Python numpy.repeat() Examples

def offset_to_pts(self, center_list, pred_list):
        """Change from point offset to point coordinate."""
        pts_list = []
        for i_lvl in range(len(self.point_strides)):
            pts_lvl = []
            for i_img in range(len(center_list)):
                pts_center = center_list[i_img][i_lvl][:, :2].repeat(
                    1, self.num_points)
                pts_shift = pred_list[i_lvl][i_img]
                yx_pts_shift = pts_shift.permute(1, 2, 0).view(
                    -1, 2 * self.num_points)
                y_pts_shift = yx_pts_shift[..., 0::2]
                x_pts_shift = yx_pts_shift[..., 1::2]
                xy_pts_shift = torch.stack([x_pts_shift, y_pts_shift], -1)
                xy_pts_shift = xy_pts_shift.view(*yx_pts_shift.shape[:-1], -1)
                pts = xy_pts_shift * self.point_strides[i_lvl] + pts_center
                pts_lvl.append(pts)
            pts_lvl = torch.stack(pts_lvl, 0)
            pts_list.append(pts_lvl)
        return pts_list 

def evaluate_sliding_window(img_filename, crops):
    img = io.imread(img_filename).astype(np.float32)/255
    if img.ndim == 2: # Handle B/W images
        img = np.expand_dims(img, axis=-1)
        img = np.repeat(img, 3, 2)

    img_crops = np.zeros((batch_size, 227, 227, 3))
    for i in xrange(len(crops)):
        crop = crops[i]
        img_crop = transform.resize(img[crop[1]:crop[1]+crop[3],crop[0]:crop[0]+crop[2]], (227, 227))-0.5
        img_crop = np.expand_dims(img_crop, axis=0)
        img_crops[i,:,:,:] = img_crop

    # compute ranking scores
    scores = sess.run([score_func], feed_dict={image_placeholder: img_crops})

    # find the optimal crop
    idx = np.argmax(scores[:len(crops)])
    best_window = crops[idx]

    # return the best crop
    return (best_window[0], best_window[1], best_window[2], best_window[3]) 

def testMultilabelMatch3(self):
    predictions = np.random.randint(1, 5, size=(100, 1, 1, 1))
    targets = np.random.randint(1, 5, size=(100, 10, 1, 1))
    weights = np.random.randint(0, 2, size=(100, 1, 1, 1))
    targets *= weights

    predictions_repeat = np.repeat(predictions, 10, axis=1)
    expected = (predictions_repeat == targets).astype(float)
    expected = np.sum(expected, axis=(1, 2, 3))
    expected = np.minimum(expected / 3.0, 1.)
    expected = np.sum(expected * weights[:, 0, 0, 0]) / weights.shape[0]
    with self.test_session() as session:
      scores, weights_ = metrics.multilabel_accuracy_match3(
          tf.one_hot(predictions, depth=5, dtype=tf.float32),
          tf.constant(targets, dtype=tf.int32))
      a, a_op = tf.metrics.mean(scores, weights_)
      session.run(tf.local_variables_initializer())
      session.run(tf.global_variables_initializer())
      _ = session.run(a_op)
      actual = session.run(a)
    self.assertAlmostEqual(actual, expected, places=6) 

def zxy_grid(co_y, tymin, tymax, subs, c, t, c_peat, t_peat):
    # create linespace grid between bottom and top of tri z
    #subs = 7
    t_min = np.min(tymin)
    t_max = np.max(tymax)
    divs = np.linspace(t_min, t_max, num=subs, dtype=np.float32)            
    
    # figure out which triangles and which co are in each section
    co_bools = (co_y > divs[:-1][:, nax]) & (co_y < divs[1:][:, nax])
    tri_bools = (tymin < divs[1:][:, nax]) & (tymax > divs[:-1][:, nax])

    for i, j in zip(co_bools, tri_bools):
        if (np.sum(i) > 0) & (np.sum(j) > 0):
            c3 = c[i]
            t3 = t[j]
        
            c_peat.append(np.repeat(c3, t3.shape[0]))
            t_peat.append(np.tile(t3, c3.shape[0])) 

def repair_out_of_bounds_manually(X, xl, xu):

    only_1d = (X.ndim == 1)
    X = at_least_2d_array(X)

    if xl is not None:
        xl = np.repeat(xl[None, :], X.shape[0], axis=0)
        X[X < xl] = xl[X < xl]

    if xu is not None:
        xu = np.repeat(xu[None, :], X.shape[0], axis=0)
        X[X > xu] = xu[X > xu]

    if only_1d:
        return X[0, :]
    else:
        return X 

def bounds_back(problem, X):
    only_1d = (X.ndim == 1)
    X = at_least_2d_array(X)

    if problem.xl is not None and problem.xu is not None:
        xl = np.repeat(problem.xl[None, :], X.shape[0], axis=0)
        xu = np.repeat(problem.xu[None, :], X.shape[0], axis=0)

        # otherwise bounds back into the feasible space
        _range = xu - xl
        X[X < xl] = (xl + np.mod((xl - X), _range))[X < xl]
        X[X > xu] = (xu - np.mod((X - xu), _range))[X > xu]

    if only_1d:
        return X[0, :]
    else:
        return X 

def computeUVN_vec(n, in_, planeID):
    '''
    vectorization version of computeUVN
    @n         N x 3
    @in_      MN x 1
    @planeID   N
    '''
    n = n.copy()
    if (planeID == 2).sum():
        n[planeID == 2] = np.roll(n[planeID == 2], 2, axis=1)
    if (planeID == 3).sum():
        n[planeID == 3] = np.roll(n[planeID == 3], 1, axis=1)
    n = np.repeat(n, in_.shape[0] // n.shape[0], axis=0)
    assert n.shape[0] == in_.shape[0]
    bc = n[:, [0]] * np.sin(in_) + n[:, [1]] * np.cos(in_)
    bs = n[:, [2]]
    out = np.arctan(-bc / (bs + 1e-9))
    return out 

def np_sample(img, coords):
    # a numpy implementation of ImageSample layer
    coords = np.maximum(coords, 0)
    coords = np.minimum(coords, np.array([img.shape[0] - 1, img.shape[1] - 1]))

    lcoor = np.floor(coords).astype('int32')
    ucoor = lcoor + 1
    ucoor = np.minimum(ucoor, np.array([img.shape[0] - 1, img.shape[1] - 1]))
    diff = coords - lcoor
    neg_diff = 1.0 - diff

    lcoory, lcoorx = np.split(lcoor, 2, axis=2)
    ucoory, ucoorx = np.split(ucoor, 2, axis=2)
    diff = np.repeat(diff, 3, 2).reshape((diff.shape[0], diff.shape[1], 2, 3))
    neg_diff = np.repeat(neg_diff, 3, 2).reshape((diff.shape[0], diff.shape[1], 2, 3))
    diffy, diffx = np.split(diff, 2, axis=2)
    ndiffy, ndiffx = np.split(neg_diff, 2, axis=2)

    ret = img[lcoory, lcoorx, :] * ndiffx * ndiffy + \
        img[ucoory, ucoorx, :] * diffx * diffy + \
        img[lcoory, ucoorx, :] * ndiffy * diffx + \
        img[ucoory, lcoorx, :] * diffy * ndiffx
    return ret[:, :, 0, :] 

def setup_mnist(self, img_res):

        print ("Setting up MNIST...")

        if not os.path.exists('datasets/mnist_x.npy'):
            # Load the dataset
            (mnist_X, mnist_y), (_, _) = mnist.load_data()

            # Normalize and rescale images
            mnist_X = self.normalize(mnist_X)
            mnist_X = np.array([imresize(x, img_res) for x in mnist_X])
            mnist_X = np.expand_dims(mnist_X, axis=-1)
            mnist_X = np.repeat(mnist_X, 3, axis=-1)

            self.mnist_X, self.mnist_y = mnist_X, mnist_y

            # Save formatted images
            np.save('datasets/mnist_x.npy', self.mnist_X)
            np.save('datasets/mnist_y.npy', self.mnist_y)
        else:
            self.mnist_X = np.load('datasets/mnist_x.npy')
            self.mnist_y = np.load('datasets/mnist_y.npy')

        print ("+ Done.") 

def _project(self, matrix):
        if self.is_scalar():
            return 1
        else:
            # Note that we use Munkres algorithm, but expand columns from n to m
            # by replicating each column by group size.
            mm = np.repeat(matrix, self.col_sum, axis=1)
            indexes = lap.lapjv(np.asarray(-mm))
            result = np.zeros(self.shape)
            reduce = np.repeat(range(len(self.col_sum)), self.col_sum)
            for row, column in enumerate(indexes[1]):
                # map expanded column index to reduced group index.
                result[row, reduce[column]] = 1
            return result

    # Constrain all entries to be zero that correspond to
    # zeros in the matrix. 

def __init__(self, env, k, channel_order='hwc'):
        """Stack k last frames.

