Python numpy.maximum() Examples

def update_alg_state_kwargs(
        **alg_state_kwargs):
    alg_state_kwargs['cur_step'] += 1
    alg_state_kwargs['cur_lap'] = \
        alg_state_kwargs['cur_step'] / float(alg_state_kwargs['n_batches'])
    alg_state_kwargs['elapsed_time_sec'] = \
        time.time() - alg_state_kwargs['start_time_sec']
    alg_state_kwargs['cur_mem_MiB'] = getMemUsageOfCurProcess_MiB('rss')
    alg_state_kwargs['cur_swp_MiB'] = getMemUsageOfCurProcess_MiB('vms')
    alg_state_kwargs['max_mem_MiB'] = np.maximum(
        alg_state_kwargs['cur_mem_MiB'],
        alg_state_kwargs['max_mem_MiB'])
    alg_state_kwargs['max_swp_MiB'] = np.maximum(
        alg_state_kwargs['cur_swp_MiB'],
        alg_state_kwargs['max_swp_MiB'])
    return alg_state_kwargs 

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 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 _next_power_of_two(number):
    return int(2 ** np.ceil(np.log2(np.maximum(1., number)))) 

def fft_features(signal, fps, **kwargs):
    wav_chunks = _split_into_chunks(signal=signal, chunks_per_second=fps)
    fft = np.fft.fft(wav_chunks)
    sound_dB = 10 * np.log10(np.maximum(1., np.square(np.real(fft)) + np.square(np.imag(fft))))
    np.random.shuffle(sound_dB)

    return sound_dB 

def project(self, x_0):
        return x_0 if self.contains(x_0) else np.maximum(x_0, 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 __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 testAllReduce(self):
    self._Test(partial(_NcclAllReduce, nccl.all_sum), lambda x, y: x + y)
    self._Test(partial(_NcclAllReduce, nccl.all_prod), lambda x, y: x * y)
    self._Test(partial(_NcclAllReduce, nccl.all_min), np.minimum)
    self._Test(partial(_NcclAllReduce, nccl.all_max), np.maximum) 

def clip_boxes(boxes, im_shape):
    """
    Clip boxes to image boundaries.
    """

    # x1 >= 0
    boxes[:, 0::4] = np.maximum(np.minimum(boxes[:, 0::4], im_shape[1] - 1), 0)
    # y1 >= 0
    boxes[:, 1::4] = np.maximum(np.minimum(boxes[:, 1::4], im_shape[0] - 1), 0)
    # x2 < im_shape[1]
    boxes[:, 2::4] = np.maximum(np.minimum(boxes[:, 2::4], im_shape[1] - 1), 0)
    # y2 < im_shape[0]
    boxes[:, 3::4] = np.maximum(np.minimum(boxes[:, 3::4], im_shape[0] - 1), 0)
    return boxes 

def nms(dets, thresh):
    x1 = dets[:, 0]
    y1 = dets[:, 1]
    x2 = dets[:, 2]
    y2 = dets[:, 3]
    scores = dets[:, 4]

    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
    order = scores.argsort()[::-1]

    keep = []
    while order.size > 0:
        i = order[0]
        keep.append(i)
        xx1 = np.maximum(x1[i], x1[order[1:]])
        yy1 = np.maximum(y1[i], y1[order[1:]])
        xx2 = np.minimum(x2[i], x2[order[1:]])
        yy2 = np.minimum(y2[i], y2[order[1:]])

        w = np.maximum(0.0, xx2 - xx1 + 1)
        h = np.maximum(0.0, yy2 - yy1 + 1)
        inter = w * h
        ovr = inter / (areas[i] + areas[order[1:]] - inter)

        inds = np.where(ovr <= thresh)[0]
        order = order[inds + 1]

    return keep 

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) & (scores > cfg.FLAGS.DET_THRESHOLD))[0]
            dets = dets[inds, :]
            if dets == []:
                continue

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

def py_cpu_nms(dets, thresh):
    """Pure Python NMS baseline."""
    x1 = dets[:, 0]
    y1 = dets[:, 1]
    x2 = dets[:, 2]
    y2 = dets[:, 3]
    scores = dets[:, 4]

    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
    order = scores.argsort()[::-1]

    keep = []
    while order.size > 0:
        i = order[0]
        keep.append(i)
        xx1 = np.maximum(x1[i], x1[order[1:]])
        yy1 = np.maximum(y1[i], y1[order[1:]])
        xx2 = np.minimum(x2[i], x2[order[1:]])
        yy2 = np.minimum(y2[i], y2[order[1:]])

        w = np.maximum(0.0, xx2 - xx1 + 1)
        h = np.maximum(0.0, yy2 - yy1 + 1)
        inter = w * h
        ovr = inter / (areas[i] + areas[order[1:]] - inter)

        inds = np.where(ovr <= thresh)[0]
        order = order[inds + 1]

    return keep 

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.grilimn_iter):
        X_t = invert_spectrogram(X_best)
        est = librosa.stft(X_t, Hp.num_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 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 mask_target_single(pos_proposals, pos_assigned_gt_inds, gt_masks, cfg):
    mask_size = _pair(cfg.mask_size)
    num_pos = pos_proposals.size(0)
    mask_targets = []
    if num_pos > 0:
        proposals_np = pos_proposals.cpu().numpy()
        _, maxh, maxw = gt_masks.shape
        proposals_np[:, [0, 2]] = np.clip(proposals_np[:, [0, 2]], 0, maxw - 1)
        proposals_np[:, [1, 3]] = np.clip(proposals_np[:, [1, 3]], 0, maxh - 1)
        pos_assigned_gt_inds = pos_assigned_gt_inds.cpu().numpy()
        for i in range(num_pos):
            gt_mask = gt_masks[pos_assigned_gt_inds[i]]
            bbox = proposals_np[i, :].astype(np.int32)
            x1, y1, x2, y2 = bbox
            w = np.maximum(x2 - x1 + 1, 1)
            h = np.maximum(y2 - y1 + 1, 1)
            # mask is uint8 both before and after resizing
            # mask_size (h, w) to (w, h)
            target = mmcv.imresize(gt_mask[y1:y1 + h, x1:x1 + w],
                                   mask_size[::-1])
            mask_targets.append(target)
        mask_targets = torch.from_numpy(np.stack(mask_targets)).float().to(
            pos_proposals.device)
    else:
        mask_targets = pos_proposals.new_zeros((0, ) + mask_size)
    return mask_targets 

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)) 

def _amp_to_db(x):
    return 20 * np.log10(np.maximum(1e-5, x)) 

def relu(vector):
    return np.maximum(vector, 0) 

def relu(vector):
    return np.maximum(vector, 0) 

def ReLU(x):
    return np.maximum(0, x) 

def relu(x):
    return np.maximum(0, x)

# 广播 

def relu(x):
    return np.maximum(0, x) 

def pad_img_to_fit_bbox(img, x1, x2, y1, y2):
    if img.ndim==3:
        img = np.pad(img, ((np.abs(np.minimum(0, y1)), np.maximum(y2 - img.shape[0], 0)),
               (np.abs(np.minimum(0, x1)), np.maximum(x2 - img.shape[1], 0)), (0,0)), mode="constant")
    else:
        img = np.pad(img, ((np.abs(np.minimum(0, y1)), np.maximum(y2 - img.shape[0], 0)),
                  (np.abs(np.minimum(0, x1)), np.maximum(x2 - img.shape[1], 0))), mode="constant")

    y1 += np.abs(np.minimum(0, y1))
    y2 += np.abs(np.minimum(0, y1))
    x1 += np.abs(np.minimum(0, x1))
    x2 += np.abs(np.minimum(0, x1))
    return img, x1, x2, y1, y2 

def forward(self, is_train, req, in_data, out_data, aux):
        x = in_data[0]
        y = nd.maximum(x, nd.zeros_like(x))
        self.assign(out_data[0], req[0], y) 

def get_cam(imggrad, conv_out):
    """Compute CAM. Refer section 3 of https://arxiv.org/abs/1610.02391 for details"""
    weights = np.mean(imggrad, axis=(1, 2))
    cam = np.ones(conv_out.shape[1:], dtype=np.float32)
    for i, w in enumerate(weights):
        cam += w * conv_out[i, :, :]
    cam = cv2.resize(cam, (imggrad.shape[1], imggrad.shape[2]))
    cam = np.maximum(cam, 0)
    cam = (cam - np.min(cam)) / (np.max(cam) - np.min(cam)) 
    cam = np.uint8(cam * 255)
    return cam 

def _average_precision(self, rec, prec):
        """
        calculate average precision

