Python numpy.maximum() Examples

def generate_moving_mnist(self, num_digits=2):
    '''
    Get random trajectories for the digits and generate a video.
    '''
    data = np.zeros((self.n_frames_total, self.image_size_, self.image_size_), dtype=np.float32)
    for n in range(num_digits):
      # Trajectory
      start_y, start_x = self.get_random_trajectory(self.n_frames_total)
      ind = random.randint(0, self.mnist.shape[0] - 1)
      digit_image = self.mnist[ind]
      for i in range(self.n_frames_total):
        top    = start_y[i]
        left   = start_x[i]
        bottom = top + self.digit_size_
        right  = left + self.digit_size_
        # Draw digit
        data[i, top:bottom, left:right] = np.maximum(data[i, top:bottom, left:right], digit_image)

    data = data[..., np.newaxis]
    return data 

def detect(self, img):
        """
        img: rgb 3 channel
        """
        minsize = 20  # minimum size of face
        threshold = [0.6, 0.7, 0.7]  # three steps's threshold
        factor = 0.709  # scale factor

        bounding_boxes, _ = FaceDet.detect_face(
                img, minsize, self.pnet, self.rnet, self.onet, threshold, factor)
        area = (bounding_boxes[:, 2] - bounding_boxes[:, 0]) * (bounding_boxes[:, 3] - bounding_boxes[:, 1])
        face_idx = area.argmax()
        bbox = bounding_boxes[face_idx][:4]  # xy,xy

        margin = 32
        x0 = np.maximum(bbox[0] - margin // 2, 0)
        y0 = np.maximum(bbox[1] - margin // 2, 0)
        x1 = np.minimum(bbox[2] + margin // 2, img.shape[1])
        y1 = np.minimum(bbox[3] + margin // 2, img.shape[0])
        x0, y0, x1, y1 = bbox = [int(k + 0.5) for k in [x0, y0, x1, y1]]
        cropped = img[y0:y1, x0:x1, :]
        scaled = cv2.resize(cropped, (160, 160), interpolation=cv2.INTER_LINEAR)
        return scaled, bbox 

def crop(self, bbox):
        """See :func:`BaseInstanceMasks.crop`."""
        assert isinstance(bbox, np.ndarray)
        assert bbox.ndim == 1

        # clip the boundary
        bbox = bbox.copy()
        bbox[0::2] = np.clip(bbox[0::2], 0, self.width)
        bbox[1::2] = np.clip(bbox[1::2], 0, self.height)
        x1, y1, x2, y2 = bbox
        w = np.maximum(x2 - x1, 1)
        h = np.maximum(y2 - y1, 1)

        if len(self.masks) == 0:
            cropped_masks = np.empty((0, h, w), dtype=np.uint8)
        else:
            cropped_masks = self.masks[:, y1:y1 + h, x1:x1 + w]
        return BitmapMasks(cropped_masks, h, w) 

def apply_perturbations(i, j, X, increase, theta, clip_min, clip_max):
    """
    TensorFlow implementation for apply perturbations to input features based
    on salency maps
    :param i: index of first selected feature
    :param j: index of second selected feature
    :param X: a matrix containing our input features for our sample
    :param increase: boolean; true if we are increasing pixels, false otherwise
    :param theta: delta for each feature adjustment
    :param clip_min: mininum value for a feature in our sample
    :param clip_max: maximum value for a feature in our sample
    : return: a perturbed input feature matrix for a target class
    """

