GNNにおける学習時にエラーが起きた
ImportError: 'NeighborSampler' requires either 'pyg-lib' or 'torch-sparse'

code

問題が起きたのは上の記事での学習と評価...の部分

# 学習と評価
def train(model, loader, device, optimizer, epoch):
    model.train()
    for epoch in range(1, epoch):
        total_loss = total_samples = 0
        for batch_data in tqdm(loader):
            optimizer.zero_grad()
            batch_data = batch_data.to(device)
            pred = model(batch_data)
            loss = F.binary_cross_entropy_with_logits(
                pred, batch_data["blog", "tagged", "tag"].edge_label
            )
            loss.backward()
            optimizer.step()
            total_loss += float(loss) * pred.numel()
            total_samples += pred.numel()
        print(f"Epoch: {epoch:04d}, Loss: {total_loss / total_samples:.4f}")

def validation(model, loader, device, optimizer):
    y_preds = []
    y_trues = []
    model.eval()
    for batch_data in tqdm(loader):
        with torch.no_grad():
            batch_data = batch_data.to(device)
            pred = model(batch_data)
            y_preds.append(pred)
            y_trues.append(batch_data["blog", "tagged", "tag"].edge_label)

    y_pred = torch.cat(y_preds, dim=0).cpu().numpy()
    y_true = torch.cat(y_trues, dim=0).cpu().numpy()
    auc = roc_auc_score(y_true, y_pred)
    return auc, y_pred, y_true


# パラメータセット
model = Model(hidden_channels=64)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
model = model.to(device)

Error

ImportError                               Traceback (most recent call last)
Cell In[41], line 43
     40 model = model.to(device)
     42 # 学習・評価
---> 43 train(model, train_loader, device, optimizer, 6)
     44 auc, y_pred, y_true = validation(model, val_loader, device, optimizer)
     46 # 精度確認（ROC-AUC曲線）

Cell In[41], line 6, in train(model, loader, device, optimizer, epoch)
      4 for epoch in range(1, epoch):
      5     total_loss = total_samples = 0
----> 6     for batch_data in tqdm(loader):
      7         optimizer.zero_grad()
      8         batch_data = batch_data.to(device)

File /opt/conda/lib/python3.10/site-packages/tqdm/std.py:1182, in tqdm.__iter__(self)
   1179 time = self._time
   1181 try:
-> 1182     for obj in iterable:
   1183         yield obj
   1184         # Update and possibly print the progressbar.
   1185         # Note: does not call self.update(1) for speed optimisation.

File /opt/conda/lib/python3.10/site-packages/torch/utils/data/dataloader.py:631, in _BaseDataLoaderIter.__next__(self)
    628 if self._sampler_iter is None:
    629     # TODO(https://github.com/pytorch/pytorch/issues/76750)
    630     self._reset()  # type: ignore[call-arg]
--> 631 data = self._next_data()
    632 self._num_yielded += 1
    633 if self._dataset_kind == _DatasetKind.Iterable and \
    634         self._IterableDataset_len_called is not None and \
    635         self._num_yielded > self._IterableDataset_len_called:

File /opt/conda/lib/python3.10/site-packages/torch/utils/data/dataloader.py:675, in _SingleProcessDataLoaderIter._next_data(self)
    673 def _next_data(self):
    674     index = self._next_index()  # may raise StopIteration
--> 675     data = self._dataset_fetcher.fetch(index)  # may raise StopIteration
    676     if self._pin_memory:
    677         data = _utils.pin_memory.pin_memory(data, self._pin_memory_device)

File /opt/conda/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py:54, in _MapDatasetFetcher.fetch(self, possibly_batched_index)
     52 else:
     53     data = self.dataset[possibly_batched_index]
---> 54 return self.collate_fn(data)

File /opt/conda/lib/python3.10/site-packages/torch_geometric/loader/link_loader.py:211, in LinkLoader.collate_fn(self, index)
    208 r"""Samples a subgraph from a batch of input edges."""
    209 input_data: EdgeSamplerInput = self.input_data[index]
--> 211 out = self.link_sampler.sample_from_edges(
    212     input_data, neg_sampling=self.neg_sampling)
    214 if self.filter_per_worker:  # Execute `filter_fn` in the worker process
    215     out = self.filter_fn(out)

File /opt/conda/lib/python3.10/site-packages/torch_geometric/sampler/neighbor_sampler.py:334, in NeighborSampler.sample_from_edges(self, inputs, neg_sampling)
    329 def sample_from_edges(
    330     self,
    331     inputs: EdgeSamplerInput,
    332     neg_sampling: Optional[NegativeSampling] = None,
    333 ) -> Union[SamplerOutput, HeteroSamplerOutput]:
--> 334     out = edge_sample(inputs, self._sample, self.num_nodes, self.disjoint,
    335                       self.node_time, neg_sampling)
    336     if self.subgraph_type == SubgraphType.bidirectional:
    337         out = out.to_bidirectional()

File /opt/conda/lib/python3.10/site-packages/torch_geometric/sampler/neighbor_sampler.py:666, in edge_sample(inputs, sample_fn, num_nodes, disjoint, node_time, neg_sampling)
    661     if edge_label_time is not None:  # Always disjoint.
    662         seed_time_dict = {
    663             input_type[0]: torch.cat([src_time, dst_time], dim=0),
    664         }
--> 666 out = sample_fn(seed_dict, seed_time_dict)
    668 # Enhance `out` by label information ##################################
    669 if disjoint:

File /opt/conda/lib/python3.10/site-packages/torch_geometric/sampler/neighbor_sampler.py:431, in NeighborSampler._sample(self, seed, seed_time, **kwargs)
    428     num_sampled_nodes = num_sampled_edges = None
    430 else:
--> 431     raise ImportError(f"'{self.__class__.__name__}' requires "
    432                       f"either 'pyg-lib' or 'torch-sparse'")
    434 if num_sampled_edges is not None:
    435     num_sampled_edges = remap_keys(
    436         num_sampled_edges,
    437         self.to_edge_type,
    438     )

ImportError: 'NeighborSampler' requires either 'pyg-lib' or 'torch-sparse'

原因と解決策

原因

単純にpytorch-geometricのオプションライブラリ群が入っていない時は入れる

# Install required packages.
import os
os.environ['TORCH'] = torch.__version__

!pip install torch-scatter -f https://data.pyg.org/whl/torch-${TORCH}.html
!pip install torch-sparse -f https://data.pyg.org/whl/torch-${TORCH}.html
!pip install pyg-lib -f https://data.pyg.org/whl/nightly/torch-${TORCH}.html

pythorh-geometric,pyg-libなどを問題なくimportできる場合は，インストールする順番がおかしい

check

import torch_geometric
print(torch_geometric.typing.WITH_PYG_LIB)
import pyg_lib

torch -> pytorch-geometric -> pyg_lib torch_scatter torch_sparse torch_cluster などのオプションライブラリ の順番でインストールすると，参照するpytorch-geometricのversionが異なることでエラーになるらしい

解決策

torch -> pyg_lib torch_scatter torch_sparse torch_cluster などのオプションライブラリ -> pytorch-geometric の順番でインストールする

!pip install torch==2.2.0
# Install required packages.
import os
os.environ['TORCH'] = torch.__version__

!pip install torch-scatter -f https://data.pyg.org/whl/torch-${TORCH}.html
!pip install torch-sparse -f https://data.pyg.org/whl/torch-${TORCH}.html
!pip install pyg-lib -f https://data.pyg.org/whl/nightly/torch-${TORCH}.html
!pip install git+https://github.com/pyg-team/pytorch_geometric.git

参考