Python torch.arange() Examples

def plot_wh_methods():  # from utils.utils import *; plot_wh_methods()
    # Compares the two methods for width-height anchor multiplication
    # https://github.com/ultralytics/yolov3/issues/168
    x = np.arange(-4.0, 4.0, .1)
    ya = np.exp(x)
    yb = torch.sigmoid(torch.from_numpy(x)).numpy() * 2

    fig = plt.figure(figsize=(6, 3), dpi=150)
    plt.plot(x, ya, '.-', label='yolo method')
    plt.plot(x, yb ** 2, '.-', label='^2 power method')
    plt.plot(x, yb ** 2.5, '.-', label='^2.5 power method')
    plt.xlim(left=-4, right=4)
    plt.ylim(bottom=0, top=6)
    plt.xlabel('input')
    plt.ylabel('output')
    plt.legend()
    fig.tight_layout()
    fig.savefig('comparison.png', dpi=200) 

def _feature_window_function(window_type: str,
                             window_size: int,
                             blackman_coeff: float,
                             device: torch.device,
                             dtype: int,
                             ) -> Tensor:
    r"""Returns a window function with the given type and size
    """
    if window_type == HANNING:
        return torch.hann_window(window_size, periodic=False, device=device, dtype=dtype)
    elif window_type == HAMMING:
        return torch.hamming_window(window_size, periodic=False, alpha=0.54, beta=0.46, device=device, dtype=dtype)
    elif window_type == POVEY:
        # like hanning but goes to zero at edges
        return torch.hann_window(window_size, periodic=False, device=device, dtype=dtype).pow(0.85)
    elif window_type == RECTANGULAR:
        return torch.ones(window_size, device=device, dtype=dtype)
    elif window_type == BLACKMAN:
        a = 2 * math.pi / (window_size - 1)
        window_function = torch.arange(window_size, device=device, dtype=dtype)
        # can't use torch.blackman_window as they use different coefficients
        return (blackman_coeff - 0.5 * torch.cos(a * window_function) +
                (0.5 - blackman_coeff) * torch.cos(2 * a * window_function)).to(device=device, dtype=dtype)
    else:
        raise Exception('Invalid window type ' + window_type) 

def decimate(tensor, m):
    """
    Decimate a tensor by a factor 'm', i.e. downsample by keeping every 'm'th value.

    This is used when we convert FC layers to equivalent Convolutional layers, BUT of a smaller size.

    :param tensor: tensor to be decimated
    :param m: list of decimation factors for each dimension of the tensor; None if not to be decimated along a dimension
    :return: decimated tensor
    """
    assert tensor.dim() == len(m)
    for d in range(tensor.dim()):
        if m[d] is not None:
            tensor = tensor.index_select(dim=d,
                                         index=torch.arange(start=0, end=tensor.size(d), step=m[d]).long())

    return tensor 

def forward(self, x):
        # x is of shape: batchSize x dimInFeatures x numberNodesIn
        B = x.shape[0]
        F = x.shape[1]
        Nin = x.shape[2]
        # And now we add the zero padding
        if Nin < self.N:
            x = torch.cat((x,
                           torch.zeros(B, F, self.N-Nin)\
                                   .type(x.dtype).to(x.device)
                          ), dim = 2)
        # Compute the filter output
        u = jARMA(self.inverseWeight, self.directWeight, self.filterWeight,
                  self.S, x, b = self.bias, tMax = self.tMax)
        # So far, u is of shape batchSize x dimOutFeatures x numberNodes
        # And we want to return a tensor of shape
        # batchSize x dimOutFeatures x numberNodesIn
        # since the nodes between numberNodesIn and numberNodes are not required
        if Nin < self.N:
            u = torch.index_select(u, 2, torch.arange(Nin).to(u.device))
        return u 

