Python torch.max() Examples

def plot_one_box(x, img, color=None, label=None, line_thickness=None):
    # Plots one bounding box on image img
    tl = line_thickness or round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 1  # line thickness
    color = color or [random.randint(0, 255) for _ in range(3)]
    c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3]))
    cv2.rectangle(img, c1, c2, color, thickness=tl)
    if label:
        tf = max(tl - 1, 1)  # font thickness
        t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0]
        c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3
        cv2.rectangle(img, c1, c2, color, -1)  # filled
        cv2.putText(img, label, (c1[0], c1[1] - 2), 0, tl / 3, [225, 255, 255], thickness=tf, lineType=cv2.LINE_AA) 

def plot_evolution_results(hyp):  # from utils.utils import *; plot_evolution_results(hyp)
    # Plot hyperparameter evolution results in evolve.txt
    x = np.loadtxt('evolve.txt', ndmin=2)
    f = fitness(x)
    weights = (f - f.min()) ** 2  # for weighted results
    fig = plt.figure(figsize=(12, 10))
    matplotlib.rc('font', **{'size': 8})
    for i, (k, v) in enumerate(hyp.items()):
        y = x[:, i + 5]
        # mu = (y * weights).sum() / weights.sum()  # best weighted result
        mu = y[f.argmax()]  # best single result
        plt.subplot(4, 5, i + 1)
        plt.plot(mu, f.max(), 'o', markersize=10)
        plt.plot(y, f, '.')
        plt.title('%s = %.3g' % (k, mu), fontdict={'size': 9})  # limit to 40 characters
        print('%15s: %.3g' % (k, mu))
    fig.tight_layout()
    plt.savefig('evolve.png', dpi=200) 

def iou(source: Tensor, other: Tensor) -> Tensor:
        source, other = source.unsqueeze(dim=-2).repeat(1, 1, other.shape[-2], 1), \
                        other.unsqueeze(dim=-3).repeat(1, source.shape[-2], 1, 1)

        source_area = (source[..., 2] - source[..., 0]) * (source[..., 3] - source[..., 1])
        other_area = (other[..., 2] - other[..., 0]) * (other[..., 3] - other[..., 1])

        intersection_left = torch.max(source[..., 0], other[..., 0])
        intersection_top = torch.max(source[..., 1], other[..., 1])
        intersection_right = torch.min(source[..., 2], other[..., 2])
        intersection_bottom = torch.min(source[..., 3], other[..., 3])
        intersection_width = torch.clamp(intersection_right - intersection_left, min=0)
        intersection_height = torch.clamp(intersection_bottom - intersection_top, min=0)
        intersection_area = intersection_width * intersection_height

        return intersection_area / (source_area + other_area - intersection_area) 

def get(self):
        """Get top cost recorded.

        You should call step before get at least once.

        Returns:
            A float number of cost.

        """
        assert len(self.coll) != 0, 'Please call step before get'
        if self.type == 'mean':
            ret = np.mean(self.coll)
        elif self.type == 'max':
            ret = np.max(self.coll)
        else:
            ret = np.min(self.coll)
        return ret 

def greedy_decode(self, latent, max_len, start_id):
        '''
        latent: (batch_size, max_src_seq, d_model)
        src_mask: (batch_size, 1, max_src_len)
        '''
        batch_size = latent.size(0)
        ys = get_cuda(torch.ones(batch_size, 1).fill_(start_id).long())  # (batch_size, 1)
        for i in range(max_len - 1):
            # input("==========")
            # print("="*10, i)
            # print("ys", ys.size())  # (batch_size, i)
            # print("tgt_mask", subsequent_mask(ys.size(1)).size())  # (1, i, i)
            out = self.decode(latent.unsqueeze(1), to_var(ys), to_var(subsequent_mask(ys.size(1)).long()))
            prob = self.generator(out[:, -1])
            # print("prob", prob.size())  # (batch_size, vocab_size)
            _, next_word = torch.max(prob, dim=1)
            # print("next_word", next_word.size())  # (batch_size)