        Params:
        ----------
        rec : numpy.array
            cumulated recall
        prec : numpy.array
            cumulated precision
        Returns:
        ----------
        ap as float
        """
        # 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 voc_ap(rec, prec, use_07_metric=False):
    """
    average precision calculations
    [precision integrated to recall]
    :param rec: recall
    :param prec: precision
    :param use_07_metric: 2007 metric is 11-recall-point based AP
    :return: average precision
    """
    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 nms(dets, thresh):
    """
    greedily select boxes with high confidence and overlap with current maximum <= thresh
    rule out overlap >= thresh
    :param dets: [[x1, y1, x2, y2 score]]
    :param thresh: retain overlap < thresh
    :return: indexes to keep
    """
    x1 = dets[:, 0]
    y1 = dets[:, 1]
    x2 = dets[:, 2]
    y2 = dets[:, 3]
    scores = dets[:, 4]

    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
    order = scores.argsort()[::-1]

    keep = []
    while order.size > 0:
        i = order[0]
        keep.append(i)
        xx1 = np.maximum(x1[i], x1[order[1:]])
        yy1 = np.maximum(y1[i], y1[order[1:]])
        xx2 = np.minimum(x2[i], x2[order[1:]])
        yy2 = np.minimum(y2[i], y2[order[1:]])

        w = np.maximum(0.0, xx2 - xx1 + 1)
        h = np.maximum(0.0, yy2 - yy1 + 1)
        inter = w * h
        ovr = inter / (areas[i] + areas[order[1:]] - inter)

        inds = np.where(ovr <= thresh)[0]
        order = order[inds + 1]

    return keep 

def _calculate_areas(self, label):
        """Calculate areas for multiple labels"""
        heights = np.maximum(0, label[:, 3] - label[:, 1])
        widths = np.maximum(0, label[:, 2] - label[:, 0])
        return heights * widths 

def _intersect(self, label, xmin, ymin, xmax, ymax):
        """Calculate intersect areas, normalized."""
        left = np.maximum(label[:, 0], xmin)
        right = np.minimum(label[:, 2], xmax)
        top = np.maximum(label[:, 1], ymin)
        bot = np.minimum(label[:, 3], ymax)
        invalid = np.where(np.logical_or(left >= right, top >= bot))[0]
        out = label.copy()
        out[:, 0] = left
        out[:, 1] = top
        out[:, 2] = right
        out[:, 3] = bot
        out[invalid, :] = 0
        return out 

def _generate_objects():
    num = np.random.randint(1, 10)
    xy = np.random.rand(num, 2)
    wh = np.random.rand(num, 2) / 2
    left = (xy[:, 0] - wh[:, 0])[:, np.newaxis]
    right = (xy[:, 0] + wh[:, 0])[:, np.newaxis]
    top = (xy[:, 1] - wh[:, 1])[:, np.newaxis]
    bot = (xy[:, 1] + wh[:, 1])[:, np.newaxis]
    boxes = np.maximum(0., np.minimum(1., np.hstack((left, top, right, bot))))
    cid = np.random.randint(0, 20, size=num)
    label = np.hstack((cid[:, np.newaxis], boxes)).ravel().tolist()
    return [2, 5] + label 

def test_relu():
    def frelu(x):
        return np.maximum(x, 0.0)
    def frelu_grad(x):
        return 1.0 * (x > 0.0)
    shape = (3, 4)
    x = mx.symbol.Variable("x")
    y = mx.sym.relu(x)
    xa = np.random.uniform(low=-1.0,high=1.0,size=shape)
    eps = 1e-4
    # Avoid finite difference method inaccuracies due to discontinuous gradient at the origin.
    # Here we replace small problematic inputs with 1.0.  Repro issue with seed 97264195.
    xa[abs(xa) < eps] = 1.0
    ya = frelu(xa)
    ga = frelu_grad(xa)
    check_numeric_gradient(y, [xa], numeric_eps=eps)
    check_symbolic_forward(y, [xa], [ya])
    check_symbolic_backward(y, [xa], [np.ones(shape)], [ga])


# NOTE(haojin2): Skipping the numeric check tests for float16 data type due to precision issues,
# the analytical checks are still performed on each and every data type to verify the correctness. 

def test_hard_sigmoid():
    def fhardsigmoid(a, alpha=0.2, beta=0.5):
        return np.maximum(np.zeros(a.shape, dtype=a.dtype),
                          np.minimum(np.ones(a.shape, dtype=a.dtype), alpha*a+beta))
    def fhardsigmoid_grad(a, out_grad, alpha=0.2, beta=0.5):
        orig_out = fhardsigmoid(a, alpha, beta)
        res = out_grad * alpha
        res[orig_out <= 0.0] = 0.0
        res[orig_out >= 1.0] = 0.0
        return res
    shape = (3, 4)
    x = mx.symbol.Variable("x")
    y = mx.sym.hard_sigmoid(x)
    for dtype in [np.float16, np.float32, np.float64]:
        if dtype is np.float16:
            rtol = 1e-2
        else:
            rtol = 1e-3
        atol = 1e-3
        eps = 1e-3
        xa = np.random.uniform(low=-3.0,high=3.0,size=shape).astype(dtype)
        # function not differentiable at x=2.5 and -2.5
        xa[abs(xa-2.5) < eps] -= 2 * eps
        xa[abs(xa+2.5) < eps] += 2 * eps
        ya = fhardsigmoid(xa)
        grad_xa = fhardsigmoid_grad(xa, np.ones(shape))
        if dtype is not np.float16:
            check_numeric_gradient(y, [xa], numeric_eps=eps, rtol=rtol, atol=atol, dtype=dtype)
        check_symbolic_forward(y, [xa], [ya], rtol=rtol, atol=atol, dtype=dtype)
        check_symbolic_backward(y, [xa], [np.ones(shape)], [grad_xa], rtol=rtol, atol=atol, dtype=dtype) 

def test_maximum_minimum():
    data1 = mx.symbol.Variable('data')
    data2 = mx.symbol.Variable('data')
    shape = (3, 4)
    data_tmp1 = np.random.rand(3,4)
    data_tmp2 = np.random.rand(3,4)
    data_tmp1[:] = 2
    data_tmp2[:] = 3

    arr_data1 = mx.nd.array(data_tmp1)
    arr_data2 = mx.nd.array(data_tmp2)

    arr_grad1 = mx.nd.empty(shape)
    arr_grad2 = mx.nd.empty(shape)

    test =  mx.sym.maximum(data1,data2) + mx.sym.minimum(data1,data2);
    exe_test = test.bind(default_context(), args=[arr_data1,arr_data2], args_grad=[arr_grad1,arr_grad2])
    exe_test.forward(is_train=True)
    out = exe_test.outputs[0].asnumpy()
    npout =  np.maximum(data_tmp1,data_tmp2) + np.minimum(data_tmp1,data_tmp2)
    assert_almost_equal(out, npout)

    out_grad = mx.nd.empty(shape)
    out_grad[:] = 2
    exe_test.backward(out_grad)

    npout_grad = np.ones(shape)
    npout_grad[:] = 2
    mask1 = (data_tmp1 > data_tmp2).astype('float')
    mask2 = (data_tmp1 < data_tmp2).astype('float')
    npout_grad1 = npout_grad * mask1 + npout_grad * mask2
    npout_grad2 = (npout_grad - npout_grad * mask1) + (npout_grad - npout_grad * mask2)

    assert_almost_equal(arr_grad1.asnumpy(), npout_grad1)
    assert_almost_equal(arr_grad2.asnumpy(), npout_grad2) 

def test_maximum_minimum_scalar():
    data1 = mx.symbol.Variable('data')
    shape = (3, 4)
    data_tmp1 = np.random.rand(3,4)
    data_tmp1[:] = 2

    arr_data1 = mx.nd.array(data_tmp1)
    arr_grad1 = mx.nd.empty(shape)

    test =  mx.sym.maximum(data1,3) + mx.sym.maximum(9,data1) + mx.sym.minimum(5,data1) + mx.sym.minimum(data1,4)
    exe_test = test.bind(default_context(), args=[arr_data1], args_grad=[arr_grad1])
    exe_test.forward(is_train=True)
    out = exe_test.outputs[0].asnumpy()
    npout =  np.maximum(data_tmp1,3) + np.maximum(9,data_tmp1) + np.minimum(5,data_tmp1) + np.minimum(data_tmp1,4)
    assert_almost_equal(out, npout)

    out_grad = mx.nd.empty(shape)
    out_grad[:] = 2
    exe_test.backward(out_grad)

    npout_grad = np.ones(shape)
    npout_grad[:] = 2
    mask1 = (data_tmp1 > 3).astype('float')
    mask2 = (9 > data_tmp1).astype('float')
    mask3 = (5 < data_tmp1).astype('float')
    mask4 = (data_tmp1 < 4).astype('float')
    npout_grad1 = npout_grad * mask1 + (npout_grad - npout_grad * mask2) + (npout_grad - npout_grad * mask3) + npout_grad * mask4

    assert_almost_equal(arr_grad1.asnumpy(), npout_grad1) 

def test_bind():
    def check_bind(disable_bulk_exec):
        if disable_bulk_exec:
            prev_bulk_inf_val = mx.test_utils.set_env_var("MXNET_EXEC_BULK_EXEC_INFERENCE", "0", "1")
            prev_bulk_train_val = mx.test_utils.set_env_var("MXNET_EXEC_BULK_EXEC_TRAIN", "0", "1")

        nrepeat = 10
        maxdim = 4
        for repeat in range(nrepeat):
            for dim in range(1, maxdim):
                check_bind_with_uniform(lambda x, y: x + y,
                                        lambda g, x, y: (g, g),
                                        dim)
                check_bind_with_uniform(lambda x, y: x - y,
                                        lambda g, x, y: (g, -g),
                                        dim)
                check_bind_with_uniform(lambda x, y: x * y,
                                        lambda g, x, y: (y * g, x * g),
                                        dim)
                check_bind_with_uniform(lambda x, y: x / y,
                                        lambda g, x, y: (g / y, -x * g/ (y**2)),
                                        dim)

                check_bind_with_uniform(lambda x, y: np.maximum(x, y),
                                        lambda g, x, y: (g * (x>=y), g * (y>x)),
                                        dim,
                                        sf=mx.symbol.maximum)
                check_bind_with_uniform(lambda x, y: np.minimum(x, y),
                                        lambda g, x, y: (g * (x<=y), g * (y<x)),
                                        dim,
                                        sf=mx.symbol.minimum)
        if disable_bulk_exec:
           mx.test_utils.set_env_var("MXNET_EXEC_BULK_EXEC_INFERENCE", prev_bulk_inf_val)
           mx.test_utils.set_env_var("MXNET_EXEC_BULK_EXEC_TRAIN", prev_bulk_train_val)

    check_bind(True)
    check_bind(False)


# @roywei: Removing fixed seed as flakiness in this test is fixed
# tracked at https://github.com/apache/incubator-mxnet/issues/11686 

def _average_precision(self, rec, prec):
        """
        calculate average precision

        Params:
        ----------
        rec : numpy.array
            cumulated recall
        prec : numpy.array
            cumulated precision
        Returns:
        ----------
        ap as float
        """
        if rec is None or prec is None:
            return np.nan

        # append sentinel values at both ends
        mrec = np.concatenate(([0.], rec, [1.]))
        mpre = np.concatenate(([0.], np.nan_to_num(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 bbox_iou(bbox_a, bbox_b, offset=0):
    """Calculate Intersection-Over-Union(IOU) of two bounding boxes.

    Parameters
    ----------
    bbox_a : numpy.ndarray
        An ndarray with shape :math:`(N, 4)`.
    bbox_b : numpy.ndarray
        An ndarray with shape :math:`(M, 4)`.
    offset : float or int, default is 0
        The ``offset`` is used to control the whether the width(or height) is computed as
        (right - left + ``offset``).
        Note that the offset must be 0 for normalized bboxes, whose ranges are in ``[0, 1]``.

    Returns
    -------
    numpy.ndarray
        An ndarray with shape :math:`(N, M)` indicates IOU between each pairs of
        bounding boxes in `bbox_a` and `bbox_b`.

    """
    if bbox_a.shape[1] < 4 or bbox_b.shape[1] < 4:
        raise IndexError("Bounding boxes axis 1 must have at least length 4")

    tl = np.maximum(bbox_a[:, None, :2], bbox_b[:, :2])
    br = np.minimum(bbox_a[:, None, 2:4], bbox_b[:, 2:4])

    area_i = np.prod(br - tl + offset, axis=2) * (tl < br).all(axis=2)
    area_a = np.prod(bbox_a[:, 2:4] - bbox_a[:, :2] + offset, axis=1)
    area_b = np.prod(bbox_b[:, 2:4] - bbox_b[:, :2] + offset, axis=1)
    return area_i / (area_a[:, None] + area_b - area_i) 

def bbox_xywh_to_xyxy(xywh):
    """Convert bounding boxes from format (x, y, w, h) to (xmin, ymin, xmax, ymax)

    Parameters
    ----------
    xywh : list, tuple or numpy.ndarray
        The bbox in format (x, y, w, h).
        If numpy.ndarray is provided, we expect multiple bounding boxes with
        shape `(N, 4)`.

    Returns
    -------
    tuple or numpy.ndarray
        The converted bboxes in format (xmin, ymin, xmax, ymax).
        If input is numpy.ndarray, return is numpy.ndarray correspondingly.

    """
    if isinstance(xywh, (tuple, list)):
        if not len(xywh) == 4:
            raise IndexError(
                "Bounding boxes must have 4 elements, given {}".format(len(xywh)))
        w, h = np.maximum(xywh[2] - 1, 0), np.maximum(xywh[3] - 1, 0)
        return (xywh[0], xywh[1], xywh[0] + w, xywh[1] + h)
    elif isinstance(xywh, np.ndarray):
        if not xywh.size % 4 == 0:
            raise IndexError(
                "Bounding boxes must have n * 4 elements, given {}".format(xywh.shape))
        xyxy = np.hstack((xywh[:, :2], xywh[:, :2] + np.maximum(0, xywh[:, 2:4] - 1)))
        return xyxy
    else:
        raise TypeError(
            'Expect input xywh a list, tuple or numpy.ndarray, given {}'.format(type(xywh))) 

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 box_iou(b1, b2):
    '''Return iou tensor

    Parameters
    ----------
    b1: tensor, shape=(i1,...,iN, 4), xywh
    b2: tensor, shape=(j, 4), xywh

    Returns
    -------
    iou: tensor, shape=(i1,...,iN, j)

    '''

    # Expand dim to apply broadcasting.
    b1 = K.expand_dims(b1, -2)
    b1_xy = b1[..., :2]
    b1_wh = b1[..., 2:4]
    b1_wh_half = b1_wh/2.
    b1_mins = b1_xy - b1_wh_half
    b1_maxes = b1_xy + b1_wh_half

    # Expand dim to apply broadcasting.
    b2 = K.expand_dims(b2, 0)
    b2_xy = b2[..., :2]
    b2_wh = b2[..., 2:4]
    b2_wh_half = b2_wh/2.
    b2_mins = b2_xy - b2_wh_half
    b2_maxes = b2_xy + b2_wh_half

    intersect_mins = K.maximum(b1_mins, b2_mins)
    intersect_maxes = K.minimum(b1_maxes, b2_maxes)
    intersect_wh = K.maximum(intersect_maxes - intersect_mins, 0.)
    intersect_area = intersect_wh[..., 0] * intersect_wh[..., 1]
    b1_area = b1_wh[..., 0] * b1_wh[..., 1]
    b2_area = b2_wh[..., 0] * b2_wh[..., 1]
    iou = intersect_area / (b1_area + b2_area - intersect_area)

    return iou 

def _relu(x, deriv=False):
        """
        Rectified linear unit activation function
        """
        if deriv:
            return 1.0 * (x >= 0)
        return np.maximum(x, 0) 

def _leakyrelu(x, deriv=False):
        """
        Rectified linear unit activation function with small value for negatives
        """
        if deriv:
            return 1.0 * (x >= 0) - 0.01 * (x < 0)
        return np.maximum(x, 0.01 * x) 

def compute_losses_multi_or(logits, actions_one_hot, weights=None,
                            num_actions=-1, data_loss_wt=1., reg_loss_wt=1.,
                            ewma_decay=0.99, reg_loss_op=None):
  assert(num_actions > 0), 'num_actions must be specified and must be > 0.'
  
