Python tensorflow.keras.backend.batch_dot() Examples

def call(self, inputs):
        if self.data_mode == 'disjoint':
            X, I = inputs
            if K.ndim(I) == 2:
                I = I[:, 0]
        else:
            X = inputs
        attn_coeff = K.dot(X, self.attn_kernel)
        attn_coeff = K.squeeze(attn_coeff, -1)
        attn_coeff = K.softmax(attn_coeff)
        if self.data_mode == 'single':
            output = K.dot(attn_coeff[None, ...], X)
        elif self.data_mode == 'batch':
            output = K.batch_dot(attn_coeff, X)
        else:
            output = attn_coeff[:, None] * X
            output = tf.math.segment_sum(output, I)

        return output 

def _build_tf_cosine_similarity(max_rank=0, offset=1, eps=1e-12):
    # We build the graph (See utils.generic_utils.tf_recall_at_k for original implementation):
    tf_db = K.placeholder(ndim=2, dtype=K.floatx())  # Where to find
    tf_labels = K.placeholder(ndim=1, dtype=K.floatx())  # and their labels

    tf_batch_query = K.placeholder(ndim=2, dtype=K.floatx())  # Used in case of memory issues
    batch_labels = K.placeholder(ndim=2, dtype=K.floatx())  # and their labels

    all_representations_T = K.expand_dims(tf_db, axis=0)  # 1 x D x N
    batch_representations = K.expand_dims(tf_batch_query, axis=0)  # 1 x n x D
    sim = K.batch_dot(batch_representations, all_representations_T)  # 1 x n x N
    sim = K.squeeze(sim, axis=0)  # n x N
    sim /= tf.linalg.norm(tf_batch_query, axis=1, keepdims=True) + eps
    sim /= tf.linalg.norm(tf_db, axis=0, keepdims=True) + eps

    if max_rank > 0:  # computing r@K or mAP@K
        index_ranking = tf.nn.top_k(sim, k=max_rank + offset).indices
    else:
        index_ranking = tf.contrib.framework.argsort(sim, axis=-1, direction='DESCENDING', stable=True)

    top_k = index_ranking[:, offset:]
    tf_ranking = tf.gather(tf_labels, top_k)

    return tf_db, tf_labels, tf_batch_query, batch_labels, tf_ranking 

def call(self, x, **kwargs):
        assert isinstance(x, list)
        inp_a, inp_b = x
        last_state = K.expand_dims(inp_b[:, -1, :], 1)
        m = []
        for i in range(self.output_dim):
            outp_a = inp_a * self.W[i]
            outp_last = last_state * self.W[i]
            outp_a = K.l2_normalize(outp_a, -1)
            outp_last = K.l2_normalize(outp_last, -1)
            outp = K.batch_dot(outp_a, outp_last, axes=[2, 2])
            m.append(outp)
        if self.output_dim > 1:
            persp = K.concatenate(m, 2)
        else:
            persp = m[0]
        return [persp, persp] 

def call(self, x, **kwargs):
        assert isinstance(x, list)
        inp_a, inp_b = x
        m = []
        for i in range(self.output_dim):
            outp_a = inp_a * self.W[i]
            outp_b = inp_b * self.W[i]
            outp_a = K.l2_normalize(outp_a, -1)
            outp_b = K.l2_normalize(outp_b, -1)
            outp = K.batch_dot(outp_a, outp_b, axes=[2, 2])
            outp = K.max(outp, -1, keepdims=True)
            m.append(outp)
        if self.output_dim > 1:
            persp = K.concatenate(m, 2)
        else:
            persp = m[0]
        return [persp, persp] 

def call(self, x, **kwargs):
        assert isinstance(x, list)
        inp_a, inp_b = x