        Returns lazy array, which is much more memory efficient.

        See Also
        --------
        baselines.common.atari_wrappers.LazyFrames
        """
        gym.Wrapper.__init__(self, env)
        self.k = k
        self.frames = deque([], maxlen=k)
        self.stack_axis = {'hwc': 2, 'chw': 0}[channel_order]
        orig_obs_space = env.observation_space
        low = np.repeat(orig_obs_space.low, k, axis=self.stack_axis)
        high = np.repeat(orig_obs_space.high, k, axis=self.stack_axis)
        self.observation_space = spaces.Box(
            low=low, high=high, dtype=orig_obs_space.dtype) 

def render_cubemap(self, out_size, mode='blend'):
        # Create coordinate arrays.
        cvec = np.arange(out_size, dtype='float32') - out_size/2        # Coordinate range [-S/2, S/2)
        vec0 = np.ones(out_size*out_size, dtype='float32') * out_size/2 # Constant vector +S/2
        vec1 = np.repeat(cvec, out_size)                                # Increment every N steps
        vec2 = np.tile(cvec, out_size)                                  # Sweep N times
        # Create XYZ coordinate vectors and render each cubemap face.
        render = lambda(xyz): self._render(xyz, out_size, out_size, mode)
        xm = render(np.matrix([-vec0, vec1, vec2]))     # -X face
        xp = render(np.matrix([vec0, vec1, -vec2]))     # +X face
        ym = render(np.matrix([-vec1, -vec0, vec2]))    # -Y face
        yp = render(np.matrix([vec1, vec0, vec2]))      # +Y face
        zm = render(np.matrix([-vec2, vec1, -vec0]))    # -Z face
        zp = render(np.matrix([vec2, vec1, vec0]))      # +Z face
        # Concatenate the individual faces in canonical order:
        # https://en.wikipedia.org/wiki/Cube_mapping#Memory_Addressing
        img_mat = np.concatenate([zp, zm, ym, yp, xm, xp], axis=0)
        return Image.fromarray(img_mat)

    # Get XYZ vectors for an equirectangular render, in raster order.
    # (Each row left to right, with rows concatenates from top to bottom.) 

def _optimize_2D(nodes, triangles, stay=[]):
    ''' Optimize the locations of the points by moving them towards the center
    of their patch. This is done iterativally for all points for a number of
    iterations and using a .05 step length'''
    edges, tr_edges, adjacency_list = _edge_list(triangles)
    boundary = edges[adjacency_list[:, 1] == -1].reshape(-1)
    stay = np.union1d(boundary, stay)
    stay = stay.astype(int)
    n_iter = 5
    step_length = .05
    mean_bar = np.zeros_like(nodes)
    new_nodes = np.copy(nodes)
    k = np.bincount(triangles.reshape(-1), minlength=len(nodes))
    for n in range(n_iter):
        bar = np.mean(new_nodes[triangles], axis=1)
        for i in range(2):
            mean_bar[:, i] = np.bincount(triangles.reshape(-1),
                                         weights=np.repeat(bar[:, i], 3),
                                         minlength=len(nodes))
        mean_bar /= k[:, None]
        new_nodes += step_length * (mean_bar - new_nodes)
        new_nodes[stay] = nodes[stay]
    return new_nodes 

def nodes_areas(self):
        ''' Areas for all nodes in a surface

        Returns
        ---------
        nd: NodeData
            NodeData structure with normals for each node

        '''
        areas = self.elements_volumes_and_areas()[self.elm.triangles]
        triangle_nodes = self.elm[self.elm.triangles, :3] - 1
        nd = np.bincount(
            triangle_nodes.reshape(-1),
            np.repeat(areas/3., 3), self.nodes.nr
        )

        return NodeData(nd, 'areas') 

def combvec(arrays, out=None):

    arrays = [np.asarray(x) for x in arrays]
    dtype = arrays[0].dtype

    n = np.prod([x.size for x in arrays])
    if out is None:
        out = np.zeros([n, len(arrays)], dtype=dtype)

    m = n // arrays[0].size
    out[:,0] = np.repeat(arrays[0], m)
    if arrays[1:]:
        combvec(arrays[1:], out=out[0:m,1:])
        for j in range(1, arrays[0].size):
            out[j*m:(j+1)*m,1:] = out[0:m,1:]
    return out 

def gen_grid_from_reg(self, reg, previous_boxes):
        """Base on the previous bboxes and regression values, we compute the
        regressed bboxes and generate the grids on the bboxes.

