Making ScaleAndShift blocks begin capable of having parameters that are broadcast by construction, e.g. batch norm with scale parameters [1, 1, 1, d].

PiperOrigin-RevId: 456070961

Author: botev

kfac_jax/_src/curvature_blocks.py

@@ -1509,6 +1509,25 @@ def update_curvature_matrix_estimate(
 #
 
 
+def compatible_shapes(ref_shape, target_shape):
+  if len(target_shape) > len(ref_shape):
+    raise ValueError("Target shape should be smaller.")
+  for ref_d, target_d in zip(reversed(ref_shape), reversed(target_shape)):
+    if ref_d != target_d and target_d != 1:
+      raise ValueError(f"{target_shape} is incompatible with {ref_shape}.")
+
+
+def compatible_sum(tensor, target_shape, skip_axes):
+  compatible_shapes(tensor.shape, target_shape)
+  n = tensor.ndim - len(target_shape)
+  axis = [i + n for i, t in enumerate(target_shape)
+          if t == 1 and i + n not in skip_axes]
+  tensor = jnp.sum(tensor, axis=axis, keepdims=True)
+  axis = [i for i in range(tensor.ndim - len(target_shape))
+          if i not in skip_axes]
+  return jnp.sum(tensor, axis=axis)
+
+
 class ScaleAndShiftDiagonal(Diagonal):
   """A diagonal approximation specifically for a scale and shift layers."""
 
@@ -1539,18 +1558,20 @@ def _update_curvature_matrix_estimate(
       assert (state.diagonal_factors[0].raw_value.shape ==
               self.parameters_shapes[0])
       scale_shape = estimation_data["params"][0].shape
-      axis = range(x.ndim)[1:(x.ndim - len(scale_shape))]
-      d_scale = jnp.sum(x * dy, axis=tuple(axis))
-      scale_diag_update = jnp.sum(d_scale * d_scale, axis=0) / batch_size
+      d_scale = compatible_sum(x * dy, scale_shape, skip_axes=[0])
+      scale_diag_update = jnp.sum(
+          d_scale * d_scale,
+          axis=0, keepdims=d_scale.ndim == len(scale_shape)
+      ) / batch_size
       state.diagonal_factors[0].update(scale_diag_update, ema_old, ema_new)
 
     if self.has_shift:
-      assert (state.diagonal_factors[-1].raw_value.shape ==
-              self.parameters_shapes[-1])
       shift_shape = estimation_data["params"][-1].shape
-      axis = range(x.ndim)[1:(x.ndim - len(shift_shape))]
-      d_shift = jnp.sum(dy, axis=tuple(axis))
-      shift_diag_update = jnp.sum(d_shift * d_shift, axis=0) / batch_size
+      d_shift = compatible_sum(dy, shift_shape, skip_axes=[0])
+      shift_diag_update = jnp.sum(
+          d_shift * d_shift,
+          axis=0, keepdims=d_shift.ndim == len(shift_shape)
+      ) / batch_size
       state.diagonal_factors[-1].update(shift_diag_update, ema_old, ema_new)
 
     return state
@@ -1587,16 +1608,14 @@ def update_curvature_matrix_estimate(
     if self._has_scale:
       # Scale tangent
       scale_shape = estimation_data["params"][0].shape
-      axis = range(x.ndim)[1:(x.ndim - len(scale_shape))]
-      d_scale = jnp.sum(x * dy, axis=tuple(axis))
+      d_scale = compatible_sum(x * dy, scale_shape, skip_axes=[0])
       d_scale = d_scale.reshape([batch_size, -1])
       tangents.append(d_scale)
 
     if self._has_shift:
       # Shift tangent
       shift_shape = estimation_data["params"][-1].shape
-      axis = range(x.ndim)[1:(x.ndim - len(shift_shape))]
-      d_shift = jnp.sum(dy, axis=tuple(axis))
+      d_shift = compatible_sum(dy, shift_shape, skip_axes=[0])
       d_shift = d_shift.reshape([batch_size, -1])
       tangents.append(d_shift)