        outp_a = K.l2_normalize(inp_a, -1)
        outp_b = K.l2_normalize(inp_b, -1)
        alpha = K.batch_dot(outp_b, outp_a, axes=[1, 1])
        alpha = K.l2_normalize(alpha, 1)
        hmean = K.batch_dot(outp_b, alpha, axes=[2, 1])
        kcon = K.eye(K.int_shape(inp_a)[1], dtype='float32')

        m = []
        for i in range(self.output_dim):
            outp_a = inp_a * self.W[i]
            outp_hmean = hmean * self.W[i]
            outp_a = K.l2_normalize(outp_a, -1)
            outp_hmean = K.l2_normalize(outp_hmean, -1)
            outp = K.batch_dot(outp_hmean, outp_a, axes=[2, 2])
            outp = K.sum(outp * kcon, -1, keepdims=True)
            m.append(outp)
        if self.output_dim > 1:
            persp = K.concatenate(m, 2)
        else:
            persp = m[0]
        return [persp, persp] 

def call(self, x):
        x_orig = x

        # x reshape
        this_bs_int = K.shape(x)[0]
        this_bs = tf.cast(this_bs_int, 'float32')  # this batch size
        prev_count = self.count
        x = K.batch_flatten(x)  # B x N

        # update mean
        new_mean, new_count = _mean_update(self.mean, self.count, x, self.cap)        

        # new C update. Should be B x N x N
        x = K.expand_dims(x, -1)
        C_delta = K.batch_dot(x, K.permute_dimensions(x, [0, 2, 1]))

        # update cov
        prev_cap = K.minimum(prev_count, self.cap)
        C = self.cov * (prev_cap - 1) + K.sum(C_delta, 0)
        new_cov = C / (prev_cap + this_bs - 1)

        # updates
        updates = [(self.count, new_count), (self.mean, new_mean), (self.cov, new_cov)]
        self.add_update(updates, x_orig)

        # prep for broadcasting :(
        p = tf.concat((K.reshape(this_bs_int, (1,)), K.shape(self.cov)), 0)
        z = tf.ones(p)

        return K.minimum(1., new_count/self.cap) * (z * K.expand_dims(new_cov, 0)) 

def mi_loss(self, y_true, y_pred):
        """Mutual information loss computed from the joint
           distribution matrix and the marginals

        Arguments:
            y_true (tensor): Not used since this is
                unsupervised learning
            y_pred (tensor): stack of softmax predictions for
                the Siamese latent vectors (Z and Zbar)
        """
        size = self.args.batch_size
        n_labels = y_pred.shape[-1]
        # lower half is Z
        Z = y_pred[0: size, :]
        Z = K.expand_dims(Z, axis=2)
        # upper half is Zbar
        Zbar = y_pred[size: y_pred.shape[0], :]
        Zbar = K.expand_dims(Zbar, axis=1)
        # compute joint distribution (Eq 10.3.2 & .3)
        P = K.batch_dot(Z, Zbar)
        P = K.sum(P, axis=0)
        # enforce symmetric joint distribution (Eq 10.3.4)
        P = (P + K.transpose(P)) / 2.0
        # normalization of total probability to 1.0
        P = P / K.sum(P)
        # marginal distributions (Eq 10.3.5 & .6)
        Pi = K.expand_dims(K.sum(P, axis=1), axis=1)
        Pj = K.expand_dims(K.sum(P, axis=0), axis=0)
        Pi = K.repeat_elements(Pi, rep=n_labels, axis=1)
        Pj = K.repeat_elements(Pj, rep=n_labels, axis=0)
        P = K.clip(P, K.epsilon(), np.finfo(float).max)
        Pi = K.clip(Pi, K.epsilon(), np.finfo(float).max)
        Pj = K.clip(Pj, K.epsilon(), np.finfo(float).max)
        # negative MI loss (Eq 10.3.7)
        neg_mi = K.sum((P * (K.log(Pi) + K.log(Pj) - K.log(P))))
        # each head contribute 1/n_heads to the total loss
        return neg_mi/self.args.heads 

def call(self, inputs, mask=None):
        """
        convert to query, key, value vectors, shaped [batch_size*num_head, time_step, embed_dim]
        """
        multihead_query = K.concatenate(tf.split(K.dot(inputs, self.w_q),
                                                 self.num_heads, axis=2), axis=0)
        multihead_key = K.concatenate(tf.split(K.dot(inputs, self.w_k),
                                               self.num_heads, axis=2), axis=0)
        multihead_value = K.concatenate(tf.split(K.dot(inputs, self.w_v),
                                                 self.num_heads, axis=2), axis=0)

        """scaled dot product"""
        scaled = K.int_shape(inputs)[-1] ** -0.5
        attend = K.batch_dot(multihead_query, multihead_key, axes=2) * scaled
        # apply mask before normalization (softmax)
        if mask is not None:
            multihead_mask = K.tile(mask, [self.num_heads, 1])
            attend *= K.expand_dims(K.cast(multihead_mask, K.floatx()), 2)
            attend *= K.expand_dims(K.cast(multihead_mask, K.floatx()), 1)
        # normalization
        attend = attend / K.cast(K.sum(attend, axis=-1, keepdims=True) + K.epsilon(), K.floatx())
        # apply attention
        attend = K.batch_dot(attend, multihead_value, axes=(2, 1))
        attend = tf.concat(tf.split(attend, self.num_heads, axis=0), axis=2)
        attend = K.dot(attend, self.w_final)

        if self.residual:
            attend = attend + inputs
        if self.normalize:
            mean = K.mean(attend, axis=-1, keepdims=True)
            std = K.mean(attend, axis=-1, keepdims=True)
            attend = self.gamma * (attend - mean) / (std + K.epsilon()) + self.beta

        return attend 

def _build_tf_l2_similarity(max_rank=0, offset=1):
    # We build the graph (See utils.generic_utils.tf_recall_at_k for original implementation):
    tf_db = K.placeholder(ndim=2, dtype=K.floatx())  # Where to find
    tf_labels = K.placeholder(ndim=1, dtype=K.floatx())  # and their labels

    tf_batch_query = K.placeholder(ndim=2, dtype=K.floatx())  # Used in case of memory issues
    batch_labels = K.placeholder(ndim=2, dtype=K.floatx())  # and their labels

    all_representations_T = K.expand_dims(tf_db, axis=0)  # 1 x D x N
    batch_representations = K.expand_dims(tf_batch_query, axis=0)  # 1 x n x D
    dist = -2. * K.batch_dot(batch_representations, all_representations_T)  # 1 x n x N
    dist = K.squeeze(dist, axis=0)  # n x N
    dist += K.sum(tf_batch_query * tf_batch_query, axis=1, keepdims=True)
    dist += K.sum(tf_db * tf_db, axis=0, keepdims=True)

    if max_rank > 0:  # computing r@K or mAP@K
        # top_k finds the k greatest entries and we want the lowest. Note that distance with itself will be last ranked
        dist = -dist
        index_ranking = tf.nn.top_k(dist, k=max_rank + offset).indices
    else:
        index_ranking = tf.contrib.framework.argsort(dist, axis=-1, direction='ASCENDING', stable=True)

    index_ranking = index_ranking[:, offset:]

    tf_ranking = tf.gather(tf_labels, index_ranking)

    return tf_db, tf_labels, tf_batch_query, batch_labels, tf_ranking 

def _pairwise_distances(self, inputs: List[Tensor]) -> Tensor:
        emb_c, emb_r = inputs
        bs = K.shape(emb_c)[0]
        embeddings = K.concatenate([emb_c, emb_r], 0)
        dot_product = K.dot(embeddings, K.transpose(embeddings))
        square_norm = K.batch_dot(embeddings, embeddings, axes=1)
        distances = K.transpose(square_norm) - 2.0 * dot_product + square_norm
        distances = distances[0:bs, bs:bs+bs]
        distances = K.clip(distances, 0.0, None)
        mask = K.cast(K.equal(distances, 0.0), K.dtype(distances))
        distances = distances + mask * 1e-16
        distances = K.sqrt(distances)
        distances = distances * (1.0 - mask)
        return distances 

def biaffine_layer(deps: tf.Tensor, heads: tf.Tensor, deps_dim: int,
                   heads_dim: int, output_dim: int, name: str = "biaffine_layer") -> tf.Tensor:
    """Implements a biaffine layer from [Dozat, Manning, 2016].

    Args:
        deps: the 3D-tensor of dependency states,
        heads: the 3D-tensor of head states,
        deps_dim: the dimension of dependency states,
        heads_dim: the dimension of head_states,
        output_dim: the output dimension
        name: the name of a layer

    Returns:
        `answer` the output 3D-tensor

    """
    input_shape = [kb.shape(deps)[i] for i in range(tf.keras.backend.ndim(deps))]
    first_input = tf.reshape(deps, [-1, deps_dim])  # first_input.shape = (B*L, D1)
    second_input = tf.reshape(heads, [-1, heads_dim])  # second_input.shape = (B*L, D2)
    with tf.variable_scope(name):
        kernel_shape = (deps_dim, heads_dim * output_dim)
        kernel = tf.get_variable('kernel', shape=kernel_shape, initializer=xavier_initializer())
        first = tf.matmul(first_input, kernel)  # (B*L, D2*H)
        first = tf.reshape(first, [-1, heads_dim, output_dim])  # (B*L, D2, H)
        answer = kb.batch_dot(first, second_input, axes=[1, 1])  # (B*L, H)
        first_bias = tf.get_variable('first_bias', shape=(deps_dim, output_dim),
                                     initializer=xavier_initializer())
        answer += tf.matmul(first_input, first_bias)
        second_bias = tf.get_variable('second_bias', shape=(heads_dim, output_dim),
                                      initializer=xavier_initializer())
        answer += tf.matmul(second_input, second_bias)
        label_bias = tf.get_variable('label_bias', shape=(output_dim,),
                                     initializer=xavier_initializer())
        answer = kb.bias_add(answer, label_bias)
        answer = tf.reshape(answer, input_shape[:-1] + [output_dim])  # (B, L, H)
    return answer 

def call(self, args):

        if not isinstance(args, (list, tuple)):
            args = [args]
        self.cargs = len(args)

        # flatten
        if len(args) == 2:  # input y, m
            # get inputs
            y, y_mask = args 
            a_fact = int(y.get_shape().as_list()[-1] / y_mask.get_shape().as_list()[-1])
            y_mask = K.repeat_elements(y_mask, a_fact, -1)
            y_flat = K.batch_flatten(y)  # N x D
            y_mask_flat = K.batch_flatten(y_mask)  # N x D

            # prepare switching matrix
            W = self.W # d x D

            w_tmp = K.expand_dims(W, 0)  # 1 x d x D
            Wo = K.permute_dimensions(w_tmp, [0, 2, 1]) * K.expand_dims(y_mask_flat, -1)  # N x D x d
            WoT = K.permute_dimensions(Wo, [0, 2, 1])    # N x d x D
            WotWo_inv = tf.matrix_inverse(K.batch_dot(WoT, Wo))  # N x d x d
            pre = K.batch_dot(WotWo_inv, WoT) # N x d x D
            res = K.batch_dot(pre, y_flat)  # N x d

            if self.use_bias:
                res += K.expand_dims(self.bias, 0)

        else:
            x_data = args[0]
            shape = K.shape(x_data)

            x_data = K.batch_flatten(x_data)  # N x d

            if self.use_bias:
                x_data -= self.bias

            res = K.dot(x_data, self.W)

            # reshape
            # Here you can mix integers and symbolic elements of `shape`
            pool_shape = tf.stack([shape[0], *self.orig_input_shape])
            res = K.reshape(res, pool_shape)

        return res