FIX: bugs in evaluator

recbole/data/dataloader/general_dataloader.py

@@ -220,12 +220,11 @@ def __init__(self, config, dataset, sampler, neg_sample_args,
 
         dataset.sort(by=uid_field, ascending=True)
         last_uid = None
-        positive_item = None
+        positive_item = set()
         uid2used_item = sampler.used_ids
         for uid, iid in zip(dataset.inter_feat[uid_field].numpy(), dataset.inter_feat[iid_field].numpy()):
             if uid != last_uid:
-                if last_uid is not None:
-                    self._set_user_property(last_uid, uid2used_item[last_uid], positive_item)
+                self._set_user_property(last_uid, uid2used_item[last_uid], positive_item)
                 last_uid = uid
                 self.uid_list.append(uid)
                 positive_item = set()
@@ -238,6 +237,8 @@ def __init__(self, config, dataset, sampler, neg_sample_args,
                          batch_size=batch_size, dl_format=dl_format, shuffle=shuffle)
 
     def _set_user_property(self, uid, used_item, positive_item):
+        if uid is None:
+            return
         history_item = used_item - positive_item
         positive_item_num = len(positive_item)
         self.uid2items_num[uid] = positive_item_num
@@ -260,17 +261,16 @@ def _shuffle(self):
         self.logger.warnning('GeneralFullDataLoader can\'t shuffle')
 
     def _next_batch_data(self):
-        index = slice(self.pr, self.pr + self.step)
-        user_df = self.user_df[index]
-        pos_len_list = self.uid2items_num[self.uid_list[index]]
-        user_len_list = np.full(len(user_df), self.item_num)
-        user_df.set_additional_info(pos_len_list, user_len_list)
+        user_df = self.user_df[self.pr: self.pr + self.step]
         cur_data = self._neg_sampling(user_df)
         self.pr += self.step
         return cur_data
 
     def _neg_sampling(self, user_df):
         uid_list = list(user_df[self.dataset.uid_field])
+        pos_len_list = self.uid2items_num[uid_list]
+        user_len_list = np.full(len(uid_list), self.item_num)
+        user_df.set_additional_info(pos_len_list, user_len_list)
 
         history_item = self.uid2history_item[uid_list]
         history_row = torch.cat([torch.full_like(hist_iid, i) for i, hist_iid in enumerate(history_item)])

recbole/data/dataset/dataset.py

@@ -1340,6 +1340,11 @@ def build(self, eval_setting):
         Returns:
             list: List of builded :class:`Dataset`.
         """
+        if self.benchmark_filename_list is not None:
+            cumsum = list(np.cumsum(self.file_size_list))
+            datasets = [self.copy(self.inter_feat[start: end]) for start, end in zip([0] + cumsum[:-1], cumsum)]
+            return datasets
+
         ordering_args = eval_setting.ordering_args
         if ordering_args['strategy'] == 'shuffle':
             self.shuffle()

recbole/data/utils.py

@@ -121,7 +121,7 @@ def data_preparation(config, dataset, save=False):
         getattr(es, es_str[1])()
         if 'sampler' not in locals():
             sampler = Sampler(phases, builded_datasets, es.neg_sample_args['distribution'])
-        sampler.set_distribution(es.neg_sample_args['distribution'])   
+        sampler.set_distribution(es.neg_sample_args['distribution'])
         kwargs['sampler'] = [sampler.set_phase('valid'), sampler.set_phase('test')]
         kwargs['neg_sample_args'] = copy.deepcopy(es.neg_sample_args)
     valid_data, test_data = dataloader_construct(

recbole/evaluator/abstract_evaluator.py

@@ -4,7 +4,7 @@
 # @email   :   tsotfsk@outlook.com
 
 # UPDATE
-# @Time    :   2020/10/21, 2020/12/9
+# @Time    :   2020/10/21, 2020/12/18
 # @Author  :   Kaiyuan Li, Zhichao Feng
 # @email   :   tsotfsk@outlook.com, fzcbupt@gmail.com
 