            # print("next_word.unsqueeze(1)", next_word.unsqueeze(1).size())

            ys = torch.cat([ys, next_word.unsqueeze(1)], dim=1)
            # print("ys", ys.size())
        return ys[:, 1:] 

def shem(roi_probs_neg, negative_count, ohem_poolsize):
    """
    stochastic hard example mining: from a list of indices (referring to non-matched predictions),
    determine a pool of highest scoring (worst false positives) of size negative_count*ohem_poolsize.
    Then, sample n (= negative_count) predictions of this pool as negative examples for loss.
    :param roi_probs_neg: tensor of shape (n_predictions, n_classes).
    :param negative_count: int.
    :param ohem_poolsize: int.
    :return: (negative_count).  indices refer to the positions in roi_probs_neg. If pool smaller than expected due to
    limited negative proposals availabel, this function will return sampled indices of number < negative_count without
    throwing an error.
    """
    # sort according to higehst foreground score.
    probs, order = roi_probs_neg[:, 1:].max(1)[0].sort(descending=True)
    select = torch.tensor((ohem_poolsize * int(negative_count), order.size()[0])).min().int()
    pool_indices = order[:select]
    rand_idx = torch.randperm(pool_indices.size()[0])
    return pool_indices[rand_idx[:negative_count].cuda()] 

def greedy_decode(self, latent, max_len, start_id):
        '''
        latent: (batch_size, max_src_seq, d_model)
        src_mask: (batch_size, 1, max_src_len)
        '''
        batch_size = latent.size(0)
        ys = get_cuda(torch.ones(batch_size, 1).fill_(start_id).long())  # (batch_size, 1)
        for i in range(max_len - 1):
            # input("==========")
            # print("="*10, i)
            # print("ys", ys.size())  # (batch_size, i)
            # print("tgt_mask", subsequent_mask(ys.size(1)).size())  # (1, i, i)
            out = self.decode(latent.unsqueeze(1), to_var(ys), to_var(subsequent_mask(ys.size(1)).long()))
            prob = self.generator(out[:, -1])
            # print("prob", prob.size())  # (batch_size, vocab_size)
            _, next_word = torch.max(prob, dim=1)
            # print("next_word", next_word.size())  # (batch_size)

            # print("next_word.unsqueeze(1)", next_word.unsqueeze(1).size())

            ys = torch.cat([ys, next_word.unsqueeze(1)], dim=1)
            # print("ys", ys.size())
        return ys[:, 1:] 

def greedy_decode(self, latent, max_len, start_id):
        '''
        latent: (batch_size, max_src_seq, d_model)
        src_mask: (batch_size, 1, max_src_len)
        '''
        batch_size = latent.size(0)
        ys = get_cuda(torch.ones(batch_size, 1).fill_(start_id).long())  # (batch_size, 1)
        for i in range(max_len - 1):
            # input("==========")
            # print("="*10, i)
            # print("ys", ys.size())  # (batch_size, i)
            # print("tgt_mask", subsequent_mask(ys.size(1)).size())  # (1, i, i)
            out = self.decode(latent.unsqueeze(1), to_var(ys), to_var(subsequent_mask(ys.size(1)).long()))
            prob = self.generator(out[:, -1])
            # print("prob", prob.size())  # (batch_size, vocab_size)
            _, next_word = torch.max(prob, dim=1)
            # print("next_word", next_word.size())  # (batch_size)

            # print("next_word.unsqueeze(1)", next_word.unsqueeze(1).size())

            ys = torch.cat([ys, next_word.unsqueeze(1)], dim=1)
            # print("ys", ys.size())
        return ys[:, 1:] 

def _get_waveform_and_window_properties(waveform: Tensor,
                                        channel: int,
                                        sample_frequency: float,
                                        frame_shift: float,
                                        frame_length: float,
                                        round_to_power_of_two: bool,
                                        preemphasis_coefficient: float) -> Tuple[Tensor, int, int, int]:
    r"""Gets the waveform and window properties
    """
    channel = max(channel, 0)
    assert channel < waveform.size(0), ('Invalid channel %d for size %d' % (channel, waveform.size(0)))
    waveform = waveform[channel, :]  # size (n)
    window_shift = int(sample_frequency * frame_shift * MILLISECONDS_TO_SECONDS)
    window_size = int(sample_frequency * frame_length * MILLISECONDS_TO_SECONDS)
    padded_window_size = _next_power_of_2(window_size) if round_to_power_of_two else window_size

    assert 2 <= window_size <= len(waveform), ('choose a window size %d that is [2, %d]' % (window_size, len(waveform)))
    assert 0 < window_shift, '`window_shift` must be greater than 0'
    assert padded_window_size % 2 == 0, 'the padded `window_size` must be divisible by two.' \
                                        ' use `round_to_power_of_two` or change `frame_length`'
    assert 0. <= preemphasis_coefficient <= 1.0, '`preemphasis_coefficient` must be between [0,1]'
    assert sample_frequency > 0, '`sample_frequency` must be greater than zero'
    return waveform, window_shift, window_size, padded_window_size 