    # perturb our input sample
    if increase:
        X[0, i] = np.minimum(clip_max, X[0, i] + theta)
        X[0, j] = np.minimum(clip_max, X[0, j] + theta)
    else:
        X[0, i] = np.maximum(clip_min, X[0, i] - theta)
        X[0, j] = np.maximum(clip_min, X[0, j] - theta)

    return X 

def backward(self, req, out_grad, in_data, out_data, in_grad, aux):
        if ReluOp.guided_backprop:
            # Get output and gradients of output
            y = out_data[0]
            dy = out_grad[0]
            # Zero out the negatives in the gradients of the output
            dy_positives = nd.maximum(dy, nd.zeros_like(dy))
            # What output values were greater than 0?
            y_ones = y.__gt__(0)
            # Mask out the values for which at least one of dy or y is negative
            dx = dy_positives * y_ones
            self.assign(in_grad[0], req[0], dx)
        else:
            # Regular backward for ReLU
            x = in_data[0]
            x_gt_zero = x.__gt__(0)
            dx = out_grad[0] * x_gt_zero
            self.assign(in_grad[0], req[0], dx) 

def voc_ap(rec, prec, use_07_metric=False):
        if use_07_metric:
            ap = 0.
            for t in np.arange(0., 1.1, 0.1):
                if np.sum(rec >= t) == 0:
                    p = 0
                else:
                    p = np.max(prec[rec >= t])
                ap += p / 11.
        else:
            # append sentinel values at both ends
            mrec = np.concatenate(([0.], rec, [1.]))
            mpre = np.concatenate(([0.], prec, [0.]))

            # compute precision integration ladder
            for i in range(mpre.size - 1, 0, -1):
                mpre[i - 1] = np.maximum(mpre[i - 1], mpre[i])

            # look for recall value changes
            i = np.where(mrec[1:] != mrec[:-1])[0]

            # sum (\delta recall) * prec
            ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])
        return ap 

def test_quantize_float32_to_int8():
    shape = rand_shape_nd(4)
    data = rand_ndarray(shape, 'default', dtype='float32')
    min_range = mx.nd.min(data)
    max_range = mx.nd.max(data)
    qdata, min_val, max_val = mx.nd.contrib.quantize(data, min_range, max_range, out_type='int8')
    data_np = data.asnumpy()
    min_range = min_range.asscalar()
    max_range = max_range.asscalar()
    real_range = np.maximum(np.abs(min_range), np.abs(max_range))
    quantized_range = 127.0
    scale = quantized_range / real_range
    assert qdata.dtype == np.int8
    assert min_val.dtype == np.float32
    assert max_val.dtype == np.float32
    assert same(min_val.asscalar(), -real_range)
    assert same(max_val.asscalar(), real_range)
    qdata_np = (np.sign(data_np) * np.minimum(np.abs(data_np) * scale + 0.5, quantized_range)).astype(np.int8)
    assert_almost_equal(qdata.asnumpy(), qdata_np, atol = 1) 

def heuristic_fn_vec(n1, n2, n_ori, step_size):
  # n1 is a vector and n2 is a single point.
  dx = (n1[:,0] - n2[0,0])/step_size
  dy = (n1[:,1] - n2[0,1])/step_size
  dt = n1[:,2] - n2[0,2]
  dt = np.mod(dt, n_ori)
  dt = np.minimum(dt, n_ori-dt)

  if n_ori == 6:
    if dx*dy > 0:
      d = np.maximum(np.abs(dx), np.abs(dy))
    else:
      d = np.abs(dy-dx)
  elif n_ori == 4:
    d = np.abs(dx) + np.abs(dy)

  return (d + dt).reshape((-1,1)) 

def resize_maps(map, map_scales, resize_method):
  scaled_maps = []
  for i, sc in enumerate(map_scales):
    if resize_method == 'antialiasing':
      # Resize using open cv so that we can compute the size.
      # Use PIL resize to use anti aliasing feature.
      map_ = cv2.resize(map*1, None, None, fx=sc, fy=sc, interpolation=cv2.INTER_LINEAR)
      w = map_.shape[1]; h = map_.shape[0]

      map_img = PIL.Image.fromarray((map*255).astype(np.uint8))
      map__img = map_img.resize((w,h), PIL.Image.ANTIALIAS)
      map_ = np.asarray(map__img).astype(np.float32)
      map_ = map_/255.
      map_ = np.minimum(map_, 1.0)
      map_ = np.maximum(map_, 0.0)
    elif resize_method == 'linear_noantialiasing':
      map_ = cv2.resize(map*1, None, None, fx=sc, fy=sc, interpolation=cv2.INTER_LINEAR)
    else:
      logging.error('Unknown resizing method')
    scaled_maps.append(map_)
  return scaled_maps 

def intersection(boxes1, boxes2):
  """Compute pairwise intersection areas between boxes.

