Python mxnet.ndarray.max() Examples

def pad_packed_tensor(input, lengths, value, l_min=None):
    old_shape = input.shape
    if isinstance(lengths, nd.NDArray):
        max_len = as_scalar(input.max())
    else:
        max_len = builtins.max(lengths)

    if l_min is not None:
        max_len = builtins.max(max_len, l_min)

    batch_size = len(lengths)
    ctx = input.context
    dtype = input.dtype
    x = nd.full((batch_size * max_len, *old_shape[1:]), value, ctx=ctx, dtype=dtype)
    index = []
    for i, l in enumerate(lengths):
        index.extend(range(i * max_len, i * max_len + l))
    index = nd.array(index, ctx=ctx)
    return scatter_row(x, index, input).reshape(batch_size, max_len, *old_shape[1:]) 

def log_sum_exp(vec):
    max_score = nd.max(vec).asscalar()
    return nd.log(nd.sum(nd.exp(vec - max_score))) + max_score

# Model 

def _viterbi_decode(self, feats):
        backpointers = []

        # Initialize the viterbi variables in log space
        vvars = nd.full((1, self.tagset_size), -10000.)
        vvars[0, self.tag2idx[START_TAG]] = 0

        for feat in feats:
            bptrs_t = []  # holds the backpointers for this step
            viterbivars_t = []  # holds the viterbi variables for this step

            for next_tag in range(self.tagset_size):
                # next_tag_var[i] holds the viterbi variable for tag i at the
                # previous step, plus the score of transitioning
                # from tag i to next_tag.
                # We don't include the emission scores here because the max
                # does not depend on them (we add them in below)
                next_tag_var = vvars + self.transitions.data()[next_tag]
                best_tag_id = argmax(next_tag_var)
                bptrs_t.append(best_tag_id)
                viterbivars_t.append(next_tag_var[0, best_tag_id])
            # Now add in the emission scores, and assign vvars to the set
            # of viterbi variables we just computed
            vvars = (nd.concat(*viterbivars_t, dim=0) + feat).reshape((1, -1))
            backpointers.append(bptrs_t)

        # Transition to STOP_TAG
        terminal_var = vvars + self.transitions.data()[self.tag2idx[STOP_TAG]]
        best_tag_id = argmax(terminal_var)
        path_score = terminal_var[0, best_tag_id]

        # Follow the back pointers to decode the best path.
        best_path = [best_tag_id]
        for bptrs_t in reversed(backpointers):
            best_tag_id = bptrs_t[best_tag_id]
            best_path.append(best_tag_id)
        # Pop off the start tag (we dont want to return that to the caller)
        start = best_path.pop()
        assert start == self.tag2idx[START_TAG]  # Sanity check
        best_path.reverse()
        return path_score, best_path 

def max(input, dim):
    return nd.max(input, axis=dim) 

def reduce_max(input):
    return input.max() 

def hybrid_forward(self, F, preds, label):
        label = label.astype('float32')
        dist = F.sqrt(F.sum(F.square(preds), axis=1))

        return label * F.square(dist) + (1 - label) * F.square(F.max(self._m - dist, 0)) 

def hybrid_forward(self, F, fts, ys, ftt, yt):
        """
        Semantic Alignment Loss
        :param F: Function
        :param yt: label for the target domain [N]
        :param ftt: features for the target domain [N, K]
        :param ys: label for the source domain [M]
        :param fts: features for the source domain [M, K]
        :return:
        """
        if self._fn:
            # Normalize ft
            fts = F.L2Normalization(fts, mode='instance')
            ftt = F.L2Normalization(ftt, mode='instance')

        fts_rpt = F.broadcast_to(fts.expand_dims(axis=0), shape=(self._bs_tgt, self._bs_src, self._embed_size))
        ftt_rpt = F.broadcast_to(ftt.expand_dims(axis=1), shape=(self._bs_tgt, self._bs_src, self._embed_size))

        dists = F.sum(F.square(ftt_rpt - fts_rpt), axis=2)

        yt_rpt = F.broadcast_to(yt.expand_dims(axis=1), shape=(self._bs_tgt, self._bs_src)).astype('int32')
        ys_rpt = F.broadcast_to(ys.expand_dims(axis=0), shape=(self._bs_tgt, self._bs_src)).astype('int32')

        y_same = F.equal(yt_rpt, ys_rpt).astype('float32')
        y_diff = F.not_equal(yt_rpt, ys_rpt).astype('float32')

        intra_cls_dists = dists * y_same
        inter_cls_dists = dists * y_diff

        max_dists = F.max(dists, axis=1, keepdims=True)
        max_dists = F.broadcast_to(max_dists, shape=(self._bs_tgt, self._bs_src))
        revised_inter_cls_dists = F.where(y_same, max_dists, inter_cls_dists)

        max_intra_cls_dist = F.max(intra_cls_dists, axis=1)
        min_inter_cls_dist = F.min(revised_inter_cls_dists, axis=1)

        loss = F.relu(max_intra_cls_dist - min_inter_cls_dist + self._margin)

        return loss 

def log_sum_exp(vec):
    max_score = nd.max(vec).asscalar()
    return nd.log(nd.sum(nd.exp(vec - max_score))) + max_score