def do_eval(model, test_loader, cuda):
    model.is_training = False
    predicted = []
    true_label = []
    for X, y in test_loader:
        X = Variable(X)
        if cuda:
            X = X.cuda()
        output = model(X)
        output = output.squeeze(0)
        _, output = torch.max(output, 1)
        if cuda:
            output = output.cpu()
        predicted.extend(output.data.numpy().tolist())
        y = y.squeeze(0)
        true_label.extend(y.numpy().tolist()) 
    print("Acc: %.3f" % accuracy(predicted, true_label))
    return predicted 

def forward(self,x):
        input_size = x.size()[2]
        self.interp1 = nn.UpsamplingBilinear2d(size = (  int(input_size*0.75)+1,  int(input_size*0.75)+1  ))
        self.interp2 = nn.UpsamplingBilinear2d(size = (  int(input_size*0.5)+1,   int(input_size*0.5)+1   ))
        self.interp3 = nn.UpsamplingBilinear2d(size = (  outS(input_size),   outS(input_size)   ))

        out = []
        x2 = self.interp1(x)        
        x3 = self.interp2(x)

        out.append(self.Scale(x))                   #1.0x
        out.append(self.interp3(self.Scale(x2)))    #0.75x
        out.append(self.interp3(self.Scale(x3)))    #0.5x
        #out.append(self.Scale(x3))  # for 0.5x scale
        
        x2Out_interp = out[1]
        x3Out_interp = out[2]
        temp1 = torch.max(out[0],x2Out_interp)
        out.append(torch.max(temp1,x3Out_interp))
        return out 

def evo_norm(x, prefix, running_var, v, weight, bias,
             training, momentum, eps=0.1, groups=32):
    if prefix == 'b0':
        if training:
            var = torch.var(x, dim=(0, 2, 3), keepdim=True)
            running_var.mul_(momentum)
            running_var.add_((1 - momentum) * var)
        else:
            var = running_var
        if v is not None:
            den = torch.max((var + eps).sqrt(), v * x + instance_std(x, eps))
            x = x / den * weight + bias
        else:
            x = x * weight + bias
    else:
        if v is not None:
            x = x * torch.sigmoid(v * x) / group_std(x,
                                                     groups, eps) * weight + bias
        else:
            x = x * weight + bias

    return x 

def test(model, target_test_loader):
    model.eval()
    test_loss = utils.AverageMeter()
    correct = 0
    criterion = torch.nn.CrossEntropyLoss()
    len_target_dataset = len(target_test_loader.dataset)
    with torch.no_grad():
        for data, target in target_test_loader:
            data, target = data.to(DEVICE), target.to(DEVICE)
            s_output = model.predict(data)
            loss = criterion(s_output, target)
            test_loss.update(loss.item())
            pred = torch.max(s_output, 1)[1]
            correct += torch.sum(pred == target)

    print('{} --> {}: max correct: {}, accuracy{: .2f}%\n'.format(
        source_name, target_name, correct, 100. * correct / len_target_dataset)) 

def extract_feature(model, dataloader, save_path, load_from_disk=True, model_path=''):
    if load_from_disk:
        model = models.Network(base_net=args.model_name,
                               n_class=args.num_class)
        model.load_state_dict(torch.load(model_path))
        model = model.to(DEVICE)
    model.eval()
    correct = 0
    fea_all = torch.zeros(1,1+model.base_network.output_num()).to(DEVICE)
    with torch.no_grad():
        for inputs, labels in dataloader:
            inputs, labels = inputs.to(DEVICE), labels.to(DEVICE)
            feas = model.get_features(inputs)
            labels = labels.view(labels.size(0), 1).float()
            x = torch.cat((feas, labels), dim=1)
            fea_all = torch.cat((fea_all, x), dim=0)
            outputs = model(inputs)
            preds = torch.max(outputs, 1)[1]
            correct += torch.sum(preds == labels.data.long())
        test_acc = correct.double() / len(dataloader.dataset)
    fea_numpy = fea_all.cpu().numpy()
    np.savetxt(save_path, fea_numpy[1:], fmt='%.6f', delimiter=',')
    print('Test acc: %f' % test_acc)

