Add epoch shuffle option (#67)

graphormer/data/dataset.py

@@ -4,7 +4,7 @@ import ogb
 import numpy as np
 import torch
 from torch.nn import functional as F
-from fairseq.data import FairseqDataset
+from fairseq.data import data_utils, FairseqDataset, BaseWrapperDataset
 
 from .wrapper import MyPygGraphPropPredDataset
 from .collator import collator
@@ -92,3 +92,22 @@ class GraphormerDataset:
         self.dataset_train = self.dataset.train_data
         self.dataset_val = self.dataset.valid_data
         self.dataset_test = self.dataset.test_data
+
+
+class EpochShuffleDataset(BaseWrapperDataset):
+    def __init__(self, dataset, num_samples, seed):
+        super().__init__(dataset)
+        self.num_samples = num_samples
+        self.seed = seed
+        self.set_epoch(1)
+
+    def set_epoch(self, epoch):
+        with data_utils.numpy_seed(self.seed + epoch - 1):
+            self.sort_order = np.random.permutation(self.num_samples)
+
+    def ordered_indices(self):
+        return self.sort_order
+
+    @property
+    def can_reuse_epoch_itr_across_epochs(self):
+        return False

graphormer/tasks/graph_prediction.py

@@ -21,7 +21,12 @@ from fairseq.tasks import FairseqDataclass, FairseqTask, register_task
 
 from graphormer.pretrain import load_pretrained_model
 
-from ..data.dataset import BatchedDataDataset, TargetDataset, GraphormerDataset
+from ..data.dataset import (
+    BatchedDataDataset,
+    TargetDataset,
+    GraphormerDataset,
+    EpochShuffleDataset,
+)
 
 import torch
 from fairseq.optim.amp_optimizer import AMPOptimizer
@@ -109,6 +114,11 @@ class GraphPredictionConfig(FairseqDataclass):
         metadata={"help": "whether to load the output layer of pretrained model"},
     )
 
+    train_epoch_shuffle: bool = field(
+        default=False,
+        metadata={"help": "whether to shuffle the dataset at each epoch"},
+    )
+
 
 @register_task("graph_prediction", dataclass=GraphPredictionConfig)
 class GraphPredictionTask(FairseqTask):
@@ -161,6 +171,11 @@ class GraphPredictionTask(FairseqTask):
             sizes=data_sizes,
         )
 
+        if split == "train" and self.cfg.train_epoch_shuffle:
+            dataset = EpochShuffleDataset(
+                dataset, size=len(dataset), seed=self.cfg.seed
+            )
+
         logger.info("Loaded {0} with #samples: {1}".format(split, len(dataset)))
 
         self.datasets[split] = dataset

graphormer/tasks/is2re.py

@@ -20,27 +20,7 @@ from fairseq.data import (
 )
 from fairseq.tasks import FairseqTask, register_task
 
-
-class EpochShuffleDataset(BaseWrapperDataset):
-    def __init__(self, dataset, num_samples, seed):
-        super().__init__(dataset)
-        self.num_samples = num_samples
-        self.seed = seed
-        self.set_epoch(1)
-
-    def set_epoch(self, epoch):
-        with data_utils.numpy_seed(self.seed + epoch - 1):
-            self.sort_order = np.random.permutation(
-                self.num_samples
-            )  # random ordered_indices will break fairseq bucket by size batch iter, but we just want to reproduce...
-
-    def ordered_indices(self):
-        return self.sort_order
-
-    @property
-    def can_reuse_epoch_itr_across_epochs(self):
-        return False
-
+from ..data.dataset import EpochShuffleDataset
 
 class LMDBDataset:
     def __init__(self, db_path):