5.2. PopTorch Geometric

3.4.0

None

None

Fix scatter_reduce when input has float16 type.
Document that GeneralConv is a supported PyTorch Geometric operator. (Support was added in SDK version 3.3.0.)

None

None

Added support for the following PyTorch Geometric operators:
- SortAggregation
- QuantileAggregation
- MedianAggregation
- MFConv
- ClusterGCNConv
- knn_graph
- to_dense_batch
- HypergraphConv
- SplineBasis and SplineWeighting to support PyTorch Geometric SplineConv
Added support for PyTorch Geometric HeteroLinear and normalization layers.
Added batch_size argument to avg_pool_x and max_pool_x functions to avoid reading IPU tensor.
PopTorch now uses ONNX 11 by default.
FixedSizeDataLoader improvements:
- Added support for a FixedSizeOptions object which can be initialised from an existing dataset or data loader.
- Added support for different fixed-size strategies. Currently supports pad-to-max and stream packing strategies.
- Added support for custom samplers.
- Added support for all arguments used by torch_geometric.DataLoader, which this class is a drop-in replacement for.
Added support for FixedSizeNeighborLoader.

See the supported operations chapter in the PopTorch Geometric documentation for full details of the supported operations.

Fixed PowerMeanAggregation layer by changing power layer gradient calculations.
Fixed support for torch_geometric.Batch in PopTorch which required poptorch_geometric import.
Fixed error in node/edge stores of HeteroData object when using fixed-size loader.

None

None

Renamed FixedBatchSampler to StreamPackingSampler.
FixedBatchSampler now uses SequentialSampler as the default sampler. Previously the default was RandomSampler.