        :param reg: the regression value to previous bboxes.
        :param previous_boxes: previous bboxes.
        :return: generate grids on the regressed bboxes.
        """
        b, _, h, w = reg.shape
        bxy = (previous_boxes[:, :2, ...] + previous_boxes[:, 2:, ...]) / 2.
        bwh = (previous_boxes[:, 2:, ...] -
               previous_boxes[:, :2, ...]).clamp(min=1e-6)
        grid_topleft = bxy + bwh * reg[:, :2, ...] - 0.5 * bwh * torch.exp(
            reg[:, 2:, ...])
        grid_wh = bwh * torch.exp(reg[:, 2:, ...])
        grid_left = grid_topleft[:, [0], ...]
        grid_top = grid_topleft[:, [1], ...]
        grid_width = grid_wh[:, [0], ...]
        grid_height = grid_wh[:, [1], ...]
        intervel = torch.linspace(0., 1., self.dcn_kernel).view(
            1, self.dcn_kernel, 1, 1).type_as(reg)
        grid_x = grid_left + grid_width * intervel
        grid_x = grid_x.unsqueeze(1).repeat(1, self.dcn_kernel, 1, 1, 1)
        grid_x = grid_x.view(b, -1, h, w)
        grid_y = grid_top + grid_height * intervel
        grid_y = grid_y.unsqueeze(2).repeat(1, 1, self.dcn_kernel, 1, 1)
        grid_y = grid_y.view(b, -1, h, w)
        grid_yx = torch.stack([grid_y, grid_x], dim=2)
        grid_yx = grid_yx.view(b, -1, h, w)
        regressed_bbox = torch.cat([
            grid_left, grid_top, grid_left + grid_width, grid_top + grid_height
        ], 1)
        return grid_yx, regressed_bbox 

def extract_vggish_features(paths, path2gt, model): 
    """Extracts VGGish features and their corresponding ground_truth and identifiers (the path).

       VGGish features are extracted from non-overlapping audio patches of 0.96 seconds, 
       where each audio patch covers 64 mel bands and 96 frames of 10 ms each.

       We repeat ground_truth and identifiers to fit the number of extracted VGGish features.
    """
    # 1) Extract log-mel spectrograms
    first_audio = True
    for p in paths:
        if first_audio:
            input_data = vggish_input.wavfile_to_examples(config['audio_folder'] + p)
            ground_truth = np.repeat(path2gt[p], input_data.shape[0], axis=0)
            identifiers = np.repeat(p, input_data.shape[0], axis=0)
            first_audio = False
        else:
            tmp_in = vggish_input.wavfile_to_examples(config['audio_folder'] + p)
            input_data = np.concatenate((input_data, tmp_in), axis=0)
            tmp_gt = np.repeat(path2gt[p], tmp_in.shape[0], axis=0)
            ground_truth = np.concatenate((ground_truth, tmp_gt), axis=0)
            tmp_id = np.repeat(p, tmp_in.shape[0], axis=0)
            identifiers = np.concatenate((identifiers, tmp_id), axis=0)

    # 2) Load Tensorflow model to extract VGGish features
    with tf.Graph().as_default(), tf.Session() as sess:
        vggish_slim.define_vggish_slim(training=False)
        vggish_slim.load_vggish_slim_checkpoint(sess, 'vggish_model.ckpt')
        features_tensor = sess.graph.get_tensor_by_name(vggish_params.INPUT_TENSOR_NAME)
        embedding_tensor = sess.graph.get_tensor_by_name(vggish_params.OUTPUT_TENSOR_NAME)
        extracted_feat = sess.run([embedding_tensor], feed_dict={features_tensor: input_data})
        feature = np.squeeze(np.asarray(extracted_feat))

    return [feature, ground_truth, identifiers] 

def extract_other_features(paths, path2gt, model_type):
    """Extracts MusiCNN or OpenL3 features and their corresponding ground_truth and identifiers (the path).

       OpenL3 features are extracted from non-overlapping audio patches of 1 second, 
       where each audio patch covers 128 mel bands.

       MusiCNN features are extracted from non-overlapping audio patches of 1 second, 
       where each audio patch covers 96 mel bands.

