloatTensor in shape (C, H, W) in RGB :return: uint8 numpy in shape (H, W, C) in BGR

Convert boxes from [x0, y0, x1, y1] format to [xc, yc, w, h] format Input: - boxes: Input boxes of shape (N, 4) in [x0, y0, x1, y1] form

Input: - boxes: FloatTensor of shape (O, 4) giving boxes in the [x0, y0, x1, y1] in (0, 1) format in the [0, 1] coordinate space -

image: N, C, H, W

image: N, C, H, W

Helper method to compute padding size

(V, O, 2), (V, T, 3)

Input: - samples: FloatTensor of shape (O, D, H, W) - obj_to_img: LongTensor of shape (O,) with each element in the range [0, N) ma

Perform bilinear sampling on the features in feats using the sampling grid given by X and Y. Inputs: - feats: Tensor holding input feature m

Take differentiable crops of feats specified by bbox. Inputs: - feats: Tensor of shape (N, C, H, W) - bbox: Bounding box coordinates of shap

todo: clean!

:param im0: (dt, V, 1, 3, H, W) in mean=(0, 0, 0) std=(1, 1, 1,)

generate sample from N(mu, var). or no sample

Resize gt to the size of pred. and calculate loss

:param pred_list: list of image: [im0, im2, ...] :return:

:param box: (O, 2) :param img: :param trip: (T, 3) :return:

:param bbox: (O, 4) in [0, 1] :param feat: (O, D, h, w) :return: (O, D, H, W)

Vectorized version of torch.linspace. Inputs: - start: Tensor of any shape - end: Tensor of the same shape as start - steps: Integer Re

:param boxes: (O, 4) [x0, y0, x1, y1] in (0, 1) :return: (O, H, W, 2)

:param bg_feat: (N, Di, h, w) :param mask: (N, 1, H, W) :return: (N, Do, Hm, Wm)

.view(V, -1, bg_feat.size(-2), bg_feat.size(-1))

:param bg_feat: (N, Di, h, w) :param mask: (N, 1, H, W) :return: (N, Do, Hm, Wm)

:param holder: (V, num, D) :param valid: (V, 1, O) :return: (V, O, D)

:param bbox: (V, O, D) :param valid: (V, 1, O) :return: (V, num, D)

Apply box transforms to a set of anchor boxes. Inputs: - anchors: Anchor boxes of shape (N, 4), where each anchor is specified in