  Args:
    boxes1: a numpy array with shape [N, 4] holding N boxes
    boxes2: a numpy array with shape [M, 4] holding M boxes

  Returns:
    a numpy array with shape [N*M] representing pairwise intersection area
  """
  [y_min1, x_min1, y_max1, x_max1] = np.split(boxes1, 4, axis=1)
  [y_min2, x_min2, y_max2, x_max2] = np.split(boxes2, 4, axis=1)

  all_pairs_min_ymax = np.minimum(y_max1, np.transpose(y_max2))
  all_pairs_max_ymin = np.maximum(y_min1, np.transpose(y_min2))
  intersect_heights = np.maximum(
      np.zeros(all_pairs_max_ymin.shape),
      all_pairs_min_ymax - all_pairs_max_ymin)
  all_pairs_min_xmax = np.minimum(x_max1, np.transpose(x_max2))
  all_pairs_max_xmin = np.maximum(x_min1, np.transpose(x_min2))
  intersect_widths = np.maximum(
      np.zeros(all_pairs_max_xmin.shape),
      all_pairs_min_xmax - all_pairs_max_xmin)
  return intersect_heights * intersect_widths 

def create_random_boxes(num_boxes, max_height, max_width):
  """Creates random bounding boxes of specific maximum height and width.

  Args:
    num_boxes: number of boxes.
    max_height: maximum height of boxes.
    max_width: maximum width of boxes.

  Returns:
    boxes: numpy array of shape [num_boxes, 4]. Each row is in form
        [y_min, x_min, y_max, x_max].
  """

  y_1 = np.random.uniform(size=(1, num_boxes)) * max_height
  y_2 = np.random.uniform(size=(1, num_boxes)) * max_height
  x_1 = np.random.uniform(size=(1, num_boxes)) * max_width
  x_2 = np.random.uniform(size=(1, num_boxes)) * max_width

  boxes = np.zeros(shape=(num_boxes, 4))
  boxes[:, 0] = np.minimum(y_1, y_2)
  boxes[:, 1] = np.minimum(x_1, x_2)
  boxes[:, 2] = np.maximum(y_1, y_2)
  boxes[:, 3] = np.maximum(x_1, x_2)

  return boxes.astype(np.float32) 

def nullspace(A, atol=1e-13, rtol=0):
    """
    Compute an approximate basis for the nullspace of A.

    INPUT   (1) array 'A': 1-D array with length k will be treated
                as a 2-D with shape (1, k).
            (2) float 'atol': the absolute tolerance for a zero singular value.
                Singular values smaller than `atol` are considered to be zero.
            (3) float 'rtol': relative tolerance. Singular values less than
                rtol*smax are considered to be zero, where smax is the largest
                singular value.

                If both `atol` and `rtol` are positive, the combined tolerance
                is the maximum of the two; tol = max(atol, rtol * smax)
                Singular values smaller than `tol` are considered to be zero.
    OUTPUT  (1) array 'B': if A is an array with shape (m, k), then B will be
                an array with shape (k, n), where n is the estimated dimension
                of the nullspace of A.  The columns of B are a basis for the
                nullspace; each element in np.dot(A, B) will be
                approximately zero.
    """
    # Expand A to a matrix
    A = np.atleast_2d(A)

    # Singular value decomposition
    u, s, vh = al.svd(A)

    # Set tolerance
    tol = max(atol, rtol * s[0])

    # Compute the number of non-zero entries
    nnz = (s >= tol).sum()

    # Conjugate and transpose to ensure real numbers
    ns = vh[nnz:].conj().T

    return ns 

def project_simplex(self, v, z=1.0):
        """
        Project vector onto simplex using sorting.

        Reference: "Efficient Projections onto the L1-Ball for Learning in High
        Dimensions (Duchi, Shalev-Shwartz, Singer, Chandra, 2006)."

