fix group norm, add scaffolding for autograd.grad tests (#630)

Author: Samantha Andow

functorch/csrc/BatchRulesNorm.cpp

@@ -349,6 +349,39 @@ std::tuple<Tensor,Tensor,Tensor> native_group_norm_plumbing(
   return std::make_tuple(result0, mean, rstd);
 }
 
+std::tuple<at::Tensor,optional<int64_t>> group_norm_backward_no_weight_bias_batch_rule(
+    const at::Tensor & grad_out, optional<int64_t> grad_out_bdim,
+    const at::Tensor & input, optional<int64_t> input_bdim,
+    const at::Tensor & mean, optional<int64_t> mean_bdim,
+    const at::Tensor & rstd, optional<int64_t> rstd_bdim,
+    int64_t N, int64_t C, int64_t HxW, int64_t group) {
+  auto grad_out_ = moveBatchDimToFront(grad_out, grad_out_bdim);
+  auto input_ = moveBatchDimToFront(input, input_bdim);
+  auto mean_ = moveBatchDimToFront(mean, mean_bdim);
+  auto rstd_ = moveBatchDimToFront(rstd, rstd_bdim);
+
+  const auto bdim_size = get_bdim_size2(grad_out, grad_out_bdim, input, input_bdim);
+  grad_out_ = ensure_has_bdim(grad_out, grad_out_bdim.has_value(), bdim_size);
+  input_ = ensure_has_bdim(input_, input_bdim.has_value(), bdim_size);
+  mean_ = ensure_has_bdim(mean_, mean_bdim.has_value(), bdim_size);
+  rstd_ = ensure_has_bdim(rstd_, rstd_bdim.has_value(), bdim_size);
+
+  grad_out_ = reshape_dim_into(0, 0, grad_out_); // [B0 * N, C, *]
+  input_ = reshape_dim_into(0, 0, input_);       // [B0 * N, C, *]
+  mean_ = reshape_dim_into(0, 0, mean_);         // [B0 * N, G]
+  rstd_ = reshape_dim_into(0, 0, rstd_);         // [B0 * N, G]
+
+  const auto result = native_group_norm_backward(
+      grad_out_.contiguous(),
+      input_.contiguous(),
+      mean_.contiguous(),
+      rstd_.contiguous(),
+      nullopt, N * bdim_size, C, HxW, group, {true, false, false});
+  auto result0 = std::get<0>(result);
+  result0 = reshape_dim_outof(0, bdim_size, result0);
+  return std::make_tuple(result0, 0);
+}
+
 std::tuple<Tensor,Tensor,Tensor> native_group_norm_backward_plumbing(
   const Tensor & grad_out, const Tensor & input, const Tensor & mean,
   const Tensor & rstd, const c10::optional<Tensor> & weight_opt,
@@ -368,9 +401,6 @@ std::tuple<Tensor,Tensor,Tensor> native_group_norm_backward_plumbing(
     return at::native_group_norm_backward(grad_out, input, mean, rstd, weight_opt, N, C, HxW, group, output_mask);
   }
 
-  Tensor grad_out_value;
-  optional<int64_t> grad_out_bdim;
-  std::tie(grad_out_value, grad_out_bdim) = unwrapTensorAtLevel(grad_out, cur_level);
   Tensor input_value;
   optional<int64_t> input_bdim;
   std::tie(input_value, input_bdim) = unwrapTensorAtLevel(input, cur_level);
@@ -410,32 +440,16 @@ std::tuple<Tensor,Tensor,Tensor> native_group_norm_backward_plumbing(
     optional<int64_t> grad_normalized_input_bdim;
     std::tie(grad_normalized_input_value, grad_normalized_input_bdim) =
         unwrapTensorAtLevel(grad_normalized_input, cur_level);
-    auto grad_out_ = moveBatchDimToFront(grad_normalized_input_value, grad_normalized_input_bdim);
-    auto input_ = moveBatchDimToFront(input_value, input_bdim);
-    auto mean_ = moveBatchDimToFront(mean_value, mean_bdim);
-    auto rstd_ = moveBatchDimToFront(rstd_value, rstd_bdim);
-
-    const auto bdim_size = get_bdim_size3(grad_out_, grad_out_bdim, input_, input_bdim, weight, weight_bdim);
-    grad_out_ = ensure_has_bdim(grad_out_, grad_out_bdim.has_value(), bdim_size);
-    input_ = ensure_has_bdim(input_, input_bdim.has_value(), bdim_size);
-    mean_ = ensure_has_bdim(mean_, mean_bdim.has_value(), bdim_size);
-    rstd_ = ensure_has_bdim(rstd_, rstd_bdim.has_value(), bdim_size);
-
-    grad_out_ = reshape_dim_into(0, 0, grad_out_); // [B0 * N, C, *]
-    input_ = reshape_dim_into(0, 0, input_);       // [B0 * N, C, *]
-    mean_ = reshape_dim_into(0, 0, mean_);         // [B0 * N, G]
-    rstd_ = reshape_dim_into(0, 0, rstd_);         // [B0 * N, G]
 
     c10::impl::ExcludeDispatchKeyGuard guard(kBatchedKey);
-    const auto result = native_group_norm_backward(
-        grad_out_,
-        input_,
-        mean_,
-        rstd_,
-        nullopt, N * bdim_size, C, HxW, group, {true, false, false});
-    auto result0 = std::get<0>(result);
-    result0 = reshape_dim_outof(0, bdim_size, result0);
-    grad_input = makeBatched(result0, 0, cur_level);
+    const auto res = group_norm_backward_no_weight_bias_batch_rule(
+        grad_normalized_input_value, grad_normalized_input_bdim,
+        input_value, input_bdim,
+        mean_value, mean_bdim,
+        rstd_value, rstd_bdim,
+        N, C, HxW, group
+    );
+    grad_input = makeBatched(std::get<0>(res), std::get<1>(res), cur_level);
   }
   return std::make_tuple(grad_input, grad_weight, grad_bias);
 }