# You may want to classify with 1nn after getting features 

def record_output(self, output, output_indices, target, prev_absolutes,
                      next_absolutes, batch_size=1):
        assert output.dim() == 4
        assert target.dim() == 3

        _, predictions = output.max(3)

        # Compute per class accuracy for unbalanced data.
        sequence_length = output.size(1)
        num_label = output.size(2)
        num_class = output.size(3)
        correct_alljoint = (target == predictions).float().sum(2)
        sum_of_corrects = correct_alljoint.sum(1)
        max_value = num_label * sequence_length
        count_correct = (sum_of_corrects == max_value).float().mean()
        correct_per_seq = ((correct_alljoint == num_label - 1).sum(1).float() /
                           sequence_length).mean()
        self.meter.update(
            torch.Tensor([count_correct * 100, correct_per_seq * 100]),
            batch_size) 

def crop_images_random(path='../images/', scale=0.50):  # from utils.utils import *; crop_images_random()
    # crops images into random squares up to scale fraction
    # WARNING: overwrites images!
    for file in tqdm(sorted(glob.glob('%s/*.*' % path))):
        img = cv2.imread(file)  # BGR
        if img is not None:
            h, w = img.shape[:2]

            # create random mask
            a = 30  # minimum size (pixels)
            mask_h = random.randint(a, int(max(a, h * scale)))  # mask height
            mask_w = mask_h  # mask width

            # box
            xmin = max(0, random.randint(0, w) - mask_w // 2)
            ymin = max(0, random.randint(0, h) - mask_h // 2)
            xmax = min(w, xmin + mask_w)
            ymax = min(h, ymin + mask_h)

            # apply random color mask
            cv2.imwrite(file, img[ymin:ymax, xmin:xmax]) 

def bbox_overlaps_3D(boxes1, boxes2):
    """Computes IoU overlaps between two sets of boxes.
    boxes1, boxes2: [N, (y1, x1, y2, x2, z1, z2)].
    """
    # 1. Tile boxes2 and repeate boxes1. This allows us to compare
    # every boxes1 against every boxes2 without loops.
    # TF doesn't have an equivalent to np.repeate() so simulate it
    # using tf.tile() and tf.reshape.
    boxes1_repeat = boxes2.size()[0]
    boxes2_repeat = boxes1.size()[0]
    boxes1 = boxes1.repeat(1,boxes1_repeat).view(-1,6)
    boxes2 = boxes2.repeat(boxes2_repeat,1)

    # 2. Compute intersections
    b1_y1, b1_x1, b1_y2, b1_x2, b1_z1, b1_z2 = boxes1.chunk(6, dim=1)
    b2_y1, b2_x1, b2_y2, b2_x2, b2_z1, b2_z2 = boxes2.chunk(6, dim=1)
    y1 = torch.max(b1_y1, b2_y1)[:, 0]
    x1 = torch.max(b1_x1, b2_x1)[:, 0]
    y2 = torch.min(b1_y2, b2_y2)[:, 0]
    x2 = torch.min(b1_x2, b2_x2)[:, 0]
    z1 = torch.max(b1_z1, b2_z1)[:, 0]
    z2 = torch.min(b1_z2, b2_z2)[:, 0]
    zeros = Variable(torch.zeros(y1.size()[0]), requires_grad=False)
    if y1.is_cuda:
        zeros = zeros.cuda()
    intersection = torch.max(x2 - x1, zeros) * torch.max(y2 - y1, zeros) * torch.max(z2 - z1, zeros)

    # 3. Compute unions
    b1_volume = (b1_y2 - b1_y1) * (b1_x2 - b1_x1)  * (b1_z2 - b1_z1)
    b2_volume = (b2_y2 - b2_y1) * (b2_x2 - b2_x1)  * (b2_z2 - b2_z1)
    union = b1_volume[:,0] + b2_volume[:,0] - intersection

    # 4. Compute IoU and reshape to [boxes1, boxes2]
    iou = intersection / union
    overlaps = iou.view(boxes2_repeat, boxes1_repeat)
    return overlaps 

def __init__(self, name=None, record_type='mean'):
        super(NumericalCost, self).__init__(name)
        self.coll = []
        self.type = record_type
        assert record_type in ('mean', 'max', 'min') 

def compute_rpn_class_loss(rpn_match, rpn_class_logits, shem_poolsize):
    """
    :param rpn_match: (n_anchors). [-1, 0, 1] for negative, neutral, and positive matched anchors.
    :param rpn_class_logits: (n_anchors, 2). logits from RPN classifier.
    :param shem_poolsize: int. factor of top-k candidates to draw from per negative sample
    (stochastic-hard-example-mining).
    :return: loss: torch tensor
    :return: np_neg_ix: 1D array containing indices of the neg_roi_logits, which have been sampled for training.
    """