        Parameters
        ----------
        v : array
            vector to be projected (n dimensions by 0)
        z : float
            constant (def: 1.0)

        Returns
        -------
        w : array
            projected vector (n dimensions by 0)

        """
        # Number of dimensions
        n = v.shape[0]

        # Sort vector
        mu = np.sort(v, axis=0)[::-1]

        # Find rho
        C = np.cumsum(mu) - z
        j = np.arange(n) + 1
        rho = j[mu - C/j > 0][-1]

        # Define theta
        theta = C[mu - C/j > 0][-1] / float(rho)

        # Subtract theta from original vector and cap at 0
        w = np.maximum(v - theta, 0)

        # Return projected vector
        return w 

def iwe_kernel_densities(self, X, Z, clip=1000):
        """
        Estimate importance weights based on kernel density estimation.

        Parameters
        ----------
        X : array
            source data (N samples by D features)
        Z : array
            target data (M samples by D features)
        clip : float
            maximum allowed value for individual weights (def: 1000)

        Returns
        -------
        array
            importance weights (N samples by 1)

        """
        # Data shapes
        N, DX = X.shape
        M, DZ = Z.shape

        # Assert equivalent dimensionalities
        assert DX == DZ

        # Compute probabilities based on source kernel densities
        pT = st.gaussian_kde(Z.T).pdf(X.T)
        pS = st.gaussian_kde(X.T).pdf(X.T)

        # Check for numerics
        assert not np.any(np.isnan(pT)) or np.any(pT == 0)
        assert not np.any(np.isnan(pS)) or np.any(pS == 0)

        # Compute importance weights
        iw = pT / pS

        # Clip importance weights
        return np.minimum(clip, np.maximum(0, iw)) 

def __init__(self, nelx, nely, rmin):
        """
        Filter: Build (and assemble) the index+data vectors for the coo matrix
        format.
        """
        nfilter = int(nelx * nely * ((2 * (np.ceil(rmin) - 1) + 1)**2))
        iH = np.zeros(nfilter)
        jH = np.zeros(nfilter)
        sH = np.zeros(nfilter)
        cc = 0
        for i in range(nelx):
            for j in range(nely):
                row = i * nely + j
                kk1 = int(np.maximum(i - (np.ceil(rmin) - 1), 0))
                kk2 = int(np.minimum(i + np.ceil(rmin), nelx))
                ll1 = int(np.maximum(j - (np.ceil(rmin) - 1), 0))
                ll2 = int(np.minimum(j + np.ceil(rmin), nely))
                for k in range(kk1, kk2):
                    for l in range(ll1, ll2):
                        col = k * nely + l
                        fac = rmin - np.sqrt(
                            ((i - k) * (i - k) + (j - l) * (j - l)))
                        iH[cc] = row
                        jH[cc] = col
                        sH[cc] = np.maximum(0.0, fac)
                        cc = cc + 1
        # Finalize assembly and convert to csc format
        self.H = scipy.sparse.coo_matrix((sH, (iH, jH)),
            shape=(nelx * nely, nelx * nely)).tocsc()
        self.Hs = self.H.sum(1) 

def filter_compliance_sensitivities(self, xPhys, dc, ft):
        if ft == 0:
            dc[:] = (np.asarray((self.H * (xPhys * dc))[np.newaxis].T /
                self.Hs)[:, 0] / np.maximum(0.001, xPhys))
        elif ft == 1:
            dc[:] = np.asarray(self.H * (dc[np.newaxis].T / self.Hs))[:, 0] 

