Python torch.gather方法代码示例

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def _get_body(self, x, target):
        cos_t = torch.gather(x, 1, target.unsqueeze(1))  # cos(theta_yi)
        if self.easy_margin:
            cond = torch.relu(cos_t)
        else:
            cond_v = cos_t - self.threshold
            cond = torch.relu(cond_v)
        cond = cond.bool()
        # Apex would convert FP16 to FP32 here
        # cos(theta_yi + m)
        new_zy = torch.cos(torch.acos(cos_t) + self.m).type(cos_t.dtype)
        if self.easy_margin:
            zy_keep = cos_t
        else:
            zy_keep = cos_t - self.mm  # (cos(theta_yi) - sin(pi - m)*m)
        new_zy = torch.where(cond, new_zy, zy_keep)
        diff = new_zy - cos_t  # cos(theta_yi + m) - cos(theta_yi)
        gt_one_hot = F.one_hot(target, num_classes=self.classes)
        body = gt_one_hot * diff
        return body 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def masked_cross_entropy_(logits, target, length, take_log=False):
    if USE_CUDA:
        length = Variable(torch.LongTensor(length)).cuda()
    else:
        length = Variable(torch.LongTensor(length))

    # logits_flat: (batch * max_len, num_classes)
    # -1 means infered from other dimentions
    logits_flat = logits.view(-1, logits.size(-1))
    if take_log:
        logits_flat = torch.log(logits_flat)
    # target_flat: (batch * max_len, 1)
    target_flat = target.view(-1, 1)
    # losses_flat: (batch * max_len, 1)
    losses_flat = -torch.gather(logits_flat, dim=1, index=target_flat)
    # losses: (batch, max_len)
    losses = losses_flat.view(*target.size())
    # mask: (batch, max_len)
    mask = sequence_mask(sequence_length=length, max_len=target.size(1))
    losses = losses * mask.float()
    loss = losses.sum() / length.float().sum()
    return loss 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def masked_cross_entropy_for_slot(logits, target, mask, use_softmax=True):
    # print("logits", logits)
    # print("target", target)
    # -1 means infered from other dimentions
    logits_flat = logits.view(-1, logits.size(-1))
    # print(logits_flat.size())
    if use_softmax:
        log_probs_flat = functional.log_softmax(logits_flat, dim=1)
    else:
        log_probs_flat = logits_flat  # torch.log(logits_flat)
    # print("log_probs_flat", log_probs_flat)
    target_flat = target.view(-1, 1)
    # print("target_flat", target_flat)
    losses_flat = -torch.gather(log_probs_flat, dim=1, index=target_flat)
    losses = losses_flat.view(*target.size())  # b * |s|
    losses = losses * mask.float()
    loss = losses.sum() / (losses.size(0)*losses.size(1))
    # print("loss inside", loss)
    return loss 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def get_ranking_loss(n):
    def ranking_loss(scores, targets):
        """ Slack-rescaled max margin loss
        """
        costs = targets[1]
        true_ants = targets[2]
        weights = targets[4] if len(targets) == 5 else None
        true_ant_score = torch.gather(scores, 1, true_ants)
        top_true, _ = true_ant_score.max(dim=1)
        tmp_loss = scores.add(1).add(
            top_true.unsqueeze(1).neg()
        )  # 1 + scores - top_true
        if weights is not None:
            tmp_loss = tmp_loss.mul(weights)
        tmp_loss = tmp_loss.mul(costs)
        loss, _ = tmp_loss.max(dim=1)
        out_score = torch.sum(loss)
        return out_score / n

    return ranking_loss 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def _compute_loss(self, batch, output, target):
        scores = self.generator(self._bottle(output))

        gtruth = target.view(-1)
        if self.confidence < 1:
            tdata = gtruth.data
            mask = torch.nonzero(tdata.eq(self.padding_idx)).squeeze()
            log_likelihood = torch.gather(scores.data, 1, tdata.unsqueeze(1))
            tmp_ = self.one_hot.repeat(gtruth.size(0), 1)
            tmp_.scatter_(1, tdata.unsqueeze(1), self.confidence)
            if mask.dim() > 0:
                log_likelihood.index_fill_(0, mask, 0)
                tmp_.index_fill_(0, mask, 0)
            gtruth = Variable(tmp_, requires_grad=False)
        loss = self.criterion(scores, gtruth)
        if self.confidence < 1:
            # Default: report smoothed ppl.
            # loss_data = -log_likelihood.sum(0)
            loss_data = loss.data.clone()
        else:
            loss_data = loss.data.clone()

        stats = self._stats(loss_data, scores.data, target.view(-1).data)