  with tf.name_scope('loss'):
    if weights is None:
      weight = tf.ones_like(actions_one_hot, dtype=tf.float32, name='weight')
    
    actions_one_hot = tf.cast(tf.reshape(actions_one_hot, [-1, num_actions],
                                         're_actions_one_hot'), tf.float32)
    weights = tf.reduce_sum(tf.reshape(weights, [-1, num_actions], 're_weight'),
                            reduction_indices=1)
    total = tf.reduce_sum(weights)

    action_prob = tf.nn.softmax(logits)
    action_prob = tf.reduce_sum(tf.multiply(action_prob, actions_one_hot),
                                reduction_indices=1)
    example_loss = -tf.log(tf.maximum(tf.constant(1e-4), action_prob))

    data_loss_op = tf.reduce_sum(example_loss * weights) / total
    if reg_loss_op is None:
      if reg_loss_wt > 0:
        reg_loss_op = tf.add_n(tf.losses.get_regularization_losses())
      else:
        reg_loss_op = tf.constant(0.)
    
    if reg_loss_wt > 0:
      total_loss_op = data_loss_wt*data_loss_op + reg_loss_wt*reg_loss_op 
    else:
      total_loss_op = data_loss_wt*data_loss_op

    is_correct = tf.cast(tf.greater(action_prob, 0.5, name='pred_class'), tf.float32)
    acc_op = tf.reduce_sum(is_correct*weights) / total

    ewma_acc_op = moving_averages.weighted_moving_average(
        acc_op, ewma_decay, weight=total, name='ewma_acc')

    acc_ops = [ewma_acc_op]

  return reg_loss_op, data_loss_op, total_loss_op, acc_ops 

def __init__(self, obj_file, material_file=None, load_materials=True,
               name_prefix='', name_suffix=''):
    if material_file is not None:
      logging.error('Ignoring material file input, reading them off obj file.')
    load_flags = self.get_pyassimp_load_options()
    scene = assimp.load(obj_file, processing=load_flags)
    filter_ind = self._filter_triangles(scene.meshes)
    self.meshes = [scene.meshes[i] for i in filter_ind]
    for m in self.meshes:
      m.name = name_prefix + m.name + name_suffix

    dir_name = os.path.dirname(obj_file)
    # Load materials
    materials = None
    if load_materials:
      materials = []
      for m in self.meshes:
        file_name = os.path.join(dir_name, m.material.properties[('file', 1)])
        assert(os.path.exists(file_name)), \
            'Texture file {:s} foes not exist.'.format(file_name)
        img_rgb = cv2.imread(file_name)[::-1,:,::-1]
        if img_rgb.shape[0] != img_rgb.shape[1]:
          logging.warn('Texture image not square.')
          sz = np.maximum(img_rgb.shape[0], img_rgb.shape[1])
          sz = int(np.power(2., np.ceil(np.log2(sz))))
          img_rgb = cv2.resize(img_rgb, (sz,sz), interpolation=cv2.INTER_LINEAR)
        else:
          sz = img_rgb.shape[0]
          sz_ = int(np.power(2., np.ceil(np.log2(sz))))
          if sz != sz_:
            logging.warn('Texture image not square of power of 2 size. ' +
                         'Changing size from %d to %d.', sz, sz_)
            sz = sz_
            img_rgb = cv2.resize(img_rgb, (sz,sz), interpolation=cv2.INTER_LINEAR)
        materials.append(img_rgb)
    self.scene = scene
    self.materials = materials 

def image_whitening(data):
  """
  Subtracts mean of image and divides by adjusted standard variance (for
  stability). Operations are per image but performed for the entire array.
  :param image: 4D array (ID, Height, Weight, Channel)
  :return: 4D array (ID, Height, Weight, Channel)
  """
  assert len(np.shape(data)) == 4

  # Compute number of pixels in image
  nb_pixels = np.shape(data)[1] * np.shape(data)[2] * np.shape(data)[3]

  # Subtract mean
  mean = np.mean(data, axis=(1,2,3))

  ones = np.ones(np.shape(data)[1:4], dtype=np.float32)
  for i in xrange(len(data)):
    data[i, :, :, :] -= mean[i] * ones

  # Compute adjusted standard variance
  adj_std_var = np.maximum(np.ones(len(data), dtype=np.float32) / math.sqrt(nb_pixels), np.std(data, axis=(1,2,3))) #NOLINT(long-line)

  # Divide image
  for i in xrange(len(data)):
    data[i, :, :, :] = data[i, :, :, :] / adj_std_var[i]

  print(np.shape(data))

  return data 

def get_spectrograms(fpath):
    '''Returns normalized log(melspectrogram) and log(magnitude) from `sound_file`.
    Args:
      sound_file: A string. The full path of a sound file.

    Returns:
      mel: A 2d array of shape (T, n_mels) <- Transposed
      mag: A 2d array of shape (T, 1+n_fft/2) <- Transposed
    '''
    # Loading sound file
    y, sr = librosa.load(fpath, sr=Hp.sample_rate)

    # Trimming
    y, _ = librosa.effects.trim(y)

    # Preemphasis
    y = np.append(y[0], y[1:] - Hp.preemphasis * y[:-1])

    # stft
    linear = librosa.stft(y=y,
                          n_fft=Hp.num_fft,
                          hop_length=Hp.hop_length,
                          win_length=Hp.win_length)

    # magnitude spectrogram
    mag = np.abs(linear)  # (1+n_fft//2, T)

    # mel spectrogram
    mel_basis = librosa.filters.mel(Hp.sample_rate, Hp.num_fft, Hp.num_mels)  # (n_mels, 1+n_fft//2)
    mel = np.dot(mel_basis, mag)  # (n_mels, t)

    # to decibel
    mel = 20 * np.log10(np.maximum(1e-5, mel))
    mag = 20 * np.log10(np.maximum(1e-5, mag))

    # normalize
    mel = np.clip((mel - Hp.ref_db + Hp.max_db) / Hp.max_db, 1e-8, 1)
    mag = np.clip((mag - Hp.ref_db + Hp.max_db) / Hp.max_db, 1e-8, 1)

    # Transpose
    mel = mel.T.astype(np.float32)  # (T, n_mels)
    mag = mag.T.astype(np.float32)  # (T, 1+n_fft//2)

    return mel, mag 

def get_spectrograms(fpath):
    '''Parse the wave file in `fpath` and
    Returns normalized melspectrogram and linear spectrogram.

    Args:
      fpath: A string. The full path of a sound file.

    Returns:
      mel: A 2d array of shape (T, n_mels) and dtype of float32.
      mag: A 2d array of shape (T, 1+n_fft/2) and dtype of float32.
    '''
    # Loading sound file
    y, sr = librosa.load(fpath, sr=hp.sr)

    # Trimming
    y, _ = librosa.effects.trim(y)

    # Preemphasis
    y = np.append(y[0], y[1:] - hp.preemphasis * y[:-1])

    # stft
    linear = librosa.stft(y=y,
                          n_fft=hp.n_fft,
                          hop_length=hp.hop_length,
                          win_length=hp.win_length)

    # magnitude spectrogram
    mag = np.abs(linear)  # (1+n_fft//2, T)

    # mel spectrogram
    mel_basis = librosa.filters.mel(hp.sr, hp.n_fft, hp.n_mels)  # (n_mels, 1+n_fft//2)
    mel = np.dot(mel_basis, mag)  # (n_mels, t)

    # to decibel
    mel = 20 * np.log10(np.maximum(1e-5, mel))
    mag = 20 * np.log10(np.maximum(1e-5, mag))

    # normalize
    mel = np.clip((mel - hp.ref_db + hp.max_db) / hp.max_db, 1e-8, 1)
    mag = np.clip((mag - hp.ref_db + hp.max_db) / hp.max_db, 1e-8, 1)

    # Transpose
    mel = mel.T.astype(np.float32)  # (T, n_mels)
    mag = mag.T.astype(np.float32)  # (T, 1+n_fft//2)

    return mel, mag 

def bbox_overlaps(bboxes1, bboxes2, mode='iou'):
    """Calculate the ious between each bbox of bboxes1 and bboxes2.