@@ -99,7 +99,7 @@ class IndividualEvaluator(BaseEvaluator):
     """
     def __init__(self, config, metrics):
         super().__init__(config, metrics)
-        pass
+        self._check_args()
 
     def sample_collect(self, true_scores, pred_scores):
         """It is called when evaluation sample distribution is `uniform` or `popularity`.
@@ -127,3 +127,7 @@ def get_score_matrix(self, true_scores, pred_scores):
             scores_matrix = self.sample_collect(true_scores, pred_scores)
 
         return scores_matrix
+
+    def _check_args(self):
+        if self.full:
+            raise NotImplementedError('full sort can\'t use IndividualEvaluator')

recbole/evaluator/evaluators.py

@@ -4,7 +4,7 @@
 # @email   :   tsotfsk@outlook.com
 
 # UPDATE
-# @Time    :   2020/08/04, 2020/08/11, 2020/12/9
+# @Time    :   2020/08/04, 2020/08/11, 2020/12/18
 # @Author  :   Kaiyuan Li, Yupeng Hou, Zhichao Feng
 # @email   :   tsotfsk@outlook.com, houyupeng@ruc.edu.cn, fzcbupt@gmail.com
 
@@ -116,6 +116,8 @@ def _calculate_metrics(self, pos_len_list, topk_index):
             metric_fuc = metrics_dict[metric.lower()]
             result = metric_fuc(pos_idx_matrix, pos_len_list)
             result_list.append(result)  # n_users x len(metrics) x len(ranks)
+        import pdb;
+        pdb.set_trace()
         result = np.stack(result_list, axis=0).mean(axis=1)  # len(metrics) x len(ranks)
         return result
 
@@ -158,22 +160,40 @@ def get_user_pos_len_list(self, interaction, scores_tensor):
         user_len_list = interaction.user_len_list
         return pos_len_list, user_len_list
 
-    def get_pos_index(self, scores_tensor, pos_len_list, user_len_list):
-        """get the index of positive items
+    def average_rank(self, scores):
+        """Get the ranking of an ordered tensor, and take the average of the ranking for positions with equal values.
 
-       Args:
-           scores_tensor (tensor): the tensor of model output with size of `(N, )`
-           pos_len_list(list): number of positive items
-           user_len_list(list): number of all items
+        Args:
+            scores(tensor): an ordered tensor, with size of `(N, )`
+
+        Returns:
+            torch.Tensor: average_rank
 
-       Returns:
-           tensor: a matrix indicating whether the corresponding item is positive
+        Example:
+            >>> average_rank(tensor([[1,2,2,2,3,3,6],[2,2,2,2,4,4,5]]))
+            tensor([[1.0000, 3.0000, 3.0000, 3.0000, 5.5000, 5.5000, 7.0000],
+            [2.5000, 2.5000, 2.5000, 2.5000, 5.0000, 6.5000, 6.5000]])
 
-       """
-        scores_matrix = self.get_score_matrix(scores_tensor, user_len_list)
-        _, n_index = torch.sort(scores_matrix, dim=-1, descending=True)
-        pos_index = (n_index < pos_len_list.reshape(-1, 1))
-        return pos_index
+        Reference:
+            https://github.com/scipy/scipy/blob/v0.17.1/scipy/stats/stats.py#L5262-L5352
+
+        """
+        length, width = scores.shape
+        device = scores.device
+        true_tensor = torch.full((length, 1), True, dtype=torch.bool, device=device)
+
+        obs = torch.cat([true_tensor, scores[:, 1:] != scores[:, :-1]], dim=1)
+        # bias added to dense
+        bias = torch.arange(0, length, device=device).repeat(width).reshape(width, -1). \
+            transpose(1, 0).reshape(-1)
+        dense = obs.view(-1).cumsum(0) + bias
+
+        # cumulative counts of each unique value
+        count = torch.where(torch.cat([obs, true_tensor], dim=1))[1]
+        # get average rank
+        avg_rank = .5 * (count[dense] + count[dense - 1] + 1).view(length, -1)
+
+        return avg_rank
 
     def collect(self, interaction, scores_tensor):
         """collect the rank intermediate result of one batch, this function mainly implements ranking
@@ -185,10 +205,16 @@ def collect(self, interaction, scores_tensor):
 