        return loss, stats 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def get_ranking(predictions, labels, num_guesses=5):
    """
    Given a matrix of predictions and labels for the correct ones, get the number of guesses
    required to get the prediction right per example.
    :param predictions: [batch_size, range_size] predictions
    :param labels: [batch_size] array of labels
    :param num_guesses: Number of guesses to return
    :return: 
    """
    assert labels.size(0) == predictions.size(0)
    assert labels.dim() == 1
    assert predictions.dim() == 2

    values, full_guesses = predictions.topk(predictions.size(1), dim=1)
    _, ranking = full_guesses.topk(full_guesses.size(1), dim=1, largest=False)
    gt_ranks = torch.gather(ranking.data, 1, labels[:, None]).squeeze()

    guesses = full_guesses[:, :num_guesses]
    return gt_ranks, guesses 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def forward(self, x, y=None, batch=256):
        if y is not None:
            x, y = self.cr(x, y)
        else:
            x, y = self.cr(x)
        x = torch.nn.utils.rnn.pad_packed_sequence(x)
        r = torch.transpose(x[0], 0, 1)
        y = y.view(batch, self.hidden_size)
        ind = x[1].view(batch, 1, 1)
        ind = ind - 1
        ind = ind.expand(-1, -1, self.hidden_size)
        t = torch.gather(r, 1, ind)
        t = t.view(batch, self.hidden_size)
        t = self.fn(t)
        y = self.fn2(y)
        t = t + y
        t = self.sig(t)
        return t 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def calculate_loss(
            self, states, actions, old_policy, old_values, returns, advantages):
        """
        All parameters ought to be tensors on the appropriate compute device.
        """
        values, policy = self.model(states)
        a_policy = torch.gather(policy, -1, actions[..., np.newaxis])[..., 0]

        prob_diff = advantages.sign() * (1 - a_policy / old_policy)
        policy_loss = advantages.abs() * torch.clamp(prob_diff, min=-self.eps_policy)
        policy_loss = policy_loss.mean()

        v_clip = old_values + torch.clamp(
            values - old_values, min=-self.eps_value, max=+self.eps_value)
        value_loss = torch.max((v_clip - returns)**2, (values - returns)**2)
        value_loss = value_loss.mean()

        entropy = torch.sum(-policy * torch.log(policy + 1e-12), dim=-1)
        entropy_loss = torch.clamp(entropy.mean(), max=self.entropy_clip)
        entropy_loss *= -self.entropy_reg

        return entropy, policy_loss + value_loss * self.vf_coef + entropy_loss 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def eval(b, x, *k_b):
        original_shape = x.shape
        x = x.reshape(-1)
        n_values = x.shape[0]
        x = bit_decomposition(x)
        s, t = Array(n + 1, λ, n_values), Array(n + 1, 1, n_values)
        s[0] = k_b[0]
        # here k[1:] is (CW, CW_n)
        CW = k_b[1].unbind() + (k_b[2],)
        t[0] = b
        for i in range(0, n):
            τ = G(s[i]) ^ (t[i] * CW[i])
            τ = τ.reshape(2, λ + 1, n_values)
            x_i = x[i].unsqueeze(0).expand(λ + 1, n_values).unsqueeze(0).long()
            filtered_τ = th.gather(τ, 0, x_i).squeeze(0)
            s[i + 1], t[i + 1] = split(filtered_τ, [λ, 1])
        flat_result = (-1) ** b * (Convert(s[n]) + t[n].squeeze() * CW[n])
        return flat_result.reshape(original_shape) 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def masked_cross_entropy(logits, target, length, use_cuda=True):
    if use_cuda:
        length = Variable(torch.LongTensor(length)).cuda()
        target_flat = target.view(-1, 1).cuda()
    else:
        length = Variable(torch.LongTensor(length))
        target_flat = target.view(-1, 1)
    # logits_flat: (batch * max_len, num_classes)
    logits_flat = logits.view(-1, logits.size(-1))
    # log_probs_flat: (batch * max_len, num_classes)
    log_probs_flat = functional.log_softmax(logits_flat)
    # target_flat: (batch * max_len, 1)

    losses_flat = -torch.gather(log_probs_flat, dim=1, index=target_flat)