       We repeat ground_truth and identifiers to fit the number of extracted OpenL3 features.
    """

    if model_type == 'openl3':
        model = openl3.models.load_embedding_model(input_repr="mel128", content_type="music", embedding_size=512)

    first_audio = True
    for p in paths:
        if model_type == 'musicnn':
            taggram, tags, extracted_features = extractor(config['audio_folder'] + p, model='MSD_musicnn', extract_features=True, input_overlap=1)
            emb = extracted_features['max_pool'] # or choose any other layer, for example: emb = taggram
            # Documentation: https://github.com/jordipons/musicnn/blob/master/DOCUMENTATION.md
        elif model_type == 'openl3':
            wave, sr = wavefile_to_waveform(config['audio_folder'] + p, 'openl3')
            emb, _ = openl3.get_embedding(wave, sr, hop_size=1, model=model, verbose=False)

        if first_audio:
            features = emb
            ground_truth = np.repeat(path2gt[p], features.shape[0], axis=0)
            identifiers = np.repeat(p, features.shape[0], axis=0)
            first_audio = False
        else:
            features = np.concatenate((features, emb), axis=0)
            tmp_gt = np.repeat(path2gt[p], emb.shape[0], axis=0)
            ground_truth = np.concatenate((ground_truth, tmp_gt), axis=0)
            tmp_id = np.repeat(p, emb.shape[0], axis=0)
            identifiers = np.concatenate((identifiers, tmp_id), axis=0)

    return [features, ground_truth, identifiers] 

def to_stereo(waveform):
    """ Convert a waveform to stereo by duplicating if mono,
    or truncating if too many channels.

    :param waveform: a (N, d) numpy array.
    :returns: A stereo waveform as a (N, 1) numpy array.
    """
    if waveform.shape[1] == 1:
        return np.repeat(waveform, 2, axis=-1)
    if waveform.shape[1] > 2:
        return waveform[:, :2]
    return waveform 

def _get_mask(self, idx, in_data):
        """Returns the mask by which to multiply the parts of the embedding layer.
        In this version, we have no weights to apply.
        """
        mask = idx >= 0  # bool False for -1 values that should be removed. shape=(b,mnz)
        mask = np.expand_dims(mask,2) # shape = (b,mnz,1)
        mask = np.repeat(mask, self._proj_dim, axis=2) # shape = (b,mnz,d)
        return mask 

def _get_mask(self, idx, in_data):
        """Returns the mask by which to multiply the parts of the embedding layer.
        In this version, we apply the weights.
        """
        val = in_data[1].asnumpy()  # shape=(b,mnz)
        mask = idx >= 0  # bool False for -1 values that should be removed. shape=(b,mnz)
        mask = np.multiply(mask,val)  # All (b,mnz)
        mask = np.expand_dims(mask,2) # shape = (b,mnz,1)
        mask = np.repeat(mask, self._proj_dim, axis=2) # shape = (b,mnz,d)
        return mask 

def test_ndarray_elementwise():
    nrepeat = 10
    maxdim = 4
    all_type = [np.float32, np.float64, np.float16, np.uint8, np.int8, np.int32, np.int64]
    real_type = [np.float32, np.float64, np.float16]
    for repeat in range(nrepeat):
        for dim in range(1, maxdim):
            check_with_uniform(lambda x, y: x + y, 2, dim, type_list=all_type)
            check_with_uniform(lambda x, y: x - y, 2, dim, type_list=all_type)
            check_with_uniform(lambda x, y: x * y, 2, dim, type_list=all_type)
            check_with_uniform(lambda x, y: x / y, 2, dim, type_list=real_type)
            check_with_uniform(lambda x, y: x / y, 2, dim, rmin=1, type_list=all_type)
            check_with_uniform(mx.nd.sqrt, 1, dim, np.sqrt, rmin=0)
            check_with_uniform(mx.nd.square, 1, dim, np.square, rmin=0)
            check_with_uniform(lambda x: mx.nd.norm(x).asscalar(), 1, dim, np.linalg.norm) 

def test_ndarray_choose():
    shape = (100, 20)
    npy = np.arange(np.prod(shape)).reshape(shape)
    arr = mx.nd.array(npy)
    nrepeat = 3
    for repeat in range(nrepeat):
        indices = np.random.randint(shape[1], size=shape[0])
        assert same(npy[np.arange(shape[0]), indices],
                    mx.nd.choose_element_0index(arr, mx.nd.array(indices)).asnumpy()) 

def test_ndarray_fill():
    shape = (100, 20)
    npy = np.arange(np.prod(shape)).reshape(shape)
    arr = mx.nd.array(npy)
    new_npy = npy.copy()
    nrepeat = 3
    for repeat in range(nrepeat):
        indices = np.random.randint(shape[1], size=shape[0])
        val = np.random.randint(shape[1], size=shape[0])
        new_npy[:] = npy
        new_npy[np.arange(shape[0]), indices] = val
        assert same(new_npy,
                    mx.nd.fill_element_0index(arr, mx.nd.array(val), mx.nd.array(indices)).asnumpy()) 