# Model 

def _viterbi_decode(self, feats):
        backpointers = []

        # Initialize the viterbi variables in log space
        vvars = nd.full((1, self.tagset_size), -10000.)
        vvars[0, self.tag2idx[START_TAG]] = 0

        for feat in feats:
            bptrs_t = []  # holds the backpointers for this step
            viterbivars_t = []  # holds the viterbi variables for this step

            for next_tag in range(self.tagset_size):
                # next_tag_var[i] holds the viterbi variable for tag i at the
                # previous step, plus the score of transitioning
                # from tag i to next_tag.
                # We don't include the emission scores here because the max
                # does not depend on them (we add them in below)
                next_tag_var = vvars + self.transitions.data()[next_tag]
                best_tag_id = argmax(next_tag_var)
                bptrs_t.append(best_tag_id)
                viterbivars_t.append(next_tag_var[0, best_tag_id])
            # Now add in the emission scores, and assign vvars to the set
            # of viterbi variables we just computed
            vvars = (nd.concat(*viterbivars_t, dim=0) + feat).reshape((1, -1))
            backpointers.append(bptrs_t)

        # Transition to STOP_TAG
        terminal_var = vvars + self.transitions.data()[self.tag2idx[STOP_TAG]]
        best_tag_id = argmax(terminal_var)
        path_score = terminal_var[0, best_tag_id]

        # Follow the back pointers to decode the best path.
        best_path = [best_tag_id]
        for bptrs_t in reversed(backpointers):
            best_tag_id = bptrs_t[best_tag_id]
            best_path.append(best_tag_id)
        # Pop off the start tag (we dont want to return that to the caller)
        start = best_path.pop()
        assert start == self.tag2idx[START_TAG]  # Sanity check
        best_path.reverse()
        return path_score, best_path 

def log_sum_exp(vec):
    max_score = nd.max(vec).asscalar()
    return nd.log(nd.sum(nd.exp(vec - max_score))) + max_score 

def log_sum_exp(vec):
    max_score = nd.max(vec).asscalar()
    return nd.log(nd.sum(nd.exp(vec - max_score))) + max_score

# Model 

def _viterbi_decode(self, feats):
        backpointers = []

        # Initialize the viterbi variables in log space
        vvars = nd.full((1, self.tagset_size), -10000.)
        vvars[0, self.tag2idx[START_TAG]] = 0

        for feat in feats:
            bptrs_t = []  # holds the backpointers for this step
            viterbivars_t = []  # holds the viterbi variables for this step

            for next_tag in range(self.tagset_size):
                # next_tag_var[i] holds the viterbi variable for tag i at the
                # previous step, plus the score of transitioning
                # from tag i to next_tag.
                # We don't include the emission scores here because the max
                # does not depend on them (we add them in below)
                next_tag_var = vvars + self.transitions[next_tag]
                best_tag_id = argmax(next_tag_var)
                bptrs_t.append(best_tag_id)
                viterbivars_t.append(next_tag_var[0, best_tag_id])
            # Now add in the emission scores, and assign vvars to the set
            # of viterbi variables we just computed
            vvars = (nd.concat(*viterbivars_t, dim=0) + feat).reshape((1, -1))
            backpointers.append(bptrs_t)

        # Transition to STOP_TAG
        terminal_var = vvars + self.transitions[self.tag2idx[STOP_TAG]]
        best_tag_id = argmax(terminal_var)
        path_score = terminal_var[0, best_tag_id]

        # Follow the back pointers to decode the best path.
        best_path = [best_tag_id]
        for bptrs_t in reversed(backpointers):
            best_tag_id = bptrs_t[best_tag_id]
            best_path.append(best_tag_id)
        # Pop off the start tag (we dont want to return that to the caller)
        start = best_path.pop()
        assert start == self.tag2idx[START_TAG]  # Sanity check
        best_path.reverse()
        return path_score, best_path