def __init__(self, nelx, nely, rmin):
        """
        Filter: Build (and assemble) the index+data vectors for the coo matrix
        format.
        """
        nfilter = int(nelx * nely * ((2 * (np.ceil(rmin) - 1) + 1)**2))
        iH = np.zeros(nfilter)
        jH = np.zeros(nfilter)
        sH = np.zeros(nfilter)
        cc = 0
        for i in range(nelx):
            for j in range(nely):
                row = i * nely + j
                kk1 = int(np.maximum(i - (np.ceil(rmin) - 1), 0))
                kk2 = int(np.minimum(i + np.ceil(rmin), nelx))
                ll1 = int(np.maximum(j - (np.ceil(rmin) - 1), 0))
                ll2 = int(np.minimum(j + np.ceil(rmin), nely))
                for k in range(kk1, kk2):
                    for l in range(ll1, ll2):
                        col = k * nely + l
                        fac = rmin - np.sqrt(
                            ((i - k) * (i - k) + (j - l) * (j - l)))
                        iH[cc] = row
                        jH[cc] = col
                        sH[cc] = np.maximum(0.0, fac)
                        cc = cc + 1
        # Finalize assembly and convert to csc format
        self.H = scipy.sparse.coo_matrix((sH, (iH, jH)),
            shape=(nelx * nely, nelx * nely)).tocsc()
        self.Hs = self.H.sum(1) 

def filter_compliance_sensitivities(self, xPhys, dc, ft):
        if ft == 0:
            dc[:] = (np.asarray((self.H * (xPhys * dc))[np.newaxis].T /
                self.Hs)[:, 0] / np.maximum(0.001, xPhys))
        elif ft == 1:
            dc[:] = np.asarray(self.H * (dc[np.newaxis].T / self.Hs))[:, 0] 

def _clip_boxes(boxes, im_shape):
  """Clip boxes to image boundaries."""
  # x1 >= 0
  boxes[:, 0::4] = np.maximum(boxes[:, 0::4], 0)
  # y1 >= 0
  boxes[:, 1::4] = np.maximum(boxes[:, 1::4], 0)
  # x2 < im_shape[1]
  boxes[:, 2::4] = np.minimum(boxes[:, 2::4], im_shape[1] - 1)
  # y2 < im_shape[0]
  boxes[:, 3::4] = np.minimum(boxes[:, 3::4], im_shape[0] - 1)
  return boxes 

def apply_nms(all_boxes, thresh):
  """Apply non-maximum suppression to all predicted boxes output by the
  test_net method.
  """
  num_classes = len(all_boxes)
  num_images = len(all_boxes[0])
  nms_boxes = [[[] for _ in range(num_images)] for _ in range(num_classes)]
  for cls_ind in range(num_classes):
    for im_ind in range(num_images):
      dets = all_boxes[cls_ind][im_ind]
      if dets == []:
        continue

      x1 = dets[:, 0]
      y1 = dets[:, 1]
      x2 = dets[:, 2]
      y2 = dets[:, 3]
      scores = dets[:, 4]
      inds = np.where((x2 > x1) & (y2 > y1))[0]
      dets = dets[inds,:]
      if dets == []:
        continue

      keep = nms(torch.from_numpy(dets), thresh).numpy()
      if len(keep) == 0:
        continue
      nms_boxes[cls_ind][im_ind] = dets[keep, :].copy()
  return nms_boxes 

def _clip_boxes(boxes, im_shape):
  """Clip boxes to image boundaries."""
  # x1 >= 0
  boxes[:, 0::4] = np.maximum(boxes[:, 0::4], 0)
  # y1 >= 0
  boxes[:, 1::4] = np.maximum(boxes[:, 1::4], 0)
  # x2 < im_shape[1]
  boxes[:, 2::4] = np.minimum(boxes[:, 2::4], im_shape[1] - 1)
  # y2 < im_shape[0]
  boxes[:, 3::4] = np.minimum(boxes[:, 3::4], im_shape[0] - 1)
  return boxes 

def apply_nms(all_boxes, thresh):
  """Apply non-maximum suppression to all predicted boxes output by the
  test_net method.
  """
  num_classes = len(all_boxes)
  num_images = len(all_boxes[0])
  nms_boxes = [[[] for _ in range(num_images)] for _ in range(num_classes)]
  for cls_ind in range(num_classes):
    for im_ind in range(num_images):
      dets = all_boxes[cls_ind][im_ind]
      if dets == []:
        continue

      x1 = dets[:, 0]
      y1 = dets[:, 1]
      x2 = dets[:, 2]
      y2 = dets[:, 3]
      scores = dets[:, 4]
      inds = np.where((x2 > x1) & (y2 > y1))[0]
      dets = dets[inds,:]
      if dets == []:
        continue

      keep = nms(torch.from_numpy(dets), thresh).numpy()
      if len(keep) == 0:
        continue
      nms_boxes[cls_ind][im_ind] = dets[keep, :].copy()
  return nms_boxes 

def voc_ap(rec, prec, use_07_metric=False):
  """ ap = voc_ap(rec, prec, [use_07_metric])
  Compute VOC AP given precision and recall.
  If use_07_metric is true, uses the
  VOC 07 11 point method (default:False).
  """
  if use_07_metric:
    # 11 point metric
    ap = 0.
    for t in np.arange(0., 1.1, 0.1):
      if np.sum(rec >= t) == 0:
        p = 0
      else:
        p = np.max(prec[rec >= t])
      ap = ap + p / 11.
  else:
    # correct AP calculation
    # first append sentinel values at the end
    mrec = np.concatenate(([0.], rec, [1.]))
    mpre = np.concatenate(([0.], prec, [0.]))

    # compute the precision envelope
    for i in range(mpre.size - 1, 0, -1):
      mpre[i - 1] = np.maximum(mpre[i - 1], mpre[i])

    # to calculate area under PR curve, look for points
    # where X axis (recall) changes value
    i = np.where(mrec[1:] != mrec[:-1])[0]

    # and sum (\Delta recall) * prec
    ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])
  return ap 

def setup_text_label(text, font='Calibri', fontsize=32, padding=6, glow_size=2.0, glow_coef=3.0, glow_exp=2.0, cache_size=100): # => (alpha, glow)
    # Lookup from cache.
    key = (text, font, fontsize, padding, glow_size, glow_coef, glow_exp)
    if key in _text_label_cache:
        value = _text_label_cache[key]
        del _text_label_cache[key] # LRU policy
        _text_label_cache[key] = value
        return value

    # Limit cache size.
    while len(_text_label_cache) >= cache_size:
        _text_label_cache.popitem(last=False)

    # Render text.
    import moviepy.editor # pip install moviepy
    alpha = moviepy.editor.TextClip(text, font=font, fontsize=fontsize).mask.make_frame(0)
    alpha = np.pad(alpha, padding, mode='constant', constant_values=0.0)
    glow = scipy.ndimage.gaussian_filter(alpha, glow_size)
    glow = 1.0 - np.maximum(1.0 - glow * glow_coef, 0.0) ** glow_exp

    # Add to cache.
    value = (alpha, glow)
    _text_label_cache[key] = value
    return value

#---------------------------------------------------------------------------- 

def griffin_lim(spectrogram):
    '''Applies Griffin-Lim's raw.'''
    X_best = copy.deepcopy(spectrogram)
    for i in range(hp.n_iter):
        X_t = invert_spectrogram(X_best)
        est = librosa.stft(X_t, hp.n_fft, hp.hop_length, win_length=hp.win_length)
        phase = est / np.maximum(1e-8, np.abs(est))
        X_best = spectrogram * phase
    X_t = invert_spectrogram(X_best)
    y = np.real(X_t)

    return y 

def random_boxes(num=1, scale=1, rng=None):
    """Simple version of ``kwimage.Boxes.random``

    Returns:
        Tensor: shape (n, 4) in x1, y1, x2, y2 format.

    References:
        https://gitlab.kitware.com/computer-vision/kwimage/blob/master/kwimage/structs/boxes.py#L1390

    Example:
        >>> num = 3
        >>> scale = 512
        >>> rng = 0
        >>> boxes = random_boxes(num, scale, rng)
        >>> print(boxes)
        tensor([[280.9925, 278.9802, 308.6148, 366.1769],
                [216.9113, 330.6978, 224.0446, 456.5878],
                [405.3632, 196.3221, 493.3953, 270.7942]])
    """
    rng = ensure_rng(rng)

    tlbr = rng.rand(num, 4).astype(np.float32)

    tl_x = np.minimum(tlbr[:, 0], tlbr[:, 2])
    tl_y = np.minimum(tlbr[:, 1], tlbr[:, 3])
    br_x = np.maximum(tlbr[:, 0], tlbr[:, 2])
    br_y = np.maximum(tlbr[:, 1], tlbr[:, 3])

    tlbr[:, 0] = tl_x * scale
    tlbr[:, 1] = tl_y * scale
    tlbr[:, 2] = br_x * scale
    tlbr[:, 3] = br_y * scale

    boxes = torch.from_numpy(tlbr)
    return boxes 

def rescale(self, scale, interpolation='nearest'):
        """Rescale masks as large as possible while keeping the aspect ratio.
        For details can refer to `mmcv.imrescale`.

        Args:
            scale (tuple[int]): The maximum size (h, w) of rescaled mask.
            interpolation (str): Same as :func:`mmcv.imrescale`.

        Returns:
            BaseInstanceMasks: The rescaled masks.
        """
        pass 

def crop_and_resize(self,
                        bboxes,
                        out_shape,
                        inds,
                        device='cpu',
                        interpolation='bilinear'):
        """see :func:`BaseInstanceMasks.crop_and_resize`"""
        out_h, out_w = out_shape
        if len(self.masks) == 0:
            return PolygonMasks([], out_h, out_w)

        resized_masks = []
        for i in range(len(bboxes)):
            mask = self.masks[inds[i]]
            bbox = bboxes[i, :]
            x1, y1, x2, y2 = bbox
            w = np.maximum(x2 - x1, 1)
            h = np.maximum(y2 - y1, 1)
            h_scale = out_h / max(h, 0.1)  # avoid too large scale
            w_scale = out_w / max(w, 0.1)

            resized_mask = []
            for p in mask:
                p = p.copy()
                # crop
                # pycocotools will clip the boundary
                p[0::2] -= bbox[0]
                p[1::2] -= bbox[1]

                # resize
                p[0::2] *= w_scale
                p[1::2] *= h_scale
                resized_mask.append(p)
            resized_masks.append(resized_mask)
        return PolygonMasks(resized_masks, *out_shape) 

def vatm(model,
         x,
         logits,
         eps,
         num_iterations=1,
         xi=1e-6,
         clip_min=None,
         clip_max=None,
         scope=None):
    """
    Tensorflow implementation of the perturbation method used for virtual
    adversarial training: https://arxiv.org/abs/1507.00677
    :param model: the model which returns the network unnormalized logits
    :param x: the input placeholder
    :param logits: the model's unnormalized output tensor (the input to
                   the softmax layer)
    :param eps: the epsilon (input variation parameter)
    :param num_iterations: the number of iterations
    :param xi: the finite difference parameter
    :param clip_min: optional parameter that can be used to set a minimum
                    value for components of the example returned
    :param clip_max: optional parameter that can be used to set a maximum
                    value for components of the example returned
    :param seed: the seed for random generator
    :return: a tensor for the adversarial example
    """
    with tf.name_scope(scope, "virtual_adversarial_perturbation"):
        d = tf.random_normal(tf.shape(x), dtype=tf_dtype)
        for i in range(num_iterations):
            d = xi * utils_tf.l2_batch_normalize(d)
            logits_d = model.get_logits(x + d)
            kl = utils_tf.kl_with_logits(logits, logits_d)
            Hd = tf.gradients(kl, d)[0]
            d = tf.stop_gradient(Hd)
        d = eps * utils_tf.l2_batch_normalize(d)
        adv_x = x + d
        if (clip_min is not None) and (clip_max is not None):
            adv_x = tf.clip_by_value(adv_x, clip_min, clip_max)
        return adv_x 

def _mel_to_linear(mel_spectrogram):
    global _inv_mel_basis
    if _inv_mel_basis is None:
        _inv_mel_basis = np.linalg.pinv(_build_mel_basis())
    return np.maximum(1e-10, np.dot(_inv_mel_basis, mel_spectrogram))