def forward(self, x):
        # x is of shape: batchSize x dimInFeatures x numberNodesIn
        B = x.shape[0]
        F = x.shape[1]
        Nin = x.shape[2]
        # If we have less filter coefficients than the required ones, we need
        # to use the copying scheme
        if self.M == self.N:
            self.h = self.weight
        else:
            self.h = torch.index_select(self.weight, 4, self.copyNodes)
        # And now we add the zero padding
        if Nin < self.N:
            zeroPad = torch.zeros(B, F, self.N-Nin).type(x.dtype).to(x.device)
            x = torch.cat((x, zeroPad), dim = 2)
        # Compute the filter output
        u = NVGF(self.h, self.S, x, self.bias)
        # So far, u is of shape batchSize x dimOutFeatures x numberNodes
        # And we want to return a tensor of shape
        # batchSize x dimOutFeatures x numberNodesIn
        # since the nodes between numberNodesIn and numberNodes are not required
        if Nin < self.N:
            u = torch.index_select(u, 2, torch.arange(Nin).to(u.device))
        return u 

def forward(self, x):
        # x is of shape: batchSize x dimInFeatures x numberNodesIn
        B = x.shape[0]
        F = x.shape[1]
        Nin = x.shape[2]
        # And now we add the zero padding
        if Nin < self.N:
            x = torch.cat((x,
                           torch.zeros(B, F, self.N-Nin)\
                                   .type(x.dtype).to(x.device)
                          ), dim = 2)
        # Compute the filter output
        u = LSIGF(self.weight, self.S, x, self.bias)
        # So far, u is of shape batchSize x dimOutFeatures x numberNodes
        # And we want to return a tensor of shape
        # batchSize x dimOutFeatures x numberNodesIn
        # since the nodes between numberNodesIn and numberNodes are not required
        if Nin < self.N:
            u = torch.index_select(u, 2, torch.arange(Nin).to(u.device))
        return u 

def test(self, dataset):
        self.model.eval()
        with torch.no_grad():
            total_loss = 0.0
            predictions = torch.zeros(len(dataset), dtype=torch.float, device='cpu')
            indices = torch.arange(1, dataset.num_classes + 1, dtype=torch.float, device='cpu')
            for idx in tqdm(range(len(dataset)), desc='Testing epoch  ' + str(self.epoch) + ''):
                ltree, linput, rtree, rinput, label = dataset[idx]
                target = utils.map_label_to_target(label, dataset.num_classes)
                linput, rinput = linput.to(self.device), rinput.to(self.device)
                target = target.to(self.device)
                output = self.model(ltree, linput, rtree, rinput)
                loss = self.criterion(output, target)
                total_loss += loss.item()
                output = output.squeeze().to('cpu')
                predictions[idx] = torch.dot(indices, torch.exp(output))
        return total_loss / len(dataset), predictions 

def get_entropy(self, pred, label):
        n, c, h, w = pred.size()
        label = label.unsqueeze(3).long()
        pred = F.softmax(pred, 1).permute(0, 2, 3, 1)
        one_hot_label = ((torch.arange(c)).cuda() == label).float()

        if self.eps == 0:
            prior = 0
        else:
            if self.priorType == 'gaussian':
                tensor = (torch.arange(c).cuda() - label).float()
                prior = NormalDist(tensor, c / 10)
            elif self.priorType == 'uniform':
                prior = 1 / (c-1)

        smoothed_label = (1 - self.eps) * one_hot_label + self.eps * prior * (1 - one_hot_label)
        entropy = smoothed_label * safe_log(pred) + (1 - smoothed_label) * safe_log(1 - pred)
        return -entropy 

def get_ts_loss(self, temporal_scores, ts_labels,  answer_indices):
        """
        Args:
            temporal_scores: (N, 5, Li, 2)
            ts_labels: dict(st=(N, ), ed=(N, ))
            answer_indices: (N, )

        Returns:

        """
        bsz = len(answer_indices)
        # compute loss
        ca_temporal_scores_st_ed = \
            temporal_scores[torch.arange(bsz, dtype=torch.long), answer_indices]  # (N, Li, 2)
        loss_st = self.temporal_criterion(ca_temporal_scores_st_ed[:, :, 0], ts_labels["st"])
        loss_ed = self.temporal_criterion(ca_temporal_scores_st_ed[:, :, 1], ts_labels["ed"])
        return (loss_st + loss_ed) / 2. 

def forward(self, batch_token_ids):
        batch_size, sent_len = batch_token_ids.size()
        device = batch_token_ids.device

        batch_pos_ids = torch.arange(
            sent_len, dtype=torch.long, device=device, requires_grad=False
        )
        batch_pos_ids = batch_pos_ids.unsqueeze(0).expand_as(batch_token_ids)

        batch_token_emb = self.token_embedding(batch_token_ids)
        batch_pos_emb = self.pos_embedding(batch_pos_ids)

        batch_token_emb = batch_token_emb + batch_pos_emb

        batch_token_out = self.layer_norm(batch_token_emb)
        batch_token_out = self.dropout(batch_token_out)

        return batch_token_out 

def forward(self, batch_elem_emb, sent_pos_ids=None):
        if sent_pos_ids is None:
            num_elem = batch_elem_emb.size(-2)
            sent_pos_ids = torch.arange(
                num_elem, dtype=torch.long, device=batch_elem_emb.device, requires_grad=False
            )
        elif not isinstance(sent_pos_ids, torch.Tensor):
            sent_pos_ids = torch.tensor(
                sent_pos_ids, dtype=torch.long, device=batch_elem_emb.device, requires_grad=False
            )

        batch_pos_emb = self.embedding(sent_pos_ids)
        out = batch_elem_emb + batch_pos_emb
        out = self.dropout(self.layer_norm(out))

        return out 

def __iter__(self):
        # deterministically shuffle based on epoch
        if self.shuffle:
            g = torch.Generator()
            g.manual_seed(self.epoch)
            indices = torch.randperm(len(self.dataset), generator=g).tolist()
        else:
            indices = torch.arange(len(self.dataset)).tolist()

        # add extra samples to make it evenly divisible
        indices += indices[:(self.total_size - len(indices))]
        assert len(indices) == self.total_size

        # subsample
        indices = indices[self.rank:self.total_size:self.num_replicas]
        assert len(indices) == self.num_samples

        return iter(indices) 

def _create_data_set(self):
        # used to generate the dataset to test on. this is not used in testing (offline procedure)
        test_filepath = common_utils.get_asset_path('kaldi_file.wav')
        sr = 16000
        x = torch.arange(0, 20).float()
        # between [-6,6]
        y = torch.cos(2 * math.pi * x) + 3 * torch.sin(math.pi * x) + 2 * torch.cos(x)
        # between [-2^30, 2^30]
        y = (y / 6 * (1 << 30)).long()
        # clear the last 16 bits because they aren't used anyways
        y = ((y >> 16) << 16).float()
        torchaudio.save(test_filepath, y, sr)
        sound, sample_rate = torchaudio.load(test_filepath, normalization=False)
        print(y >> 16)
        self.assertTrue(sample_rate == sr)
        torch.testing.assert_allclose(y, sound) 

def _test_get_strided_helper(self, num_samples, window_size, window_shift, snip_edges):
        waveform = torch.arange(num_samples).float()
        output = kaldi._get_strided(waveform, window_size, window_shift, snip_edges)

        # from NumFrames in feature-window.cc
        n = window_size
        if snip_edges:
            m = 0 if num_samples < window_size else 1 + (num_samples - window_size) // window_shift
        else:
            m = (num_samples + (window_shift // 2)) // window_shift

        self.assertTrue(output.dim() == 2)
        self.assertTrue(output.shape[0] == m and output.shape[1] == n)

        window = torch.empty((m, window_size))

        for r in range(m):
            extract_window(window, waveform, r, window_size, window_shift, snip_edges)
        torch.testing.assert_allclose(window, output) 

def _get_uncertainty(self, mask_pred, labels):
        """Estimate uncertainty based on pred logits.

        We estimate uncertainty as L1 distance between 0.0 and the logits
        prediction in 'mask_pred' for the foreground class in `classes`.

        Args:
            mask_pred (Tensor): mask predication logits, shape (num_rois,
                num_classes, mask_height, mask_width).

            labels (list[Tensor]): Either predicted or ground truth label for
                each predicted mask, of length num_rois.