         """
         pos_len_list, user_len_list = self.get_user_pos_len_list(interaction, scores_tensor)
-        pos_index = self.get_pos_index(scores_tensor, pos_len_list, user_len_list)
-        index_list = torch.arange(1, pos_index.shape[1] + 1).to(pos_index.device)
-        pos_rank_sum = torch.where(pos_index, index_list, torch.zeros_like(index_list)). \
+        scores_matrix = self.get_score_matrix(scores_tensor, user_len_list)
+        desc_scores, desc_index = torch.sort(scores_matrix, dim=-1, descending=True)
+
+        # get the index of positive items in the ranking list
+        pos_index = (desc_index < pos_len_list.reshape(-1, 1))
+
+        avg_rank = self.average_rank(desc_scores)
+        pos_rank_sum = torch.where(pos_index, avg_rank, torch.zeros_like(avg_rank)). \
             sum(axis=-1).reshape(-1, 1)
+
         return pos_rank_sum
 
     def evaluate(self, batch_matrix_list, eval_data):
@@ -205,6 +231,7 @@ def evaluate(self, batch_matrix_list, eval_data):
         pos_len_list = eval_data.get_pos_len_list()
         user_len_list = eval_data.get_user_len_list()
         pos_rank_sum = torch.cat(batch_matrix_list, dim=0).cpu().numpy()
+        assert len(pos_len_list) == len(pos_rank_sum)
 
         # get metrics
         metric_dict = {}

recbole/evaluator/metrics.py

@@ -4,7 +4,7 @@
 # @email   :   tsotfsk@outlook.com
 
 # UPDATE
-# @Time    :   2020/08/12, 2020/12/9, 2020/9/16
+# @Time    :   2020/08/12, 2020/12/21, 2020/9/16
 # @Author  :   Kaiyuan Li, Zhichao Feng, Xingyu Pan
 # @email   :   tsotfsk@outlook.com, fzcbupt@gmail.com, panxy@ruc.edu.cn
 
@@ -23,7 +23,6 @@
 
 #    TopK Metrics    #
 
-
 def hit_(pos_index, pos_len):
     r"""Hit_ (also known as hit ratio at :math:`N`) is a way of calculating how many 'hits' you have
     in an n-sized list of ranked items.
@@ -129,7 +128,6 @@ def ndcg_(pos_index, pos_len):
     :math:`U^{te}` is for all users in the test set.
 
     """
-
     len_rank = np.full_like(pos_len, pos_index.shape[1])
     idcg_len = np.where(pos_len > len_rank, len_rank, pos_len)
 
@@ -166,11 +164,54 @@ def precision_(pos_index, pos_len):
 
 
 def gauc_(user_len_list, pos_len_list, pos_rank_sum):
-    frac = user_len_list - (pos_len_list - 1) / 2 - (1 / pos_len_list) * np.squeeze(pos_rank_sum)
-    neg_item_num = user_len_list - pos_len_list
-    user_auc = frac / neg_item_num
+    r"""GAUC_ (also known as Group Area Under Curve) is used to evaluate the two-class model, referring to
+    the area under the ROC curve grouped by user.
+
+    .. _GAUC: https://dl.acm.org/doi/10.1145/3219819.3219823
+
+    Note:
+        It calculates the AUC score of each user, and finally obtains GAUC by weighting the user AUC.
+        It is also not limited to k. Due to our padding for `scores_tensor` in `RankEvaluator` with
+        `-np.inf`, the padding value will influence the ranks of origin items. Therefore, we use
+        descending sort here and make an identity transformation  to the formula of `AUC`, which is
+        shown in `auc_` function. For readability, we didn't do simplification in the code.
+
+    .. math::
+        \mathrm {GAUC} = \frac {{{M} \times {(M+N+1)} - \frac{M \times (M+1)}{2}} -
+        \sum\limits_{i=1}^M rank_{i}} {{M} \times {N}}
 