    # losses: (batch, max_len)
    losses = losses_flat.view(*target.size())
    # mask: (batch, max_len)
    mask = sequence_mask(sequence_length=length, max_len=target.size(1))
    losses = losses * mask.float()
    loss = losses.sum() / length.float().sum()
    return loss 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def _expand_visual(self, visual: utils.TensorOrSequence, cur_beam_size: int, selected_beam: torch.Tensor):
        if isinstance(visual, torch.Tensor):
            visual_shape = visual.shape
            visual_exp_shape = (self.b_s, cur_beam_size) + visual_shape[1:]
            visual_red_shape = (self.b_s * self.beam_size,) + visual_shape[1:]
            selected_beam_red_size = (self.b_s, self.beam_size) + tuple(1 for _ in range(len(visual_exp_shape) - 2))
            selected_beam_exp_size = (self.b_s, self.beam_size) + visual_exp_shape[2:]
            visual_exp = visual.view(visual_exp_shape)
            selected_beam_exp = selected_beam.view(selected_beam_red_size).expand(selected_beam_exp_size)
            visual = torch.gather(visual_exp, 1, selected_beam_exp).view(visual_red_shape)
        else:
            new_visual = []
            for im in visual:
                visual_shape = im.shape
                visual_exp_shape = (self.b_s, cur_beam_size) + visual_shape[1:]
                visual_red_shape = (self.b_s * self.beam_size,) + visual_shape[1:]
                selected_beam_red_size = (self.b_s, self.beam_size) + tuple(1 for _ in range(len(visual_exp_shape) - 2))
                selected_beam_exp_size = (self.b_s, self.beam_size) + visual_exp_shape[2:]
                visual_exp = im.view(visual_exp_shape)
                selected_beam_exp = selected_beam.view(selected_beam_red_size).expand(selected_beam_exp_size)
                new_im = torch.gather(visual_exp, 1, selected_beam_exp).view(visual_red_shape)
                new_visual.append(new_im)
            visual = tuple(new_visual)
        return visual 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def forward(self, input, target, th=1.0):
        log_p = F.log_softmax(input, dim=1)
        if th < 1: # This is done while using Hardmining. Not used for our model training
            mask = F.softmax(input, dim=1) > th
            mask = mask.data
            new_target = target.data.clone()
            new_target[new_target == self.ignore] = 0
            indx = torch.gather(mask, 1, new_target.unsqueeze(1))
            indx = indx.squeeze(1)
            mod_target = target.clone()
            mod_target[indx] = self.ignore
            target = mod_target

        loss = self.nll_loss(log_p, target)
        total_valid_pixel = torch.sum(target.data != self.ignore)

        return loss, Variable(torch.FloatTensor([total_valid_pixel]).cuda()) 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def update(self, model_output, target, **kwargs):
        """
        args:
            model_output: tensor of shape (B, C) where each value is
                          either logit or class probability.
            target:       tensor of shape (B, C), which is one-hot /
                          multi-label encoded, or tensor of shape (B) /
                          (B, 1), integer encoded
        """
        # Convert target to 0/1 encoding if isn't
        target = maybe_convert_to_one_hot(target, model_output)

        _, pred = model_output.topk(max(self._topk), dim=1, largest=True, sorted=True)
        for i, k in enumerate(self._topk):
            self._curr_correct_predictions_k[i] += (
                torch.gather(target, dim=1, index=pred[:, :k])
                .max(dim=1)
                .values.sum()
                .item()
            )
        self._curr_sample_count += model_output.shape[0] 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def select_next_words(
        self, word_scores, bsz, beam_size, possible_translation_tokens
    ):
        cand_scores, cand_indices = torch.topk(word_scores.view(bsz, -1), k=beam_size)
        possible_tokens_size = self.vocab_size
        if possible_translation_tokens is not None:
            possible_tokens_size = possible_translation_tokens.size(0)
        cand_beams = torch.div(cand_indices, possible_tokens_size)
        cand_indices.fmod_(possible_tokens_size)
        # Handle vocab reduction
        if possible_translation_tokens is not None:
            possible_translation_tokens = possible_translation_tokens.view(
                1, possible_tokens_size
            ).expand(cand_indices.size(0), possible_tokens_size)
            cand_indices = torch.gather(
                possible_translation_tokens, dim=1, index=cand_indices, out=cand_indices
            )
        return cand_scores, cand_indices, cand_beams 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import gather [as 别名]
def get_kd_loss(self, sample, student_lprobs, lprobs):
        """
        The second return argument is used for unit testing.

        Args:
            * sample: batched sample that has teacher score keys (top_k_scores and
             top_k_indices)
            * student_lprobs: tensor of student log probabilities
            * lprobs: flat version of student_lprobs
        """
        top_k_teacher_probs_normalized = sample["top_k_scores"]
        indices = sample["top_k_indices"]

        assert indices.shape[0:1] == student_lprobs.shape[0:1]

        kd_loss = -torch.sum(
            torch.gather(student_lprobs, 2, indices)
            * top_k_teacher_probs_normalized.float()
        )
        return kd_loss