def test_ndarray_saveload():
    nrepeat = 10
    fname = 'tmp_list.bin'
    for repeat in range(nrepeat):
        data = []
        # test save/load as list
        for i in range(10):
            data.append(random_ndarray(np.random.randint(1, 5)))
        mx.nd.save(fname, data)
        data2 = mx.nd.load(fname)
        assert len(data) == len(data2)
        for x, y in zip(data, data2):
            assert np.sum(x.asnumpy() != y.asnumpy()) == 0
        # test save/load as dict
        dmap = {'ndarray xx %s' % i : x for i, x in enumerate(data)}
        mx.nd.save(fname, dmap)
        dmap2 = mx.nd.load(fname)
        assert len(dmap2) == len(dmap)
        for k, x in dmap.items():
            y = dmap2[k]
            assert np.sum(x.asnumpy() != y.asnumpy()) == 0
        # test save/load as ndarray
        # we expect the single ndarray to be converted into a list containing the ndarray
        single_ndarray = data[0]
        mx.nd.save(fname, single_ndarray)
        single_ndarray_loaded = mx.nd.load(fname)
        assert len(single_ndarray_loaded) == 1
        single_ndarray_loaded = single_ndarray_loaded[0]
        assert np.sum(single_ndarray.asnumpy() != single_ndarray_loaded.asnumpy()) == 0
    os.remove(fname) 

def test_arange():
    for i in range(5):
        start = np.random.rand() * 10
        stop = start + np.random.rand() * 100
        step = np.random.rand() * 4
        repeat = int(np.random.rand() * 5) + 1
        gt = np.arange(start=start, stop=stop, step=step)
        gt = np.broadcast_to(gt.reshape((gt.shape[0], 1)), shape=(gt.shape[0], repeat)).ravel()
        pred = mx.nd.arange(start=start, stop=stop, step=step, repeat=repeat).asnumpy()
        assert_almost_equal(pred, gt)
    gt = np.arange(start=0, stop=10000**2, step=10001, dtype=np.int32)
    pred = mx.nd.arange(start=0, stop=10000**2, step=10001,
                        dtype="int32").asnumpy()
    assert_almost_equal(pred, gt) 

def test_ifft():
    nrepeat = 2
    maxdim = 10
    for repeat in range(nrepeat):
        for order in [2,4]:
            shape = tuple(np.random.randint(1, maxdim, size=order))
            check_ifft(shape) 

def check_consistency_NxM(sym_list, ctx_list):
    # e.g. if sym_list=[sym1, sym2] and ctx_list=[ctx1, ctx2, ctx3], then resulting lists are:
    # sym_list=[sym1, sym1, sym1, sym2, sym2, sym2] and ctx_list=[ctx1, ctx2, ctx3, ctx1, ctx2, ctx3]
    check_consistency(np.repeat(sym_list, len(ctx_list)), ctx_list * len(sym_list), scale=0.5) 

def sample_points_on_faces(vs, fs, rng, n_samples_per_face):
  idx = np.repeat(np.arange(fs.shape[0]), n_samples_per_face)
  
  r = rng.rand(idx.size, 2)
  r1 = r[:,:1]; r2 = r[:,1:]; sqrt_r1 = np.sqrt(r1);
  
  v1 = vs[fs[idx, 0], :]; v2 = vs[fs[idx, 1], :]; v3 = vs[fs[idx, 2], :];
  pts = (1-sqrt_r1)*v1 + sqrt_r1*(1-r2)*v2 + sqrt_r1*r2*v3
  
  v1 = vs[fs[:,0], :]; v2 = vs[fs[:, 1], :]; v3 = vs[fs[:, 2], :];
  ar = 0.5*np.sqrt(np.sum(np.cross(v1-v3, v2-v3)**2, 1))
  
  return pts, ar, idx 

def observation_space(self):
    low = self._env.observation_space.low
    high = self._env.observation_space.high
    low = np.repeat(low[None, ...], len(self._past_indices), 0)
    high = np.repeat(high[None, ...], len(self._past_indices), 0)
    if self._flatten:
      low = np.reshape(low, (-1,) + low.shape[2:])
      high = np.reshape(high, (-1,) + high.shape[2:])
    return gym.spaces.Box(low, high)