        Returns:
            scores (Tensor): Uncertainty scores with the most uncertain
                locations having the highest uncertainty score,
                shape (num_rois, 1, mask_height, mask_width)
        """
        if mask_pred.shape[1] == 1:
            gt_class_logits = mask_pred.clone()
        else:
            inds = torch.arange(mask_pred.shape[0], device=mask_pred.device)
            gt_class_logits = mask_pred[inds, labels].unsqueeze(1)
        return -torch.abs(gt_class_logits) 

def _test_istft_of_sine(self, amplitude, L, n):
        # stft of amplitude*sin(2*pi/L*n*x) with the hop length and window size equaling L
        x = torch.arange(2 * L + 1, dtype=torch.get_default_dtype())
        sound = amplitude * torch.sin(2 * math.pi / L * x * n)
        # stft = torch.stft(sound, L, hop_length=L, win_length=L,
        #                   window=torch.ones(L), center=False, normalized=False)
        stft = torch.zeros((L // 2 + 1, 2, 2))
        stft_largest_val = (amplitude * L) / 2.0
        if n < stft.size(0):
            stft[n, :, 1] = -stft_largest_val

        if 0 <= L - n < stft.size(0):
            # symmetric about L // 2
            stft[L - n, :, 1] = stft_largest_val

        estimate = torchaudio.functional.istft(stft, L, hop_length=L, win_length=L,
                                               window=torch.ones(L), center=False, normalized=False)
        # There is a larger error due to the scaling of amplitude
        _compare_estimate(sound, estimate, atol=1e-3) 

def __init__(self, thresh=1e-8, projDim=8192, input_dim=512):
         super(CBP, self).__init__()
         self.thresh = thresh
         self.projDim = projDim
         self.input_dim = input_dim
         self.output_dim = projDim
         torch.manual_seed(1)
         self.h_ = [
                 torch.randint(0, self.output_dim, (self.input_dim,),dtype=torch.long),
                 torch.randint(0, self.output_dim, (self.input_dim,),dtype=torch.long)
         ]
         self.weights_ = [
             (2 * torch.randint(0, 2, (self.input_dim,)) - 1).float(),
             (2 * torch.randint(0, 2, (self.input_dim,)) - 1).float()
         ]

         indices1 = torch.cat((torch.arange(input_dim, dtype=torch.long).reshape(1, -1),
                               self.h_[0].reshape(1, -1)), dim=0)
         indices2 = torch.cat((torch.arange(input_dim, dtype=torch.long).reshape(1, -1),
                               self.h_[1].reshape(1, -1)), dim=0)

         self.sparseM = [
             torch.sparse.FloatTensor(indices1, self.weights_[0], torch.Size([self.input_dim, self.output_dim])).to_dense(),
             torch.sparse.FloatTensor(indices2, self.weights_[1], torch.Size([self.input_dim, self.output_dim])).to_dense(),
         ] 

def _matrix(weights: torch.Tensor, u: torch.Tensor):
    """
    Performs systematic resampling of a 2D array of log weights along the second axis.
    independent of the others.
    :param weights: The weights to use for resampling
    :return: Resampled indices
    """
    n = weights.shape[1]
    index_range = torch.arange(n, dtype=u.dtype, device=weights.device).unsqueeze(0)

    probs = (index_range + u) / n
    cumsum = weights.cumsum(-1)

    cumsum[..., -1] = 1.

    return torch.searchsorted(cumsum, probs) 

def _get_crop_idx_mapping(side):
    """Helper method to get the order of crops.

    :param side: how many crops live on each side.

    Example. Say you have an image that gets devided into 16 crops, i.e., the image gets cut into 16 parts:

    [[ 0,  1,  2,  3],
     [ 4,  5,  6,  7],
     [ 8,  9, 10, 11],
     [12, 13, 14, 15]],

    However, due to our recursive cropping code, this results in crops that are ordered like this:

    index of crop:
     0  1  2  3  4 ...
    corresponds to part in image:
     0, 1, 4, 5, 2, 3, 6, 7, 8, 9, 12, 13, 10, 11, 14, 15