    Args:
        bboxes1(ndarray): shape (n, 4)
        bboxes2(ndarray): shape (k, 4)
        mode(str): iou (intersection over union) or iof (intersection
            over foreground)

    Returns:
        ious(ndarray): shape (n, k)
    """

    assert mode in ['iou', 'iof']

    bboxes1 = bboxes1.astype(np.float32)
    bboxes2 = bboxes2.astype(np.float32)
    rows = bboxes1.shape[0]
    cols = bboxes2.shape[0]
    ious = np.zeros((rows, cols), dtype=np.float32)
    if rows * cols == 0:
        return ious
    exchange = False
    if bboxes1.shape[0] > bboxes2.shape[0]:
        bboxes1, bboxes2 = bboxes2, bboxes1
        ious = np.zeros((cols, rows), dtype=np.float32)
        exchange = True
    area1 = (bboxes1[:, 2] - bboxes1[:, 0] + 1) * (
        bboxes1[:, 3] - bboxes1[:, 1] + 1)
    area2 = (bboxes2[:, 2] - bboxes2[:, 0] + 1) * (
        bboxes2[:, 3] - bboxes2[:, 1] + 1)
    for i in range(bboxes1.shape[0]):
        x_start = np.maximum(bboxes1[i, 0], bboxes2[:, 0])
        y_start = np.maximum(bboxes1[i, 1], bboxes2[:, 1])
        x_end = np.minimum(bboxes1[i, 2], bboxes2[:, 2])
        y_end = np.minimum(bboxes1[i, 3], bboxes2[:, 3])
        overlap = np.maximum(x_end - x_start + 1, 0) * np.maximum(
            y_end - y_start + 1, 0)
        if mode == 'iou':
            union = area1[i] + area2 - overlap
        else:
            union = area1[i] if not exchange else area2
        ious[i, :] = overlap / union
    if exchange:
        ious = ious.T
    return ious 

def average_precision(recalls, precisions, mode='area'):
    """Calculate average precision (for single or multiple scales).

    Args:
        recalls (ndarray): shape (num_scales, num_dets) or (num_dets, )
        precisions (ndarray): shape (num_scales, num_dets) or (num_dets, )
        mode (str): 'area' or '11points', 'area' means calculating the area
            under precision-recall curve, '11points' means calculating
            the average precision of recalls at [0, 0.1, ..., 1]

    Returns:
        float or ndarray: calculated average precision
    """
    no_scale = False
    if recalls.ndim == 1:
        no_scale = True
        recalls = recalls[np.newaxis, :]
        precisions = precisions[np.newaxis, :]
    assert recalls.shape == precisions.shape and recalls.ndim == 2
    num_scales = recalls.shape[0]
    ap = np.zeros(num_scales, dtype=np.float32)
    if mode == 'area':
        zeros = np.zeros((num_scales, 1), dtype=recalls.dtype)
        ones = np.ones((num_scales, 1), dtype=recalls.dtype)
        mrec = np.hstack((zeros, recalls, ones))
        mpre = np.hstack((zeros, precisions, zeros))
        for i in range(mpre.shape[1] - 1, 0, -1):
            mpre[:, i - 1] = np.maximum(mpre[:, i - 1], mpre[:, i])
        for i in range(num_scales):
            ind = np.where(mrec[i, 1:] != mrec[i, :-1])[0]
            ap[i] = np.sum(
                (mrec[i, ind + 1] - mrec[i, ind]) * mpre[i, ind + 1])
    elif mode == '11points':
        for i in range(num_scales):
            for thr in np.arange(0, 1 + 1e-3, 0.1):
                precs = precisions[i, recalls[i, :] >= thr]
                prec = precs.max() if precs.size > 0 else 0
                ap[i] += prec
            ap /= 11
    else:
        raise ValueError(
            'Unrecognized mode, only "area" and "11points" are supported')
    if no_scale:
        ap = ap[0]
    return ap 

def fgsm(x, predictions, eps, clip_min=None, clip_max=None, y=None):
    """
    Computes symbolic TF tensor for the adversarial samples. This must
    be evaluated with a session.run call.
    :param x: the input placeholder
    :param predictions: the model's output tensor
    :param eps: the epsilon (input variation 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 y: the output placeholder. Use None (the default) to avoid the
            label leaking effect.
    :return: a tensor for the adversarial example
    """

    # Compute loss
    if y is None:
        # In this case, use model predictions as ground truth
        y = tf.to_float(
            tf.equal(predictions,
                     tf.reduce_max(predictions, 1, keep_dims=True)))
    y = y / tf.reduce_sum(y, 1, keep_dims=True)
    logits, = predictions.op.inputs
    loss = tf.reduce_mean(
        tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=y)
    )

    # Define gradient of loss wrt input
    grad, = tf.gradients(loss, x)

    # Take sign of gradient
    signed_grad = tf.sign(grad)

    # Multiply by constant epsilon
    scaled_signed_grad = eps * signed_grad

    # Add perturbation to original example to obtain adversarial example
    adv_x = tf.stop_gradient(x + scaled_signed_grad)

    # If clipping is needed, reset all values outside of [clip_min, clip_max]
    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 basic_iterative_method(sess, model, X, Y, eps, eps_iter, nb_iter=50,
                           clip_min=None, clip_max=None, batch_size=256):
    """
    TODO
    :param sess:
    :param model: predictions or after-softmax
    :param X:
    :param Y:
    :param eps:
    :param eps_iter:
    :param nb_iter:
    :param clip_min:
    :param clip_max:
    :param batch_size:
    :return:
    """
    print("nb_iter",nb_iter)
    # Define TF placeholders for the input and output
    x = tf.placeholder(tf.float32, shape=(None,)+X.shape[1:])
    y = tf.placeholder(tf.float32, shape=(None,)+Y.shape[1:])
    # results will hold the adversarial inputs at each iteration of BIM;
    # thus it will have shape (nb_iter, n_samples, n_rows, n_cols, n_channels)
    results = np.zeros((nb_iter, X.shape[0],) + X.shape[1:])
    # Initialize adversarial samples as the original samples, set upper and
    # lower bounds
    X_adv = X
    X_min = X_adv - eps
    X_max = X_adv + eps
    print('Running BIM iterations...')
    # "its" is a dictionary that keeps track of the iteration at which each
    # sample becomes misclassified. The default value will be (nb_iter-1), the
    # very last iteration.
    def f(val):
        return lambda: val
    its = defaultdict(f(nb_iter-1))
    # Out keeps track of which samples have already been misclassified
    out = set()
    for i in tqdm(range(nb_iter)):
        adv_x = fgsm(
            x, model(x), eps=eps_iter,
            clip_min=clip_min, clip_max=clip_max, y=y
        )
        X_adv, = batch_eval(
            sess, [x, y], [adv_x],
            [X_adv, Y], feed={K.learning_phase(): 0},
            args={'batch_size': batch_size}
        )
        X_adv = np.maximum(np.minimum(X_adv, X_max), X_min)
        results[i] = X_adv
        # check misclassifieds
        predictions = model.predict_classes(X_adv, batch_size=512, verbose=0)
        misclassifieds = np.where(predictions != Y.argmax(axis=1))[0]
        for elt in misclassifieds:
            if elt not in out:
                its[elt] = i
                out.add(elt)

    return its, results 

def __init__(self, sess, x, model_preds, targeted_label,
                 binary_search_steps, max_iterations, initial_const, clip_min,
                 clip_max, nb_classes, batch_size):
        """
        Return a tensor that constructs adversarial examples for the given
        input. Generate uses tf.py_func in order to operate over tensors.