    # filter out neutral anchors.
    pos_indices = torch.nonzero(rpn_match == 1)
    neg_indices = torch.nonzero(rpn_match == -1)

    # loss for positive samples
    if 0 not in pos_indices.size():
        pos_indices = pos_indices.squeeze(1)
        roi_logits_pos = rpn_class_logits[pos_indices]
        pos_loss = F.cross_entropy(roi_logits_pos, torch.LongTensor([1] * pos_indices.shape[0]).cuda())
    else:
        pos_loss = torch.FloatTensor([0]).cuda()

    # loss for negative samples: draw hard negative examples (SHEM)
    # that match the number of positive samples, but at least 1.
    if 0 not in neg_indices.size():
        neg_indices = neg_indices.squeeze(1)
        roi_logits_neg = rpn_class_logits[neg_indices]
        negative_count = np.max((1, pos_indices.cpu().data.numpy().size))
        roi_probs_neg = F.softmax(roi_logits_neg, dim=1)
        neg_ix = mutils.shem(roi_probs_neg, negative_count, shem_poolsize)
        neg_loss = F.cross_entropy(roi_logits_neg[neg_ix], torch.LongTensor([0] * neg_ix.shape[0]).cuda())
        np_neg_ix = neg_ix.cpu().data.numpy()
    else:
        neg_loss = torch.FloatTensor([0]).cuda()
        np_neg_ix = np.array([]).astype('int32')

    loss = (pos_loss + neg_loss) / 2
    return loss, np_neg_ix 

def bbox_iou(box1, box2, x1y1x2y2=True, GIoU=False):
    # Returns the IoU of box1 to box2. box1 is 4, box2 is nx4
    box2 = box2.t()

    # Get the coordinates of bounding boxes
    if x1y1x2y2:
        # x1, y1, x2, y2 = box1
        b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3]
        b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3]
    else:
        # x, y, w, h = box1
        b1_x1, b1_x2 = box1[0] - box1[2] / 2, box1[0] + box1[2] / 2
        b1_y1, b1_y2 = box1[1] - box1[3] / 2, box1[1] + box1[3] / 2
        b2_x1, b2_x2 = box2[0] - box2[2] / 2, box2[0] + box2[2] / 2
        b2_y1, b2_y2 = box2[1] - box2[3] / 2, box2[1] + box2[3] / 2

    # Intersection area
    inter_area = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \
                 (torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0)

    # Union Area
    union_area = ((b1_x2 - b1_x1) * (b1_y2 - b1_y1) + 1e-16) + \
                 (b2_x2 - b2_x1) * (b2_y2 - b2_y1) - inter_area

    iou = inter_area / union_area  # iou
    if GIoU:  # Generalized IoU https://arxiv.org/pdf/1902.09630.pdf
        c_x1, c_x2 = torch.min(b1_x1, b2_x1), torch.max(b1_x2, b2_x2)
        c_y1, c_y2 = torch.min(b1_y1, b2_y1), torch.max(b1_y2, b2_y2)
        c_area = (c_x2 - c_x1) * (c_y2 - c_y1) + 1e-16  # convex area
        return iou - (c_area - union_area) / c_area  # GIoU

    return iou 

def bbox_overlaps_2D(boxes1, boxes2):
    """Computes IoU overlaps between two sets of boxes.
    boxes1, boxes2: [N, (y1, x1, y2, x2)].
    """
    # 1. Tile boxes2 and repeate boxes1. This allows us to compare
    # every boxes1 against every boxes2 without loops.
    # TF doesn't have an equivalent to np.repeate() so simulate it
    # using tf.tile() and tf.reshape.
    boxes1_repeat = boxes2.size()[0]
    boxes2_repeat = boxes1.size()[0]
    boxes1 = boxes1.repeat(1,boxes1_repeat).view(-1,4)
    boxes2 = boxes2.repeat(boxes2_repeat,1)

