research/gam/gam/trainer/trainer_classification.py [578:702]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - self.labels_ll_left: labels_src, self.labels_ll_right: labels_tgt }) if pair_lu_iterator is not None: indices_src, indices_tgt, features_src, features_tgt, labels_src, _ = ( next(pair_lu_iterator)) feed_dict.update({ self.indices_lu_left: indices_src, self.indices_lu_right: indices_tgt, self.features_lu_left: features_src, self.features_lu_right: features_tgt, self.labels_lu_left: labels_src }) if pair_uu_iterator is not None: indices_src, indices_tgt, features_src, features_tgt, _, _ = next( pair_uu_iterator) feed_dict.update({ self.indices_uu_left: indices_src, self.indices_uu_right: indices_tgt, self.features_uu_left: features_src, self.features_uu_right: features_tgt }) return feed_dict except StopIteration: # If the iterator has finished, return None. return None def pair_iterator(self, src_indices, tgt_indices, batch_size, data): """Iterator over pairs of samples. The first element of the pair is selected from the src_indices, and the second element is selected from tgt_indices. Args: src_indices: Numpy array containing the indices available for the source node. tgt_indices: Numpy array containing the indices available for the tgt node. batch_size: An integer representing the desired batch size. data: A CotrainDataset object used to extract the features and labels. Yields: indices_src, indices_tgt, features_src, features_tgt, labels_src, labels_tgt """ def _select_from_pool(indices): """Selects batch_size indices from the provided list.""" num_indices = len(indices) if num_indices > 0: idxs = self.rng.randint(0, high=num_indices, size=(batch_size,)) indices_batch = indices[idxs] features_batch = data.get_features(indices_batch) labels_batch = data.get_labels(indices_batch) else: features_shape = [0] + list(data.features_shape) indices_batch = np.zeros(shape=(0,), dtype=np.int64) features_batch = np.zeros(shape=features_shape, dtype=np.float32) labels_batch = np.zeros(shape=(0,), dtype=np.int64) return indices_batch, features_batch, labels_batch while True: indices_src, features_src, labels_src = _select_from_pool(src_indices) indices_tgt, features_tgt, labels_tgt = _select_from_pool(tgt_indices) yield (indices_src, indices_tgt, features_src, features_tgt, labels_src, labels_tgt) def edge_iterator(self, data, batch_size, labeling): """An iterator over graph edges. Args: data: A CotrainDataset object used to extract the features and labels. batch_size: An integer representing the desired batch size. labeling: A string which can be `ll`, `lu` or `uu`, that is used to represent the type of edges to return, where `ll` refers to labeled-labeled, `lu` refers to labeled-unlabeled, and `uu` refers to unlabeled-unlabeled. Yields: indices_src, indices_tgt, features_src, features_tgt, labels_src, labels_tgt """ if labeling == 'll': edges = data.get_edges( src_labeled=True, tgt_labeled=True, label_must_match=True) elif labeling == 'lu': edges_lu = data.get_edges(src_labeled=True, tgt_labeled=False) edges_ul = data.get_edges(src_labeled=False, tgt_labeled=True) # Reverse the edges of UL to be LU. edges_ul = [e.copy(src=e.tgt, tgt=e.src) for e in edges_ul] edges = edges_lu + edges_ul elif labeling == 'uu': edges = data.get_edges(src_labeled=False, tgt_labeled=False) else: raise ValueError('Unsupported value for parameter `labeling`.') if not edges: indices = np.zeros(shape=(0,), dtype=np.int32) features = np.zeros( shape=[ 0, ] + list(data.features_shape), dtype=np.float32) labels = np.zeros(shape=(0,), dtype=np.int64) while True: yield (indices, indices, features, features, labels, labels) edges = np.stack([(e.src, e.tgt) for e in edges]) iterator = batch_iterator( inputs=edges, batch_size=batch_size, shuffle=True, allow_smaller_batch=False, repeat=True) for edge in iterator: indices_src = edge[:, 0] indices_tgt = edge[:, 1] features_src = data.get_features(indices_src) features_tgt = data.get_features(indices_tgt) labels_src = data.get_labels(indices_src) labels_tgt = data.get_labels(indices_tgt) yield (indices_src, indices_tgt, features_src, features_tgt, labels_src, labels_tgt) def _evaluate(self, indices, split, session, summary_writer): - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - research/gam/gam/trainer/trainer_classification_gcn.py [606:730]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - self.labels_ll_left: labels_src, self.labels_ll_right: labels_tgt }) if pair_lu_iterator is not None: indices_src, indices_tgt, features_src, features_tgt, labels_src, _ = ( next(pair_lu_iterator)) feed_dict.update({ self.indices_lu_left: indices_src, self.indices_lu_right: indices_tgt, self.features_lu_left: features_src, self.features_lu_right: features_tgt, self.labels_lu_left: labels_src }) if pair_uu_iterator is not None: indices_src, indices_tgt, features_src, features_tgt, _, _ = next( pair_uu_iterator) feed_dict.update({ self.indices_uu_left: indices_src, self.indices_uu_right: indices_tgt, self.features_uu_left: features_src, self.features_uu_right: features_tgt }) return feed_dict except StopIteration: # If the iterator has finished, return None. return None def pair_iterator(self, src_indices, tgt_indices, batch_size, data): """Iterator over pairs of samples. The first element of the pair is selected from the src_indices, and the second element is selected from tgt_indices. Args: src_indices: Numpy array containing the indices available for the source node. tgt_indices: Numpy array containing the indices available for the tgt node. batch_size: An integer representing the desired batch size. data: A CotrainDataset object used to extract the features and labels. Yields: indices_src, indices_tgt, features_src, features_tgt, labels_src, labels_tgt """ def _select_from_pool(indices): """Selects batch_size indices from the provided list.""" num_indices = len(indices) if num_indices > 0: idxs = self.rng.randint(0, high=num_indices, size=(batch_size,)) indices_batch = indices[idxs] features_batch = data.get_features(indices_batch) labels_batch = data.get_labels(indices_batch) else: features_shape = [0] + list(data.features_shape) indices_batch = np.zeros(shape=(0,), dtype=np.int64) features_batch = np.zeros(shape=features_shape, dtype=np.float32) labels_batch = np.zeros(shape=(0,), dtype=np.int64) return indices_batch, features_batch, labels_batch while True: indices_src, features_src, labels_src = _select_from_pool(src_indices) indices_tgt, features_tgt, labels_tgt = _select_from_pool(tgt_indices) yield (indices_src, indices_tgt, features_src, features_tgt, labels_src, labels_tgt) def edge_iterator(self, data, batch_size, labeling): """An iterator over graph edges. Args: data: A CotrainDataset object used to extract the features and labels. batch_size: An integer representing the desired batch size. labeling: A string which can be `ll`, `lu` or `uu`, that is used to represent the type of edges to return, where `ll` refers to labeled-labeled, `lu` refers to labeled-unlabeled, and `uu` refers to unlabeled-unlabeled. Yields: indices_src, indices_tgt, features_src, features_tgt, labels_src, labels_tgt """ if labeling == 'll': edges = data.get_edges( src_labeled=True, tgt_labeled=True, label_must_match=True) elif labeling == 'lu': edges_lu = data.get_edges(src_labeled=True, tgt_labeled=False) edges_ul = data.get_edges(src_labeled=False, tgt_labeled=True) # Reverse the edges of UL to be LU. edges_ul = [e.copy(src=e.tgt, tgt=e.src) for e in edges_ul] edges = edges_lu + edges_ul elif labeling == 'uu': edges = data.get_edges(src_labeled=False, tgt_labeled=False) else: raise ValueError('Unsupported value for parameter `labeling`.') if not edges: indices = np.zeros(shape=(0,), dtype=np.int32) features = np.zeros( shape=[ 0, ] + list(data.features_shape), dtype=np.float32) labels = np.zeros(shape=(0,), dtype=np.int64) while True: yield (indices, indices, features, features, labels, labels) edges = np.stack([(e.src, e.tgt) for e in edges]) iterator = batch_iterator( inputs=edges, batch_size=batch_size, shuffle=True, allow_smaller_batch=False, repeat=True) for edge in iterator: indices_src = edge[:, 0] indices_tgt = edge[:, 1] features_src = data.get_features(indices_src) features_tgt = data.get_features(indices_tgt) labels_src = data.get_labels(indices_src) labels_tgt = data.get_labels(indices_tgt) yield (indices_src, indices_tgt, features_src, features_tgt, labels_src, labels_tgt) def _evaluate(self, indices, split, session, summary_writer): - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -