Beispiel #1
0
    def signature(self):
        """
        This property computes the signature of the dataset, which can be
        passed to `spektral.data.utils.to_tf_signature(signature)` to compute
        the TensorFlow signature. You can safely ignore this property unless
        you are creating a custom `Loader`.

        A signature consist of the TensorFlow TypeSpec, shape, and dtype of
        all characteristic matrices of the graphs in the Dataset. This is
        returned as a dictionary of dictionaries, with keys `x`, `a`, `e`, and
        `y` for the four main data matrices.

        Each sub-dictionary will have keys `spec`, `shape` and `dtype`.
        """
        if len(self.graphs) == 0:
            return None
        signature = {}
        graph = self.graphs[0]  # This is always non-empty
        if graph.x is not None:
            signature["x"] = dict()
            signature["x"]["spec"] = get_spec(graph.x)
            signature["x"]["shape"] = (None, self.n_node_features)
            signature["x"]["dtype"] = tf.as_dtype(graph.x.dtype)
        if graph.a is not None:
            signature["a"] = dict()
            signature["a"]["spec"] = get_spec(graph.a)
            signature["a"]["shape"] = (None, None)
            signature["a"]["dtype"] = tf.as_dtype(graph.a.dtype)
        if graph.e is not None:
            signature["e"] = dict()
            signature["e"]["spec"] = get_spec(graph.e)
            signature["e"]["shape"] = (None, self.n_edge_features)
            signature["e"]["dtype"] = tf.as_dtype(graph.e.dtype)
        if graph.y is not None:
            signature["y"] = dict()
            signature["y"]["spec"] = get_spec(graph.y)
            signature["y"]["shape"] = (self.n_labels, )
            signature["y"]["dtype"] = tf.as_dtype(np.array(graph.y).dtype)
        return signature
Beispiel #2
0
    def tf_signature(self):
        """
        Adjacency matrix has shape [n_nodes, n_nodes]
        Node features have shape [batch, n_nodes, n_node_features]
        Edge features have shape [batch, n_edges, n_edge_features]
        Targets have shape [batch, ..., n_labels]
        """
        signature = self.dataset.signature
        for k in ["x", "e", "y"]:
            if k in signature:
                signature[k]["shape"] = prepend_none(signature[k]["shape"])

        signature["a"] = dict()
        signature["a"]["spec"] = get_spec(self.dataset.a)
        signature["a"]["shape"] = (None, None)
        signature["a"]["dtype"] = tf.as_dtype(self.dataset.a.dtype)

        return to_tf_signature(signature)