    # 2. Compute intersections
    b1_y1, b1_x1, b1_y2, b1_x2 = boxes1.chunk(4, dim=1)
    b2_y1, b2_x1, b2_y2, b2_x2 = boxes2.chunk(4, dim=1)
    y1 = torch.max(b1_y1, b2_y1)[:, 0]
    x1 = torch.max(b1_x1, b2_x1)[:, 0]
    y2 = torch.min(b1_y2, b2_y2)[:, 0]
    x2 = torch.min(b1_x2, b2_x2)[:, 0]
    zeros = Variable(torch.zeros(y1.size()[0]), requires_grad=False)
    if y1.is_cuda:
        zeros = zeros.cuda()
    intersection = torch.max(x2 - x1, zeros) * torch.max(y2 - y1, zeros)

    # 3. Compute unions
    b1_area = (b1_y2 - b1_y1) * (b1_x2 - b1_x1)
    b2_area = (b2_y2 - b2_y1) * (b2_x2 - b2_x1)
    union = b1_area[:,0] + b2_area[:,0] - intersection

    # 4. Compute IoU and reshape to [boxes1, boxes2]
    iou = intersection / union
    overlaps = iou.view(boxes2_repeat, boxes1_repeat)
    return overlaps 

def clip(bboxes: Tensor, left: float, top: float, right: float, bottom: float) -> Tensor:
        bboxes[..., [0, 2]] = bboxes[..., [0, 2]].clamp(min=left, max=right)
        bboxes[..., [1, 3]] = bboxes[..., [1, 3]].clamp(min=top, max=bottom)
        return bboxes 

def record_output(self, output, output_indices, target, prev_absolute_imu,
                      next_absolute_imu, batch_size=1):
        prediction = torch.max(output, 1)[1]
        success = torch.sum((prediction == target).float()) / len(prediction)
        self.meter.update(success * 100, batch_size) 

def record_output(self, output, output_indices, target, prev_absolute_imu,
                      next_absolute_imu, batch_size=1):

        if self.args.no_angle_metric:
            return []

        prev_absolute_imu = prev_absolute_imu[:, output_indices]
        next_absolute_imu = next_absolute_imu[:, output_indices]

        assert output.dim() == 4
        if self.regression:
            # Output is the vectors itself.
            relative_imus = output
        else:
            output = torch.max(output, -1)[1]  #get labels from vectors
            relative_imus = self.get_diff_from_initial(output)
        if self.absolute_regress:
            resulting_imus = output
        else:
            resulting_imus = self.inverse_subtract(prev_absolute_imu,
                                                   relative_imus)

        angle_diff = self.get_angle_diff(next_absolute_imu, resulting_imus)
        whole_diff = self.get_angle_diff(next_absolute_imu[:, 0:1, :, :],
                                         next_absolute_imu[:, -1:, :, :])

        self.stats.append([whole_diff, angle_diff.mean()])
        self.outputs.append(output)
        self.targets.append(target)

        degree = angle_diff * 180 / math.pi
        self.meter.update(degree, batch_size)

        return resulting_imus 

def record_output(self, output, output_indices, target, prev_absolutes,
                      next_absolutes, batch_size=1):
        assert output.dim() == 4
        assert target.dim() == 3

        _, predictions = output.max(3)

        # Compute per class accuracy for unbalanced data.
        sequence_length = output.size(1)
        num_label = output.size(2)
        num_class = output.size(3)

        if self.confusion is None:
            # Confusion matrix is 4D because it's defined per label and sequence
            # element.
            self.confusion = torch.zeros(sequence_length, num_label, num_class,
                                         num_class)
        # Compute per class accuracy in this batch and update the confusion
        # matrix.
        per_class_acc = []
        for seq_id in range(sequence_length):
            for imu_id in range(num_label):
                imu_target = target[:, seq_id, imu_id].contiguous()
                imu_preds = predictions[:, seq_id, imu_id].contiguous()
                for label, pred in zip(imu_target.view(-1), imu_preds.view(-1)):
                    self.confusion[seq_id, imu_id, label, pred] += 1.0
                for class_id in range(num_class):
                    # Look at targets where label is class_id, and see what
                    # percentage of predictions are class_id.
                    preds_for_class = imu_preds[imu_target == class_id]
                    if len(preds_for_class) > 0:
                        per_class_acc.append(
                            (preds_for_class == class_id).float().mean())
        per_class_acc = sum(per_class_acc) / len(per_class_acc)
        accuracy = (predictions == target).float().mean()
        self.meter.update(
            torch.Tensor([100 * accuracy, 100 * per_class_acc]), batch_size) 