test/test_ops.py

@@ -7,10 +7,8 @@
 from torch.testing._internal.common_utils import TestCase, run_tests, is_iterable_of_tensors
 import torch
 from torch import Tensor
-import torch.nn.functional as F
 import functools
 import unittest
-import itertools
 from contextlib import contextmanager
 from torch.testing._internal.common_device_type import instantiate_device_type_tests
 from torch.testing._internal.common_device_type import ops
@@ -29,7 +27,6 @@
     # tol2,
     opsToleranceOverride,
     check_vmap_fallback,
-    loop,
     IS_FBCODE,
 )
 from torch.utils._pytree import tree_flatten, tree_unflatten, tree_map
@@ -198,6 +195,35 @@ def wrapped(*args):
     return wrapped, tuple(flat_args + flat_cotangents)
 
 
+# returns a new function g(*args, *cotangents)
+# that computes vjps and (*args, cotangents) using torch.autograd.grad
+def get_autograd_fn_and_args_with_cotangents(f, sample, cotangents):
+    args = tuple([sample.input] + list(sample.args))
+    kwargs = sample.kwargs
+    flat_args, args_spec = tree_flatten(args)
+    flat_cotangents, cotangents_spec = tree_flatten(cotangents)
+
+    @functools.wraps(f)
+    def wrapped(*args):
+        assert len(args) == len(flat_args) + len(flat_cotangents)
+        actual_args = args[:len(flat_args)]
+        cotangents = args[len(flat_args):]
+        actual_args = tree_unflatten(actual_args, args_spec)
+        cotangents = tree_unflatten(cotangents, cotangents_spec)
+
+        fn, primals = normalize_op_input_output3(f, actual_args, kwargs,
+                                                 flat_args,
+                                                 sample.output_process_fn_grad)
+        out = fn(*primals)
+        diff_wrt = tuple(primal for primal in primals if (primal.requires_grad or primal.grad_fn is not None))
+        if diff_wrt:
+            return torch.autograd.grad(out, diff_wrt, grad_outputs=cotangents)
+        else:
+            return (torch.ones(()),)  # uuugh hack...this will need to be more generic
+
+    return wrapped, tuple(flat_args + flat_cotangents)
+
+
 # Returns a new function g(*args, *cotangents) that computes vjps and
 # sample (*args, *cotangents)
 def get_vjpfull_variant(f, sample):
@@ -1433,28 +1459,22 @@ def test_decompositions_torchscriptable(self, device):
                 continue
             torch.jit.script(decomposition)
 
-    def test_group_norm_backward(self, device):
-        # group norm will hit the decomposable ``infinitely_differentiable_group_norm_backward`` when
-        # GradMode is on, which happens by default in the grad transform. This avoids that
-        def f(x, weight, bias, grad_out):
-            output = F.group_norm(x, 6, weight, bias)
-            inputs = []
-            for input in (x, weight, bias):
-                if input.requires_grad:
-                    inputs.append(input)
-            return torch.autograd.grad(outputs=output, inputs=inputs, grad_outputs=grad_out)
-
-        B, N, C, H, W = 2, 3, 24, 5, 7
-        for (input_grad, weight_grad, bias_grad) in itertools.product((True, False), (True, False), (True, False)):
-            if not input_grad and not weight_grad and not bias_grad:
-                continue
-            x = torch.randn(N, C, H, W, device=device, requires_grad=input_grad)
-            weight = torch.randn(C, device=device, requires_grad=weight_grad)
-            bias = torch.randn(C, device=device, requires_grad=bias_grad)
-            grad_out = torch.randn(B, N, C, H, W, device=device)
-            loop_out = loop(f, (None, None, None, 0), 0, 2, x, weight, bias, grad_out)
-            batched_out = vmap(f, (None, None, None, 0), 0)(x, weight, bias, grad_out)
-            self.assertEqual(loop_out, batched_out)
+    @ops(filter(lambda op: op.name == "nn.functional.group_norm", functorch_lagging_op_db + additional_op_db),
+         allowed_dtypes=(torch.float32, torch.double))  # TODO: generalize
+    def test_group_norm_backward(self, device, dtype, op):
+        # hacky, only works since no group norm inputs can be scalars
+        def was_skipped_from_batched_tensors(batched_out, batch_size):
+            return batched_out.shape == (batch_size,) and all(tuple(e == 1 for e in batched_out))
+
+        sample_inputs = op.sample_inputs(device, dtype, requires_grad=True)
+
+        for sample_input in sample_inputs:
+            cotangents = get_sample_cotangents(op, sample_input)
+            f, args = get_autograd_fn_and_args_with_cotangents(op, sample_input, cotangents)
+            for loop_out, batched_out in get_fallback_and_vmap_exhaustive(f, args, {}, opinfo=op):
+                if all(was_skipped_from_batched_tensors(bo, lo.shape[0]) for (bo, lo) in zip(batched_out, loop_out)):
+                    continue  # we weren't able to use the batched tensor in autograd.grad
+                self.assertEqual(loop_out, batched_out)
 
 
 only_for = ("cpu", "cuda")