Function entropy_loss

loss.py:273–306

(prob_volume, depth_gt, mask, depth_value, return_prob_map=False)

def sigmoid(x, base=2.71828):
    return 1 / (1 + torch.pow(base, -x))
def entropy_loss(prob_volume, depth_gt, mask, depth_value, return_prob_map=False):
    # from AA
    mask_true = mask
    valid_pixel_num = torch.sum(mask_true, dim=[1,2]) + 1e-6

    shape = depth_gt.shape          # B,H,W

    depth_num = depth_value.shape[1]
    if len(depth_value.shape) < 3:
        depth_value_mat = depth_value.repeat(shape[1], shape[2], 1, 1).permute(2,3,0,1)     # B,N,H,W
    else:
        depth_value_mat = depth_value

    gt_index_image = torch.argmin(torch.abs(depth_value_mat-depth_gt.unsqueeze(1)), dim=1)
    temp=gt_index_image

    gt_index_image = torch.mul(mask_true, gt_index_image.type(torch.float))
    gt_index_image = torch.round(gt_index_image).type(torch.long).unsqueeze(1) # B, 1, H, W

    # gt index map -> gt one hot volume (B x 1 x H x W )
    gt_index_volume = torch.zeros(shape[0], depth_num, shape[1], shape[2]).type(mask_true.type()).scatter_(1, gt_index_image, 1)

    # cross entropy image (B x D X H x W)
    cross_entropy_image = -torch.sum(gt_index_volume * torch.log(prob_volume + 1e-6), dim=1).squeeze(1) # B, 1, H, W

    # masked cross entropy loss
    masked_cross_entropy_image = torch.mul(mask_true, cross_entropy_image) # valid pixel
    masked_cross_entropy = torch.sum(masked_cross_entropy_image, dim=[1, 2])

    masked_cross_entropy = torch.mean(masked_cross_entropy / valid_pixel_num) # Origin use sum : aggregate with batch
    # winner-take-all depth map
    wta_index_map = torch.argmax(prob_volume, dim=1, keepdim=True).type(torch.long).squeeze(1)

    return masked_cross_entropy

def entropy_loss_expand(prob_volume, depth_gt, mask, depth_value, return_prob_map=False):
    # from AA