MinkowskiEngine Versions Save

• Fix TensorField.sparse() for no duplicate coordinates
• Skip unnecessary spmm if SparseTensor.initialize_coordinates() has no duplicate coordinates
• Model summary utility function added
• TensorField.splat function for splat features to a sparse tensor
• SparseTensor.interpolate function for extracting interpolated features
• coordinate_key property function for SparseTensor and TensorField
• Fix .dense() for GPU tensors. (PR #319)

• Updated README for pytorch 1.8.1 support
• Use custom gpu_storage instead of thrust vector for faster constructors
• pytorch installation instruction updates
• fix transpose kernel map with kernel_size == stride_size
• Update reconstruction and vae examples for v0.5 API
• stack_unet.py example, API updates
• MinkowskiToFeature layer

• spmm average cuda function
• SparseTensor list operators (cat, mean, sum, var)
• MinkowskiStack containers
• Replace all at::cuda::getCurrentCUDASparseHandle with custom getCurrentCUDASparseHandle (issue #308)
• fix coordinate manager kernel map python function
• direct max pool
  • SparseTensorQuantizationMode.MAX_POOL
• TensorField global max pool
  • origin field
  • origin field map
  • MinkowskiGlobalMaxPool CPU/GPU updates for a field input
• SparseTensor.dense() raises a value error when a coordinate is negative rather than subtracting the minimum coordinate from a sparse tensor. (issue #316)
• Added to_sparse() that removes zeros. (issue #317)
  • Previous to_sparse() was renamed to to_sparse_all()
  • MinkowskiToSparseTensor takes an optional remove_zeros boolean argument.
• Fix global max pool with batch size 1
• Use separate memory chunks for in, out map, and kernel indices for gpu_kernel_map for gpu memory misaligned error

• v0.5 documentation updates
• Nonlinear functionals and modules
• Warning when using cuda without ME cuda support
• diagnostics test
• TensorField slice
  • Cache the unique map and inverse map pair in the coordinate manager
  • generate inverse_mapping on the fly
• CoordinateManager
  • field_to_sparse_insert_and_map
  • exists_field_to_sparse
  • get_field_to_sparse_map
  • fix kernel_map with empty coordinate maps
• CoordiateFieldMap
  • quantize_coordinates
• TensorField binary ops fix
• MinkowskiSyncBatchNorm
  • Support tfield
  • convert sync batchnorm updates
• TensorField to sparse with coordinate map key
• Sparse matrix multiplication
  • force contiguous matrix
• Fix AveragePooling cudaErrorMisalignedAddress error for CUDA 10 (#246)

Major speedup with GPU coordinate map/manager backend.