Negative ratings

configs/train/siren_train_config.json

@@ -0,0 +1,167 @@
+{
+  "experiment_name": "siren_ml1m_test",
+  "dataset": {
+    "type": "negative_ratings_graph",
+    "positive_domain": "positive",
+    "graph_dir_path": "../data/ml-1m",
+    "dataset": {
+      "type": "negative_rating",
+      "positive_domain": "positive",
+      "negative_domain": "negative",
+      "path_to_data_dir": "../data",
+      "name": "ml-1m",
+      "max_sequence_length": 50,
+      "samplers": {
+        "num_negatives_val": 100,
+        "type": "negative_ratings_next_item_prediction",
+        "negative_sampler_type": "negative_ratings_negative_sampler",
+        "negative_sampler_type_graph": "random",
+        "positive_domain": "positive",
+        "offset": 4
+      }
+    }
+  },
+  "dataloader": {
+    "train": {
+      "type": "torch",
+      "batch_size": 256,
+      "batch_processor": {
+        "type": "negative_batch"
+      },
+      "drop_last": true,
+      "shuffle": true
+    },
+    "validation": {
+      "type": "torch",
+      "batch_size": 256,
+      "batch_processor": {
+        "type": "negative_batch"
+      },
+      "drop_last": false,
+      "shuffle": false
+    }
+  },
+  "model": {
+    "type": "siren",
+    "user_prefix": "user.graph",
+    "positive_prefix": "positive.graph",
+    "negative_prefix": "negative.graph",
+    "candidate_prefix": "candidates.graph",
+    "mlp_layers": 2,
+    "embedding_dim": 64,
+    "num_layers": 3,
+    "dropout": 0.5,
+    "initializer_range": 0.02
+  },
+  "optimizer": {
+    "type": "basic",
+    "optimizer": {
+      "type": "adam",
+      "lr": 0.001
+    }
+  },
+  "loss": {
+    "type": "composite",
+    "losses": [
+      {
+        "type": "bpr",
+        "positive_prefix": "positive_scores",
+        "negative_prefix": "negative_scores",
+        "output_prefix": "lightgcn_downstream_loss"
+      },
+      {
+        "type": "regularization",
+        "prefix": [
+          "item_embeddings"
+        ],
+        "weight": 1e-2,
+        "output_prefix": "lightgcn_regularization_loss"
+      }
+    ],
+    "output_prefix": "lightgcn_loss"
+  },
+  "callback": {
+    "type": "composite",
+    "callbacks": [
+      {
+        "type": "metric",
+        "on_step": 1,
+        "loss_prefix": "lightgcn_loss"
+      },
+      {
+        "type": "metric",
+        "on_step": 1,
+        "loss_prefix": "lightgcn_downstream_loss"
+      },
+      {
+        "type": "metric",
+        "on_step": 1,
+        "loss_prefix": "lightgcn_regularization_loss"
+      },
+      {
+        "type": "validation",
+        "on_step": 64,
+        "pred_prefix": "logits",
+        "labels_prefix": "labels.graph",
+        "metrics": {
+          "ndcg@5": {
+            "type": "ndcg",
+            "k": 5
+          },
+          "ndcg@10": {
+            "type": "ndcg",
+            "k": 10
+          },
+          "ndcg@20": {
+            "type": "ndcg",
+            "k": 20
+          },
+          "recall@5": {
+            "type": "recall",
+            "k": 5
+          },
+          "recall@10": {
+            "type": "recall",
+            "k": 10
+          },
+          "recall@20": {
+            "type": "recall",
+            "k": 20
+          }
+        }
+      },
+      {
+        "type": "eval",
+        "on_step": 256,
+        "pred_prefix": "logits",
+        "labels_prefix": "labels.graph",
+        "metrics": {
+          "ndcg@5": {
+            "type": "ndcg",
+            "k": 5
+          },
+          "ndcg@10": {
+            "type": "ndcg",
+            "k": 10
+          },
+          "ndcg@20": {
+            "type": "ndcg",
+            "k": 20
+          },
+          "recall@5": {
+            "type": "recall",
+            "k": 5
+          },
+          "recall@10": {
+            "type": "recall",
+            "k": 10
+          },
+          "recall@20": {
+            "type": "recall",
+            "k": 20
+          }
+        }
+      }
+    ]
+  }
+}

modeling/callbacks/__init__.py

@@ -1 +1 @@
-from .base import BaseCallback, CompositeCallback, EvalCallback
+from .base import BaseCallback, CompositeCallback, EvalCallback, ValidationCallback

modeling/dataloader/batch_processors.py

@@ -35,3 +35,32 @@ def __call__(self, batch):
             processed_batch[part] = torch.tensor(values, dtype=torch.long)
 
         return processed_batch
+
+
+class NegativeBatchProcessor(BaseBatchProcessor, config_name='negative_batch'):
+
+    def __call__(self, batch):
+        processed_batch = {}
+
+        for key in batch[0].keys():
+            if key.endswith('.ids'):
+                # prefix = key.split('.')[0]
+                prefix = key[:-4]
+                assert '{}.length'.format(prefix) in batch[0]
+
+                processed_batch[f'{prefix}.ids'] = []
+                processed_batch[f'{prefix}.length'] = []
+
+                for sample in batch:
+                    # Тк item.negative_domain.ids может не быть из-за отсутствия негативных взаимодействий пользователя
+                    if f'{prefix}.ids' in sample:
+                        processed_batch[f'{prefix}.ids'].extend(sample[f'{prefix}.ids'])
+                        processed_batch[f'{prefix}.length'].append(sample[f'{prefix}.length'])
+
+        for part, values in processed_batch.items():
+            if part == 'ratings.ids':
+                processed_batch[part] = torch.tensor(values, dtype=torch.float)
+            else:
+                processed_batch[part] = torch.tensor(values, dtype=torch.long)
+
+        return processed_batch