+    :math:`M` is the number of positive samples.
+    :math:`N` is the number of negative samples.
+    :math:`rank_i` is the descending rank of the ith positive sample.
+
+    """
+    neg_len_list = user_len_list - pos_len_list
+
+    # check positive and negative samples
+    all_with_pos = np.any(pos_len_list == 0)
+    all_with_neg = np.any(neg_len_list == 0)
+    non_zero_idx = np.full(len(user_len_list), True, dtype=np.bool)
+    if all_with_pos:
+        logger = getLogger()
+        logger.warning("No positive samples in some users, "
+                       "true positive value should be meaningless, "
+                       "these users have been removed from GAUC calculation")
+        non_zero_idx *= (pos_len_list != 0)
+    if all_with_neg:
+        logger = getLogger()
+        logger.warning("No negative samples in some users, "
+                       "false positive value should be meaningless, "
+                       "these users have been removed from GAUC calculation")
+        non_zero_idx *= (neg_len_list != 0)
+    if all_with_pos or all_with_neg:
+        item_list = user_len_list, neg_len_list, pos_len_list, pos_rank_sum
+        user_len_list, neg_len_list, pos_len_list, pos_rank_sum = \
+            map(lambda x: x[non_zero_idx], item_list)
+
+    pair_num = (user_len_list + 1) * pos_len_list - pos_len_list * (pos_len_list + 1) / 2 - np.squeeze(pos_rank_sum)
+    user_auc = pair_num / (neg_len_list * pos_len_list)
     result = (user_auc * pos_len_list).sum() / pos_len_list.sum()
+
     return result
 
 
@@ -188,11 +229,11 @@ def auc_(trues, preds):
 
     .. math::
         \mathrm {AUC} = \frac{\sum\limits_{i=1}^M rank_{i}
-        - {{M} \times {(M+1)}}} {{M} \times {N}}
+        - \frac {{M} \times {(M+1)}}{2}} {{{M} \times {N}}}
 
     :math:`M` is the number of positive samples.
     :math:`N` is the number of negative samples.
-    :math:`rank_i` is the rank of the ith positive sample.
+    :math:`rank_i` is the ascending rank of the ith positive sample.
 
     """
     fps, tps = _binary_clf_curve(trues, preds)

recbole/evaluator/proxy_evaluator.py

@@ -32,9 +32,9 @@ def build(self):
 
         """
         evaluator_list = []
-        metrics_set = {metric.lower() for metric in self.metrics}
+        metrics_list = [metric.lower() for metric in self.metrics]
         for metrics, evaluator in metric_eval_bind:
-            used_metrics = list(metrics_set.intersection(set(metrics.keys())))
+            used_metrics = [metric for metric in metrics_list if metric in metrics.keys()]
             if used_metrics:
                 evaluator_list.append(evaluator(self.config, used_metrics))
         return evaluator_list

recbole/model/knowledge_aware_recommender/kgat.py

@@ -268,7 +268,7 @@ def update_attentive_A(self):
         # Current PyTorch version does not support softmax on SparseCUDA, temporarily move to CPU to calculate softmax
         A_in = torch.sparse.FloatTensor(indices, kg_score, self.matrix_size).cpu()
         A_in = torch.sparse.softmax(A_in, dim=1).to(self.device)
-        self.A_in = copy.copy(A_in)
+        self.A_in = A_in
 
     def predict(self, interaction):
         user = interaction[self.USER_ID]

recbole/trainer/trainer.py

@@ -324,6 +324,9 @@ def evaluate(self, eval_data, load_best_model=True, model_file=None):
         Returns:
             dict: eval result, key is the eval metric and value in the corresponding metric value
         """
+        if not eval_data:
+            return
+
         if load_best_model:
             if model_file:
                 checkpoint_file = model_file
@@ -441,7 +444,9 @@ def _train_epoch(self, train_data, epoch_idx, loss_func=None):
         kg_total_loss = super()._train_epoch(train_data, epoch_idx, self.model.calculate_kg_loss)
 
         # update A
-        self.model.update_attentive_A()
+        self.model.eval()
+        with torch.no_grad():
+            self.model.update_attentive_A()
 
         return rs_total_loss, kg_total_loss