        :param sess: a TF session.
        :param x: A tensor with the inputs.
        :param model_preds: A tensor with model's predictions.
        :param targeted_label: A tensor with the target labels.
        :param binary_search_steps: The number of times we perform binary
                                    search to find the optimal tradeoff-
                                    constant between norm of the purturbation
                                    and cross-entropy loss of classification.
        :param max_iterations: The maximum number of iterations.
        :param initial_const: The initial tradeoff-constant to use to tune the
                              relative importance of size of the purturbation
                              and cross-entropy loss of the classification.
        :param clip_min: Minimum input component value
        :param clip_max: Maximum input component value
        :param num_labels: The number of classes in the model's output.
        :param batch_size: Number of attacks to run simultaneously.

        """
        self.sess = sess
        self.x = x
        self.model_preds = model_preds
        self.targeted_label = targeted_label
        self.binary_search_steps = binary_search_steps
        self.max_iterations = max_iterations
        self.initial_const = initial_const
        self.clip_min = clip_min
        self.clip_max = clip_max
        self.batch_size = batch_size

        self.repeat = self.binary_search_steps >= 10
        self.shape = shape = tuple([self.batch_size] +
                                   list(self.x.get_shape().as_list()[1:]))
        self.ori_img = tf.Variable(
            np.zeros(self.shape), dtype=tf_dtype, name='ori_img')
        self.const = tf.Variable(
            np.zeros(self.batch_size), dtype=tf_dtype, name='const')

        self.score = loss_module.attack_softmax_cross_entropy(
            self.targeted_label, self.model_preds, mean=False)
        self.l2dist = reduce_sum(tf.square(self.x - self.ori_img))
        # small self.const will result small adversarial perturbation
        self.loss = reduce_sum(self.score * self.const) + self.l2dist
        self.grad, = tf.gradients(self.loss, self.x) 

def train(self, data, targets):
        # now that we have the data, we know the shape of the weight vector
        self.coefs = np.zeros(data.shape[1])
        # generate holdout set
        train_data, holdout_data, train_targets, holdout_targets = train_test_split(
            data, targets, test_size=self.holdout_proportion)

        if self.normalize_data:
            train_mean = np.mean(train_data, axis=0)
            train_std = np.std(train_data, axis=0)
            train_data = normalize(train_data, train_mean, train_std)
            if self.holdout_proportion:
                holdout_data = normalize(holdout_data, train_mean, train_std)

        for epoch in range(self.n_epochs):
            if epoch > 0:
                # randomize order for each epoch
                train_data, train_targets = shuffle_rows(train_data, train_targets)

            for batch_data, batch_targets in self.get_minibatches(train_data, train_targets):
                # evalute the gradient on this minibatch with the current coefs
                w_gradient, b_gradient = self.loss.gradient(batch_data,
                                            self.predict(batch_data), batch_targets)
                self.coefs -= self.learning_rate * w_gradient
                self.bias -= self.learning_rate * b_gradient
                # TODO: add learning rate decay
                # TODO: add momentum, rmsprop, etc.

                # regularization
                # I'm not regularizing the bias parameter
                if self.l2:
                    self.coefs -= 2. * self.l2 * self.coefs
                if self.l1:
                    self.coefs = np.sign(self.coefs) * np.maximum(
                        0.0, np.absolute(self.coefs) - self.l1)

            # report after every 2^(n-1) epoch and at the end of training
            if self.verbose and self.holdout_proportion and (
                (epoch & (epoch - 1)) == 0 or epoch == (self.n_epochs - 1)
                ):
                # evaluate holdout set w/ current coefs
                holdout_loss = self.loss.loss(holdout_targets,
                                    self.predict(holdout_data))
                sgd_report(
                    epoch=1 + epoch,
                    loss=holdout_loss,
                    br=np.mean(holdout_targets) if isinstance(self.loss, Logistic) else "",
                    bias=self.bias
                ) 

def voc_eval(class_anno, npos, image_ids, bbox, confidence, ovthresh=0.5, use_07_metric=False):
        # sort by confidence
        if bbox.shape[0] > 0:
            sorted_inds = np.argsort(-confidence)
            sorted_scores = np.sort(-confidence)
            bbox = bbox[sorted_inds, :]
            image_ids = [image_ids[x] for x in sorted_inds]

        # go down detections and mark true positives and false positives
        nd = len(image_ids)
        tp = np.zeros(nd)
        fp = np.zeros(nd)
        for d in range(nd):
            r = class_anno[image_ids[d]]
            bb = bbox[d, :].astype(float)
            ovmax = -np.inf
            bbgt = r['bbox'].astype(float)

            if bbgt.size > 0:
                # compute overlaps
                # intersection
                ixmin = np.maximum(bbgt[:, 0], bb[0])
                iymin = np.maximum(bbgt[:, 1], bb[1])
                ixmax = np.minimum(bbgt[:, 2], bb[2])
                iymax = np.minimum(bbgt[:, 3], bb[3])
                iw = np.maximum(ixmax - ixmin + 1., 0.)
                ih = np.maximum(iymax - iymin + 1., 0.)
                inters = iw * ih

                # union
                uni = ((bb[2] - bb[0] + 1.) * (bb[3] - bb[1] + 1.) +
                       (bbgt[:, 2] - bbgt[:, 0] + 1.) *
                       (bbgt[:, 3] - bbgt[:, 1] + 1.) - inters)

                overlaps = inters / uni
                ovmax = np.max(overlaps)
                jmax = np.argmax(overlaps)

            if ovmax > ovthresh:
                if not r['difficult'][jmax]:
                    if not r['det'][jmax]:
                        tp[d] = 1.
                        r['det'][jmax] = 1
                    else:
                        fp[d] = 1.
            else:
                fp[d] = 1.

        # compute precision recall
        fp = np.cumsum(fp)
        tp = np.cumsum(tp)
        rec = tp / float(npos)
        # avoid division by zero in case first detection matches a difficult ground ruth
        prec = tp / np.maximum(tp + fp, np.finfo(np.float64).eps)
        ap = PascalVOC.voc_ap(rec, prec, use_07_metric)

        return rec, prec, ap 

def test_activation():
    shape=(9, 10)
    dtype_l = [np.float64, np.float32, np.float16]
    rtol_l = [1e-7, 1e-6, 1e-2]
    atol_l = [1e-7, 1e-6, 1e-2]
    rtol_fd = 1e-5
    atol_fd = 1e-6
    num_eps = 1e-6
    unary_ops = {
        'relu': [lambda x: mx.sym.Activation(x, act_type='relu'),
                 lambda x: np.maximum(x, 0.),
                 lambda x: 1. * (x > 0.),
                 -5.0, 5.0],
        'sigmoid': [lambda x: mx.sym.Activation(x, act_type='sigmoid'),
                    lambda x: 1. / (np.exp(-x) + 1.),
                    lambda x: 1. / (np.exp(-x) + 1.) / (np.exp(x) + 1.),
                    -3.0, 3.0],
        'tanh': [lambda x: mx.sym.Activation(x, act_type='tanh'),
                 lambda x: np.tanh(x),
                 lambda x: 1. - np.tanh(x) ** 2,
                 -4.0, 4.0],
        'softrelu': [lambda x: mx.sym.Activation(x, act_type='softrelu'),
                    lambda x: np.log(1. + np.exp(x)),
                    lambda x: 1. - 1 / (1 + np.exp(x)),
                    -3.0, 3.0],
        'softsign': [lambda x: mx.sym.Activation(x, act_type='softsign'),
                     lambda x: x / (1. + np.abs(x)),
                     lambda x: 1. / np.square(1. + np.abs(x)),
                     -3.0, 3.0],
    }
    # Loop over operators
    for name, op in unary_ops.items():
        # Loop over dtype's
        for ind in range(len(dtype_l)):
            dtype = dtype_l[ind]
            rtol = rtol_l[ind]
            atol = atol_l[ind]
            compare_forw_backw_unary_op(
                name, op[0], op[1], op[2], shape, op[3], op[4], rtol, atol,
                dtype)
        # Finite difference testing
        finite_diff_unary_op(
            name, op[0], shape, op[3], op[4], rtol_fd, atol_fd, num_eps) 

def test_box_iou_op():
    def numpy_box_iou(a, b, fmt='corner'):
        def area(left, top, right, bottom):
            return np.maximum(0, right - left) * np.maximum(0, bottom - top)

        assert a.shape[-1] == 4
        assert b.shape[-1] == 4
        oshape = a.shape[:-1] + b.shape[:-1]
        a = a.reshape((-1, 4))
        ashape = a.shape
        b = b.reshape((-1, 4))
        a = np.tile(a, reps=[1, b.shape[0]]).reshape((-1, 4))
        b = np.tile(b, reps=[ashape[0], 1]).reshape((-1, 4))
        if fmt == 'corner':
            al, at, ar, ab = np.split(a, 4, axis=-1)
            bl, bt, br, bb = np.split(b, 4, axis=-1)
        elif fmt == 'center':
            ax, ay, aw, ah = np.split(a, 4, axis=-1)
            bx, by, bw, bh = np.split(b, 4, axis=-1)
            al, at, ar, ab = ax - aw / 2, ay - ah / 2, ax + aw / 2, ay + ah / 2
            bl, bt, br, bb = bx - bw / 2, by - bh / 2, bx + bw / 2, by + bh / 2
        else:
            raise NotImplementedError("Fmt {} not supported".format(fmt))
        width = np.maximum(0, np.minimum(ar, br) - np.maximum(al, bl))
        height = np.maximum(0, np.minimum(ab, bb) - np.maximum(at, bt))
        intersect = width * height
        union = area(al, at, ar, ab) + area(bl, bt, br, bb) - intersect
        union[np.where(intersect <= 0)] = 1e-12
        iou = intersect / union
        return iou.reshape(oshape)

    def generate_boxes(dims):
        s1, off1, s2, off2 = np.random.rand(4) * 100
        xy = np.random.rand(*(dims + [2])) * s1 + off1
        wh = np.random.rand(*(dims + [2])) * s2 + off2
        xywh = np.concatenate([xy, wh], axis=-1)
        ltrb = np.concatenate([xy - wh / 2, xy + wh / 2], axis=-1)
        return xywh, ltrb


    for ndima in range(1, 6):
        for ndimb in range(1, 6):
            dims_a = np.random.randint(low=1, high=3, size=ndima).tolist()
            dims_b = np.random.randint(low=1, high=3, size=ndimb).tolist()
            # generate left, top, right, bottom
            xywh_a, ltrb_a = generate_boxes(dims_a)
            xywh_b, ltrb_b = generate_boxes(dims_b)

            iou_np = numpy_box_iou(ltrb_a, ltrb_b, fmt='corner')
            iou_np2 = numpy_box_iou(xywh_a, xywh_b, fmt='center')
            iou_mx = mx.nd.contrib.box_iou(mx.nd.array(ltrb_a), mx.nd.array(ltrb_b), format='corner')
            iou_mx2 = mx.nd.contrib.box_iou(mx.nd.array(xywh_a), mx.nd.array(xywh_b), format='center')
            assert_allclose(iou_np, iou_np2, rtol=1e-5, atol=1e-5)
            assert_allclose(iou_np, iou_mx.asnumpy(), rtol=1e-5, atol=1e-5)
            assert_allclose(iou_np, iou_mx2.asnumpy(), rtol=1e-5, atol=1e-5) 

def test_requantize_int32_to_int8():
    def quantized_int32_to_float(qdata, min_range, max_range):
        assert qdata.dtype == 'int32'
        quantized_range = np.iinfo('int32').max
        real_range = np.maximum(np.abs(min_range), np.abs(max_range))
        scale = float(real_range) / float(quantized_range)
        return qdata.astype('float32') * scale

    def float_to_quantized_int8(data, min_range, max_range):
        assert data.dtype == 'float32'
        real_range = np.maximum(np.abs(min_range), np.abs(max_range))
        quantized_range = np.iinfo('int8').max
        scale = float(quantized_range) / float(real_range)
        return (np.sign(data) * np.minimum(np.abs(data) * scale + 0.5, quantized_range)).astype('int8')

    def requantize(qdata, min_data, max_data, real_range):
        data = quantized_int32_to_float(qdata, min_data, max_data)
        output = float_to_quantized_int8(data, -real_range, real_range)
        return output, -real_range, real_range

    def requantize_baseline(qdata, min_data, max_data, min_calib_range=None, max_calib_range=None):
        if min_calib_range is not None and max_calib_range is not None:
            real_range = np.maximum(np.abs(min_calib_range), np.abs(max_calib_range))
            return requantize(qdata, min_data, max_data, real_range)
        else:
            min_range = quantized_int32_to_float(np.min(qdata), min_data, max_data)
            max_range = quantized_int32_to_float(np.max(qdata), min_data, max_data)
            return requantize(qdata, min_data, max_data, np.maximum(np.abs(min_range), np.abs(max_range)))

    def check_requantize(shape, min_calib_range=None, max_calib_range=None):
        qdata = mx.nd.random.uniform(low=-1000.0, high=1000.0, shape=shape).astype('int32')
        min_range = mx.nd.array([-1010.0])
        max_range = mx.nd.array([1020.0])
        if min_calib_range is None or max_calib_range is None:
            qdata_int8, min_output, max_output = mx.nd.contrib.requantize(qdata, min_range, max_range)
        else:
            qdata_int8, min_output, max_output = mx.nd.contrib.requantize(qdata, min_range, max_range,
                                                                          min_calib_range, max_calib_range)

        qdata_int8_np, min_output_np, max_output_np = requantize_baseline(qdata.asnumpy(), min_range.asscalar(),
                                                                          max_range.asscalar(),
                                                                          min_calib_range=min_calib_range,
                                                                          max_calib_range=max_calib_range)
        assert_almost_equal(qdata_int8.asnumpy(), qdata_int8_np, atol = 1)
        assert_almost_equal(min_output.asnumpy(), np.array([min_output_np]))
        assert_almost_equal(max_output.asnumpy(), np.array([max_output_np]))

    check_requantize((3, 4, 10, 10))
    check_requantize((32, 3, 23, 23))
    check_requantize((3, 4, 10, 10), min_calib_range=-1050.0, max_calib_range=1040.0)
    check_requantize((32, 3, 23, 23), min_calib_range=-134.349, max_calib_range=523.43) 

def crop(bbox, crop_box=None, allow_outside_center=True):
    """Crop bounding boxes according to slice area.

    This method is mainly used with image cropping to ensure bonding boxes fit
    within the cropped image.

    Parameters
    ----------
    bbox : numpy.ndarray
        Numpy.ndarray with shape (N, 4+) where N is the number of bounding boxes.
        The second axis represents attributes of the bounding box.
        Specifically, these are :math:`(x_{min}, y_{min}, x_{max}, y_{max})`,
        we allow additional attributes other than coordinates, which stay intact
        during bounding box transformations.
    crop_box : tuple
        Tuple of length 4. :math:`(x_{min}, y_{min}, width, height)`
    allow_outside_center : bool
        If `False`, remove bounding boxes which have centers outside cropping area.

    Returns
    -------
    numpy.ndarray
        Cropped bounding boxes with shape (M, 4+) where M <= N.
    """
    bbox = bbox.copy()
    if crop_box is None:
        return bbox
    if not len(crop_box) == 4:
        raise ValueError(
            "Invalid crop_box parameter, requires length 4, given {}".format(str(crop_box)))
    if sum([int(c is None) for c in crop_box]) == 4:
        return bbox

    l, t, w, h = crop_box

    left = l if l else 0
    top = t if t else 0
    right = left + (w if w else np.inf)
    bottom = top + (h if h else np.inf)
    crop_bbox = np.array((left, top, right, bottom))

    if allow_outside_center:
        mask = np.ones(bbox.shape[0], dtype=bool)
    else:
        centers = (bbox[:, :2] + bbox[:, 2:4]) / 2
        mask = np.logical_and(crop_bbox[:2] <= centers, centers < crop_bbox[2:]).all(axis=1)