def _print_diagnostic(self, output, expect_output):
        # given an output and expected output, it will print the absolute/relative errors (max and mean squared)
        abs_error = output - expect_output
        abs_mse = abs_error.pow(2).sum() / output.numel()
        abs_max_error = torch.max(abs_error.abs())

        relative_error = abs_error / expect_output
        relative_mse = relative_error.pow(2).sum() / output.numel()
        relative_max_error = torch.max(relative_error.abs())

        print('abs_mse:', abs_mse.item(), 'abs_max_error:', abs_max_error.item())
        print('relative_mse:', relative_mse.item(), 'relative_max_error:', relative_max_error.item()) 

def decode_greedy(self, logits, seq_lens):
        decoded = []
        tlogits = logits.transpose(0, 1)
        _, tokens = torch.max(tlogits, 2)
        for i in range(tlogits.size(0)):
            output_str = self.convert_to_string(tokens[i], seq_lens[i])
            decoded.append(output_str)
        return decoded 

def _get_log_energy(strided_input: Tensor,
                    epsilon: Tensor,
                    energy_floor: float) -> Tensor:
    r"""Returns the log energy of size (m) for a strided_input (m,*)
    """
    device, dtype = strided_input.device, strided_input.dtype
    log_energy = torch.max(strided_input.pow(2).sum(1), epsilon).log()  # size (m)
    if energy_floor == 0.0:
        return log_energy
    return torch.max(
        log_energy, torch.tensor(math.log(energy_floor), device=device, dtype=dtype)) 

def amplitude_to_DB(
        x: Tensor,
        multiplier: float,
        amin: float,
        db_multiplier: float,
        top_db: Optional[float] = None
) -> Tensor:
    r"""Turn a tensor from the power/amplitude scale to the decibel scale.

    This output depends on the maximum value in the input tensor, and so
    may return different values for an audio clip split into snippets vs. a
    a full clip.

    Args:
        x (Tensor): Input tensor before being converted to decibel scale
        multiplier (float): Use 10. for power and 20. for amplitude
        amin (float): Number to clamp ``x``
        db_multiplier (float): Log10(max(reference value and amin))
        top_db (float or None, optional): Minimum negative cut-off in decibels. A reasonable number
            is 80. (Default: ``None``)

    Returns:
        Tensor: Output tensor in decibel scale
    """
    x_db = multiplier * torch.log10(torch.clamp(x, min=amin))
    x_db -= multiplier * db_multiplier

    if top_db is not None:
        x_db = x_db.clamp(min=x_db.max().item() - top_db)

    return x_db 

def forward(self, features, rois):
        batch_size, num_channels, data_height, data_width = features.size()
        num_rois = rois.size()[0]
        outputs = Variable(torch.zeros(num_rois, num_channels, self.pooled_height, self.pooled_width)).cuda()

        for roi_ind, roi in enumerate(rois):
            batch_ind = int(roi[0].data[0])
            roi_start_w, roi_start_h, roi_end_w, roi_end_h = np.round(
                roi[1:].data.cpu().numpy() * self.spatial_scale).astype(int)
            roi_width = max(roi_end_w - roi_start_w + 1, 1)
            roi_height = max(roi_end_h - roi_start_h + 1, 1)
            bin_size_w = float(roi_width) / float(self.pooled_width)
            bin_size_h = float(roi_height) / float(self.pooled_height)

            for ph in range(self.pooled_height):
                hstart = int(np.floor(ph * bin_size_h))
                hend = int(np.ceil((ph + 1) * bin_size_h))
                hstart = min(data_height, max(0, hstart + roi_start_h))
                hend = min(data_height, max(0, hend + roi_start_h))
                for pw in range(self.pooled_width):
                    wstart = int(np.floor(pw * bin_size_w))
                    wend = int(np.ceil((pw + 1) * bin_size_w))
                    wstart = min(data_width, max(0, wstart + roi_start_w))
                    wend = min(data_width, max(0, wend + roi_start_w))

                    is_empty = (hend <= hstart) or(wend <= wstart)
                    if is_empty:
                        outputs[roi_ind, :, ph, pw] = 0
                    else:
                        data = features[batch_ind]
                        outputs[roi_ind, :, ph, pw] = torch.max(
                            torch.max(data[:, hstart:hend, wstart:wend], 1)[0], 2)[0].view(-1)

        return outputs