add support for batched mm

torchsparsegradutils/sparse_matmul.py

@@ -1,7 +1,40 @@
 import torch
+from torchsparsegradutils.utils import sparse_block_diag, sparse_block_diag_split, stack_csr
 
 
 def sparse_mm(A, B):
+    """
+    Performs a matrix multiplication between a sparse matrix A and a dense matrix B,
+    preserving the sparsity of the gradient with respect to A, permitting sparse backpropagation.
+    
+    The sparse matrix A can be in either COO or CSR format, and is expected
+    to be 2-dimensional, with an optional leading batch dimension. The dense matrix B 
+    should also be 2-dimensional, with a matching optional leading batch dimension. 
+    The batch size must be the same for both A and B.
+    
+    Args:
+        A (torch.Tensor): The sparse matrix in COO or CSR format.
+        B (torch.Tensor): The dense matrix.
+        
+    Returns:
+        torch.Tensor: The result of the matrix multiplication.
+    """
+
+    if not isinstance(A, torch.Tensor) or not isinstance(B, torch.Tensor):
+        raise ValueError("Both A and B should be instances of torch.Tensor")
+
+    if A.dim() < 2 or B.dim() < 2:
+        raise ValueError("Both A and B should be at least 2-dimensional tensors")
+
+    if A.dim() != B.dim():
+        raise ValueError("Both A and B should have the same number of dimensions")
+
+    if A.layout not in {torch.sparse_coo, torch.sparse_csr}:
+        raise ValueError("A should be in either COO or CSR format")
+
+    if A.dim() == 3 and A.size(0) != B.size(0):
+        raise ValueError("If A and B have a leading batch dimension, they should have the same batch size")
+
     return SparseMatMul.apply(A, B)
 
 
@@ -18,11 +51,26 @@ class SparseMatMul(torch.autograd.Function):
 
     @staticmethod
     def forward(ctx, A, B):
+        ctx.batch_size = B.size()[0] if B.dim() == 3 else None
+        ctx.A_shape = A.size()  # (b), n, m
+        ctx.B_shape = B.size()  # (b), m, p
+
         grad_flag = A.requires_grad or B.requires_grad
+
         A, B = A.detach(), B.detach()
+
+        if ctx.batch_size is not None:
+            A = sparse_block_diag(*A)
+            B = torch.cat([*B])
+
         x = torch.sparse.mm(A, B)
-        x.requires_grad = grad_flag
+
         ctx.save_for_backward(A, B)
+
+        if ctx.batch_size is not None:
+            x = x.view(ctx.batch_size, ctx.A_shape[-2], ctx.B_shape[-1])
+
+        x.requires_grad = grad_flag
         return x
 
     @staticmethod
@@ -40,7 +88,7 @@ def backward(ctx, grad):
         # We start by getting the i and j indices:
 
         if A.layout == torch.sparse_coo:
-            A_row_idx, A_col_idx = A.indices()
+            A_row_idx, A_col_idx = A._indices()
         elif A.layout == torch.sparse_csr:
             A_col_idx = A.col_indices()
             A_crow_idx = A.crow_indices()
@@ -50,6 +98,9 @@ def backward(ctx, grad):
             )
         else:
             raise ValueError(f"Unsupported layout: {A.layout}")
+
+        if ctx.batch_size is not None:
+            grad = torch.cat([*grad])
 
         grad_select = grad.index_select(0, A_row_idx)  # grad[i, :]
         B_select = B.index_select(0, A_col_idx)  # B[j, :]
@@ -60,10 +111,20 @@ def backward(ctx, grad):
 
         # Create a sparse matrix of the gradient with respect to the nnz of A
         if A.layout == torch.sparse_coo:
-            gradA = torch.sparse_coo_tensor(A.indices(), gradA, A.shape)
+            gradA = torch.sparse_coo_tensor(A._indices(), gradA, A.shape)
         elif A.layout == torch.sparse_csr:
             gradA = torch.sparse_csr_tensor(A_crow_idx, A_col_idx, gradA, A.shape)
 
         # Now compute the dense gradient with respect to B
         gradB = torch.sparse.mm(A.t(), grad)
+
+        if ctx.batch_size is not None:
+            shapes = ctx.A_shape[0]*(ctx.A_shape[-2:],)
+            gradA = sparse_block_diag_split(gradA, *shapes)
+            if A.layout == torch.sparse_coo:
+                gradA = torch.stack([*gradA])
+            else:
+                gradA = stack_csr([*gradA])  # NOTE: torch.stack does not work for csr tensors
+
+            gradB = gradB.view(ctx.B_shape)
         return gradA, gradB