    This method returns the inverse, going from the index of the crop back to the index in the image.
    """
    a = torch.arange(side * side).reshape(1, 1, side, side)
    a = torch.cat((a, a, a), dim=1)
    # Create mapping
    #   Index of crop in original image -> index of crop in the order it was extracted,
    # E.g. 2 -> 4  means it's the 2nd crop, but in the image, it's at position 4 (see above).
    crops = {i: crop[0, 0, ...].flatten().item()
             for i, crop in enumerate(iter_crops(a, 1))}
    return crops 

def _vector(weights: torch.Tensor, u: torch.Tensor):
    """
    Performs systematic resampling of a 1D array log weights.
    :param weights: The weights to use for resampling
    :return: Resampled indices
    """
    n = weights.shape[0]
    probs = (torch.arange(n, dtype=u.dtype, device=weights.device) + u) / n

    cumsum = weights.cumsum(0)
    cumsum[..., -1] = 1.

    return torch.searchsorted(cumsum, probs) 

def sample(self, inds=None):
        inds = inds if inds is not None else torch.arange(self._means.shape[0], device=self._means.device)
        return _jitter(self._means[inds], self._bw_fac * self._cov.sqrt()) 

def __init__(self, d_model, max_len=512):
        super(PositionalEmbedding, self).__init__()

        # Compute the positional encodings once in log space.
        pe = torch.zeros(max_len, d_model).float()
        pe.require_grad = False

        position = torch.arange(0, max_len).float().unsqueeze(1)
        div_term = (torch.arange(0, d_model, 2).float() * -(math.log(10000.0) / d_model)).exp()

        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)

        pe = pe.unsqueeze(0)
        self.register_buffer('pe', pe) 

def flip(x, dim):
    dim = x.dim() + dim if dim < 0 else dim
    return x[tuple(slice(None, None) if i != dim
             else torch.arange(x.size(i)-1, -1, -1).long()
             for i in range(x.dim()))] 

def _renorm_cast_cdf_(cdf, precision):
    Lp = cdf.shape[-1]
    finals = 1  # NHW1
    # RENORMALIZATION_FACTOR in cuda
    f = torch.tensor(2, dtype=torch.float32, device=cdf.device).pow_(precision)
    cdf = cdf.mul((f - (Lp - 1)) / finals)  # TODO
    cdf = cdf.round()
    cdf = cdf.to(dtype=torch.int16, non_blocking=True)
    r = torch.arange(Lp, dtype=torch.int16, device=cdf.device)
    cdf.add_(r)
    return cdf 

def _get_lifter_coeffs(num_ceps: int, cepstral_lifter: float) -> Tensor:
    # returns size (num_ceps)
    # Compute liftering coefficients (scaling on cepstral coeffs)
    # coeffs are numbered slightly differently from HTK: the zeroth index is C0, which is not affected.
    i = torch.arange(num_ceps)
    return 1.0 + 0.5 * cepstral_lifter * torch.sin(math.pi * i / cepstral_lifter) 

def create_dct(
        n_mfcc: int,
        n_mels: int,
        norm: Optional[str]
) -> Tensor:
    r"""Create a DCT transformation matrix with shape (``n_mels``, ``n_mfcc``),
    normalized depending on norm.

    Args:
        n_mfcc (int): Number of mfc coefficients to retain
        n_mels (int): Number of mel filterbanks
        norm (str or None): Norm to use (either 'ortho' or None)

    Returns:
        Tensor: The transformation matrix, to be right-multiplied to
        row-wise data of size (``n_mels``, ``n_mfcc``).
    """
    # http://en.wikipedia.org/wiki/Discrete_cosine_transform#DCT-II
    n = torch.arange(float(n_mels))
    k = torch.arange(float(n_mfcc)).unsqueeze(1)
    dct = torch.cos(math.pi / float(n_mels) * (n + 0.5) * k)  # size (n_mfcc, n_mels)
    if norm is None:
        dct *= 2.0
    else:
        assert norm == "ortho"
        dct[0] *= 1.0 / math.sqrt(2.0)
        dct *= math.sqrt(2.0 / float(n_mels))
    return dct.t() 

def mask_along_axis_iid(
        specgrams: Tensor,
        mask_param: int,
        mask_value: float,
        axis: int
) -> Tensor:
    r"""
    Apply a mask along ``axis``. Mask will be applied from indices ``[v_0, v_0 + v)``, where
    ``v`` is sampled from ``uniform(0, mask_param)``, and ``v_0`` from ``uniform(0, max_v - v)``.

