batch norm forward over reverse coverage with decomposition (#877)

Author: Samantha Andow

functorch/_src/decompositions.py

@@ -47,6 +47,18 @@ def log_sigmoid_forward(self: Tensor) -> Tuple[Tensor, Tensor]:
     return min - torch.log1p(z), buffer
 
 
+def recompute_mean_var(input: Tensor, rstd: Tensor, inner_dim_indices: List[int], keepdim: bool):
+    # for most norm decompositions, it will be the same as the core version except for here.
+    # We recompute the mean and variance so that they track gradients through input
+
+    mean = torch.mean(input, dim=inner_dim_indices, keepdim=keepdim)
+    var = torch.var(input, dim=inner_dim_indices, unbiased=False, keepdim=keepdim)
+    eps = torch.pow(1 / rstd, 2) - var  # this makes me so sad inside
+    eps = eps.detach()
+    rstd = 1 / torch.sqrt(var + eps)
+    return mean, rstd
+
+
 @register_decomposition_for_jvp(aten.native_layer_norm_backward)
 def native_layer_norm_backward(
     grad_out: Tensor,
@@ -80,13 +92,7 @@ def native_layer_norm_backward(
             input.new_zeros(input_shape[axis:]),
         )
 
-    # this is exactly the same as the other decomposition except for here. We recompute the mean and variance
-    # so that they track gradients through input
-    mean_ = torch.mean(input, dim=inner_dim_indices, keepdim=True)
-    var = torch.var(input, dim=inner_dim_indices, unbiased=False, keepdim=True)
-    eps = torch.pow(1 / rstd, 2) - var  # this makes me so sad inside
-    eps = eps.detach()
-    rstd_ = 1 / torch.sqrt(var + eps)
+    mean_, rstd_ = recompute_mean_var(input, rstd, inner_dim_indices, keepdim=True)
 
     x_hat = (input - mean_) * rstd_
     if weight is not None:
@@ -128,3 +134,84 @@ def native_layer_norm_backward(
         d_bias = torch.zeros(())  # should be None but doesn't work with vjp
 
     return (d_input, d_weight, d_bias)
+
+
+def prod(x: List[int]):
+    r = 1
+    for i in x:
+        r *= i
+    return r
+
+
+@register_decomposition(aten.native_batch_norm_backward)  # @register_decomposition_for_jvp after in core
+def native_batch_norm_backward(
+    grad_out: Tensor,
+    input: Tensor,
+    weight: Optional[Tensor],
+    running_mean: Optional[Tensor],
+    running_var: Optional[Tensor],
+    save_mean: Optional[Tensor],
+    save_invstd: Optional[Tensor],
+    train: bool,
+    eps: float,
+    output_mask: List[bool],
+) -> Tuple[Tensor, Optional[Tensor], Optional[Tensor]]:
+    input_shape = input.shape
+    input_rank = input.dim()
+    assert input_rank >= 2, "rank of the input must be at least 2"
+
+    axis = 1
+    num_features = prod(input_shape) / input_shape[axis]
+    mean = save_mean
+    invstd = save_invstd
+    if train:
+        assert save_mean is not None and save_invstd is not None, "when train=True, save_mean and save_invstd are required"
+
+        reduciton_dims = [0] + list(range(2, input.dim()))
+        assert invstd is not None  # for typing
+        mean, invstd = recompute_mean_var(input, invstd, reduciton_dims, keepdim=False)
+    else:
+        assert running_mean is not None and running_var is not None
+        mean = running_mean
+        invstd = torch.rsqrt(running_var + eps)
+
+    broadcast_mask = [1] * input_rank
+    broadcast_mask[axis] = input_shape[axis]
+
+    reduction_axes: List[int] = []
+    for i in range(input_rank):
+        if i != axis:
+            reduction_axes.append(i)
+
+    mean = torch.reshape(mean, broadcast_mask)
+    norm = 1.0 / num_features
+    grad_output_sum = torch.sum(grad_out, reduction_axes)
+    dot_p = torch.sum(grad_out * (input - mean), reduction_axes)
+
+    grad_mean = torch.reshape(grad_output_sum * norm, broadcast_mask)
+    proj_scale = torch.reshape(torch.mul(dot_p * norm, invstd * invstd), broadcast_mask)
+
+    if weight is None:
+        grad_scale = torch.reshape(invstd, broadcast_mask) * 1.0
+    else:
+        grad_scale = torch.reshape(invstd * weight, broadcast_mask)
+
+    if train:
+        proj = (input - mean) * proj_scale
+        grad_input = ((grad_out - proj) - grad_mean) * grad_scale
+    else:
+        grad_input = grad_out * grad_scale
+
+    if output_mask[1]:
+        grad_weight = dot_p * invstd
+    elif weight is not None:
+        grad_weight = torch.zeros_like(weight)  # should be None but doesn't work with vjp
+    else:
+        grad_weight = torch.zeros(())  # should be None but doesn't work with vjp
+
+    if output_mask[2]:
+        grad_bias = grad_output_sum
+    else:
+        grad_bias = torch.zeros_like(grad_output_sum)  # should be None but doesn't work with vjp
+
+    return (grad_input, grad_weight, grad_bias)

functorch/_src/eager_transforms.py

@@ -1491,5 +1491,7 @@ def _register_python_decomposition_vmap(decomp):
 _register_jit_decomposition(torch.ops.aten._softmax_backward_data.default)
 _register_jit_decomposition(torch.ops.aten.log_sigmoid_forward.default)
 _register_jit_decomposition(torch.ops.aten.native_layer_norm_backward.default)
+_register_jit_decomposition(torch.ops.aten.native_batch_norm_backward.default)
+_register_jit_decomposition(torch.ops.aten.cudnn_batch_norm_backward.default)
 _register_python_decomposition_vmap(torch.ops.aten.mse_loss_backward.default)
 _register_python_decomposition_vmap(torch.ops.aten.addr.default)

functorch/csrc/DynamicLayer.cpp

@@ -502,6 +502,8 @@ TORCH_LIBRARY_IMPL(aten, FT_DYNAMIC_LAYER_FRONT_MODE_KEY, m) {
   OP_DECOMPOSE(log_sigmoid);
   JVP_DECOMP(log_sigmoid_forward);
   JVP_DECOMP(native_layer_norm_backward);
+  JVP_DECOMP(native_batch_norm_backward);
+  JVP_DECOMP(cudnn_batch_norm_backward);
 }
 
 

test/test_ops.py

@@ -1149,8 +1149,6 @@ def get_vjp(cotangents, *primals):
         xfail('logdet', ''),
         xfail('nanmean', ''),
         xfail('nansum', ''),
-        xfail('nn.functional.batch_norm', ''),
-        xfail('nn.functional.batch_norm', 'without_cudnn', device_type='cuda'),
         xfail('nn.functional.embedding'),
         xfail('nn.functional.embedding', 'functorch'),
         xfail('nn.functional.embedding_bag', ''),
@@ -1249,7 +1247,8 @@ def reference(primals, cotangents, primals_tangents, cotangents_tangents):
                 'softmax',
                 'log_softmax',
                 'nn.functional.cross_entropy',
-                'nn.functional.layer_norm'
+                'nn.functional.layer_norm',
+                'nn.functional.batch_norm',
             }
             if op.name in FUNCTORCH_HAS_FORMULA_BUT_NOT_PYTORCH:
                 self.assertFalse(op.supports_fwgrad_bwgrad,