    # transform borders
    bbox[:, :2] = np.maximum(bbox[:, :2], crop_bbox[:2])
    bbox[:, 2:4] = np.minimum(bbox[:, 2:4], crop_bbox[2:4])
    bbox[:, :2] -= crop_bbox[:2]
    bbox[:, 2:4] -= crop_bbox[:2]

    mask = np.logical_and(mask, (bbox[:, :2] < bbox[:, 2:4]).all(axis=1))
    bbox = bbox[mask]
    return bbox 

def _check_load_bbox(self, coco, entry):
        """Check and load ground-truth labels"""
        ann_ids = coco.getAnnIds(imgIds=entry['id'], iscrowd=None)
        objs = coco.loadAnns(ann_ids)
        # check valid bboxes
        valid_objs = []
        valid_segs = []
        width = entry['width']
        height = entry['height']
        for obj in objs:
            if obj.get('ignore', 0) == 1:
                continue
            # crowd objs cannot be used for segmentation
            if obj.get('iscrowd', 0) == 1:
                continue
            # need accurate floating point box representation
            x1, y1, w, h = obj['bbox']
            x2, y2 = x1 + np.maximum(0, w), y1 + np.maximum(0, h)
            # clip to image boundary
            x1 = np.minimum(width, np.maximum(0, x1))
            y1 = np.minimum(height, np.maximum(0, y1))
            x2 = np.minimum(width, np.maximum(0, x2))
            y2 = np.minimum(height, np.maximum(0, y2))
            # require non-zero seg area and more than 1x1 box size
            if obj['area'] > self._min_object_area and x2 > x1 and y2 > y1 \
                    and (x2 - x1) * (y2 - y1) >= 4:
                contiguous_cid = self.json_id_to_contiguous[obj['category_id']]
                valid_objs.append([x1, y1, x2, y2, contiguous_cid])

                segs = obj['segmentation']
                assert isinstance(segs, list), '{}'.format(obj.get('iscrowd', 0))
                valid_segs.append([np.asarray(p).reshape(-1, 2).astype('float32')
                                   for p in segs if len(p) >= 6])
        # there is no easy way to return a polygon placeholder: None is returned
        # in validation, None cannot be used for batchify -> drop label in transform
        # in training: empty images should be be skipped
        if not valid_objs:
            valid_objs = None
            valid_segs = None
        else:
            valid_objs = np.asarray(valid_objs).astype('float32')
        return valid_objs, valid_segs 

def get_spectrograms(fpath):
    '''Returns normalized log(melspectrogram) and log(magnitude) from `sound_file`.
    Args:
      sound_file: A string. The full path of a sound file.
    Returns:
      mel: A 2d array of shape (T, n_mels) <- Transposed
      mag: A 2d array of shape (T, 1+n_fft/2) <- Transposed
    '''

    # Loading sound file
    y, sr = librosa.load(fpath, sr=hp.sr)

    # Trimming
    y, _ = librosa.effects.trim(y)

    # Preemphasis
    y = np.append(y[0], y[1:] - hp.preemphasis * y[:-1])

    # stft
    linear = librosa.stft(y=y,
                          n_fft=hp.n_fft,
                          hop_length=hp.hop_length,
                          win_length=hp.win_length)

    # magnitude spectrogram
    mag = np.abs(linear)  # (1+n_fft//2, T)

    # mel spectrogram
    mel_basis = librosa.filters.mel(hp.sr, hp.n_fft, hp.n_mels)  # (n_mels, 1+n_fft//2)
    mel = np.dot(mel_basis, mag)  # (n_mels, t)

    # to decibel
    mel = 20 * np.log10(np.maximum(1e-5, mel))
    mag = 20 * np.log10(np.maximum(1e-5, mag))

    # normalize
    mel = np.clip((mel - hp.ref_db + hp.max_db) / hp.max_db, 1e-8, 1)
    mag = np.clip((mag - hp.ref_db + hp.max_db) / hp.max_db, 1e-8, 1)

    # Transpose
    mel = mel.T.astype(np.float32)  # (T, n_mels)
    mag = mag.T.astype(np.float32)  # (T, 1+n_fft//2)

    return mel, mag 

def plot_trajectories(outputs, global_step, output_dir, metric_summary, N):
  """Processes the collected outputs during validation to plot the trajectories
  in the top view.
  
  Args:
    outputs        : [locs, orig_maps, goal_loc].
    global_step    : global_step.
    output_dir     : where to store results.
    metric_summary : summary object to add summaries to.
    N              : number of outputs to process.
  """
  if N >= 0:
    outputs = outputs[:N]
  N = len(outputs)

  plt.set_cmap('gray')
  fig, axes = utils.subplot(plt, (N, outputs[0][1].shape[0]), (5,5))
  axes = axes.ravel()[::-1].tolist()
  for i in range(N):
    locs, orig_maps, goal_loc = outputs[i]
    is_semantic = np.isnan(goal_loc[0,0,1])
    for j in range(orig_maps.shape[0]):
      ax = axes.pop();
      ax.plot(locs[j,0,0], locs[j,0,1], 'ys')
      # Plot one by one, so that they come in different colors.
      for k in range(goal_loc.shape[1]):
        if not is_semantic:
          ax.plot(goal_loc[j,k,0], goal_loc[j,k,1], 's')
      if False:
        ax.plot(locs[j,:,0], locs[j,:,1], 'r.', ms=3)
        ax.imshow(orig_maps[j,0,:,:,0], origin='lower')
        ax.set_axis_off();
      else:
        ax.scatter(locs[j,:,0], locs[j,:,1], c=np.arange(locs.shape[1]),
                   cmap='jet', s=10, lw=0)
        ax.imshow(orig_maps[j,0,:,:,0], origin='lower', vmin=-1.0, vmax=2.0)
        if not is_semantic:
          xymin = np.minimum(np.min(goal_loc[j,:,:], axis=0), np.min(locs[j,:,:], axis=0))
          xymax = np.maximum(np.max(goal_loc[j,:,:], axis=0), np.max(locs[j,:,:], axis=0))
        else:
          xymin = np.min(locs[j,:,:], axis=0)
          xymax = np.max(locs[j,:,:], axis=0)
        xy1 = (xymax+xymin)/2. - np.maximum(np.max(xymax-xymin), 12)
        xy2 = (xymax+xymin)/2. + np.maximum(np.max(xymax-xymin), 12)
        ax.set_xlim([xy1[0], xy2[0]])
        ax.set_ylim([xy1[1], xy2[1]])
        ax.set_axis_off()
  file_name = os.path.join(output_dir, 'trajectory_{:d}.png'.format(global_step))
  with fu.fopen(file_name, 'w') as f:
    fig.savefig(f, bbox_inches='tight', transparent=True, pad_inches=0)
  plt.close(fig)
  return None 

def compute_average_precision(precision, recall):
  """Compute Average Precision according to the definition in VOCdevkit.

  Precision is modified to ensure that it does not decrease as recall
  decrease.

  Args:
    precision: A float [N, 1] numpy array of precisions
    recall: A float [N, 1] numpy array of recalls

  Raises:
    ValueError: if the input is not of the correct format

  Returns:
    average_precison: The area under the precision recall curve. NaN if
      precision and recall are None.

  """
  if precision is None:
    if recall is not None:
      raise ValueError("If precision is None, recall must also be None")
    return np.NAN

  if not isinstance(precision, np.ndarray) or not isinstance(recall,
                                                             np.ndarray):
    raise ValueError("precision and recall must be numpy array")
  if precision.dtype != np.float or recall.dtype != np.float:
    raise ValueError("input must be float numpy array.")
  if len(precision) != len(recall):
    raise ValueError("precision and recall must be of the same size.")
  if not precision.size:
    return 0.0
  if np.amin(precision) < 0 or np.amax(precision) > 1:
    raise ValueError("Precision must be in the range of [0, 1].")
  if np.amin(recall) < 0 or np.amax(recall) > 1:
    raise ValueError("recall must be in the range of [0, 1].")
  if not all(recall[i] <= recall[i + 1] for i in moves.range(len(recall) - 1)):
    raise ValueError("recall must be a non-decreasing array")

  recall = np.concatenate([[0], recall, [1]])
  precision = np.concatenate([[0], precision, [0]])

  # Preprocess precision to be a non-decreasing array
  for i in range(len(precision) - 2, -1, -1):
    precision[i] = np.maximum(precision[i], precision[i + 1])

  indices = np.where(recall[1:] != recall[:-1])[0] + 1
  average_precision = np.sum(
      (recall[indices] - recall[indices - 1]) * precision[indices])
  return average_precision