    Args:
        specgrams (Tensor): Real spectrograms (batch, channel, freq, time)
        mask_param (int): Number of columns to be masked will be uniformly sampled from [0, mask_param]
        mask_value (float): Value to assign to the masked columns
        axis (int): Axis to apply masking on (2 -> frequency, 3 -> time)

    Returns:
        Tensor: Masked spectrograms of dimensions (batch, channel, freq, time)
    """

    if axis != 2 and axis != 3:
        raise ValueError('Only Frequency and Time masking are supported')

    device = specgrams.device
    dtype = specgrams.dtype

    value = torch.rand(specgrams.shape[:2], device=device, dtype=dtype) * mask_param
    min_value = torch.rand(specgrams.shape[:2], device=device, dtype=dtype) * (specgrams.size(axis) - value)

    # Create broadcastable mask
    mask_start = min_value[..., None, None]
    mask_end = (min_value + value)[..., None, None]
    mask = torch.arange(0, specgrams.size(axis), device=device, dtype=dtype)

    # Per batch example masking
    specgrams = specgrams.transpose(axis, -1)
    specgrams.masked_fill_((mask >= mask_start) & (mask < mask_end), mask_value)
    specgrams = specgrams.transpose(axis, -1)

    return specgrams 

def _test_resample_waveform_accuracy(self, up_scale_factor=None, down_scale_factor=None,
                                         atol=1e-1, rtol=1e-4):
        # resample the signal and compare it to the ground truth
        n_to_trim = 20
        sample_rate = 1000
        new_sample_rate = sample_rate

        if up_scale_factor is not None:
            new_sample_rate *= up_scale_factor

        if down_scale_factor is not None:
            new_sample_rate //= down_scale_factor

        duration = 5  # seconds
        original_timestamps = torch.arange(0, duration, 1.0 / sample_rate)

        sound = 123 * torch.cos(2 * math.pi * 3 * original_timestamps).unsqueeze(0)
        estimate = kaldi.resample_waveform(sound, sample_rate, new_sample_rate).squeeze()

        new_timestamps = torch.arange(0, duration, 1.0 / new_sample_rate)[:estimate.size(0)]
        ground_truth = 123 * torch.cos(2 * math.pi * 3 * new_timestamps)

        # trim the first/last n samples as these points have boundary effects
        ground_truth = ground_truth[..., n_to_trim:-n_to_trim]
        estimate = estimate[..., n_to_trim:-n_to_trim]

        torch.testing.assert_allclose(estimate, ground_truth, atol=atol, rtol=rtol) 

def _test_1_save_sine(self):

        # save created file
        sinewave_filepath = os.path.join(self.test_dirpath, "assets",
                                         "sinewave.wav")
        sr = 16000
        freq = 440
        volume = 0.3

        y = (torch.cos(
            2 * math.pi * torch.arange(0, 4 * sr).float() * freq / sr))
        y.unsqueeze_(0)
        # y is between -1 and 1, so must scale
        y = (y * volume * (2**31)).long()
        torchaudio.save(sinewave_filepath, y, sr)
        self.assertTrue(os.path.isfile(sinewave_filepath))

        # test precision
        new_precision = 32
        new_filepath = os.path.join(self.test_dirpath, "test.wav")
        si, ei = torchaudio.info(sinewave_filepath)
        torchaudio.save(new_filepath, y, sr, new_precision)
        si32, ei32 = torchaudio.info(new_filepath)
        self.assertEqual(si.precision, 16)
        self.assertEqual(si32.precision, new_precision)
        os.unlink(new_filepath) 

def __init__(self, n_filters=128, max_len=500):
        """
        :param n_filters: same with input hidden size
        :param max_len: maximum sequence length
        """
        super(PositionEncoding, self).__init__()
        # Compute the positional encodings once in log space.
        pe = torch.zeros(max_len, n_filters)  # (L, D)
        position = torch.arange(0, max_len).float().unsqueeze(1)
        div_term = torch.exp(torch.arange(0, n_filters, 2).float() * - (math.log(10000.0) / n_filters))
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        self.register_buffer('pe', pe)  # buffer is a tensor, not a variable, (L, D)