From 288c2e08b52f215eada60e91d79469bb097e6071 Mon Sep 17 00:00:00 2001
From: Megvii Engine Team <megengine@megvii.com>
Date: Mon, 4 Jan 2021 16:44:44 +0800
Subject: [PATCH] fix(mge/autodiff): fix expand_dims and grad rule fallback

GitOrigin-RevId: 4aae771222aa1e1a0d8bebd589f4e32c59044f4c
---
 imperative/python/src/grad.cpp                |  4 ++-
 imperative/python/src/grad_override.cpp       | 28 +++++++++++++++----
 .../python/test/unit/core/test_autodiff.py    | 22 +++++++++++++++
 3 files changed, 47 insertions(+), 7 deletions(-)

diff --git a/imperative/python/src/grad.cpp b/imperative/python/src/grad.cpp
index ba340ab3..f4ae8cfc 100644
--- a/imperative/python/src/grad.cpp
+++ b/imperative/python/src/grad.cpp
@@ -309,6 +309,8 @@ public:
     auto& emplace(Args&&... args) {
         return get()->backward.emplace<T>(std::forward<Args>(args)...);
     }
+
+    void reset() { grad_fn = nullptr; }
 };
 
 apply_result_t backward_graph_grad_rule(ApplyContext& ctx, GradFnHelper& ret_grad_fn) {
@@ -398,7 +400,7 @@ apply_result_t apply_grad(ApplyContext& ctx) {
                 maker.finalize();
                 return ret;
             } catch (GradRuleFallback&) {
-                grad_fn_holder.emplace<std::monostate>();
+                grad_fn_holder.reset();
             }
         }
         return backward_graph_grad_rule(ctx, grad_fn_holder);
diff --git a/imperative/python/src/grad_override.cpp b/imperative/python/src/grad_override.cpp
index 9691290d..64217d44 100644
--- a/imperative/python/src/grad_override.cpp
+++ b/imperative/python/src/grad_override.cpp
@@ -177,11 +177,27 @@ apply_result_t reduce_grad_rule(ApplyContext& ctx, CustomBackward::Maker& maker)
     throw GradRuleFallback();
 }
 
-template<typename T, typename U>
-apply_result_t axisAddRemove_grad_rule(ApplyContext& ctx, CustomBackward::Maker& maker) {
-    auto&& op = ctx.op->cast_final_safe<T>();
+apply_result_t addAxis_grad_rule(ApplyContext& ctx, CustomBackward::Maker& maker) {
+    auto&& op = ctx.op->cast_final_safe<AddAxis>();
     mgb_assert(ctx.nargs == 1);
-    auto&& grad_op = U::make(op.axis);
+    auto&& grad_op = RemoveAxis::make(op.axis);
+    std::sort(grad_op->axis.begin(), grad_op->axis.end(), std::greater<int32_t>());
+    maker.output_size(1).output_captured(0, false);
+    maker.backward([grad_op_=std::move(grad_op)](BackwardContext&, Tensor*const* grads, size_t ngrads) {
+        mgb_assert(ngrads == 1);
+        Tensor* grad = grads[0];
+        apply_result_t ret(1);
+        ret[0] = python::apply(grad_op_, grad)[0];
+        return ret;
+    });
+    return apply(ctx);
+}
+
+apply_result_t removeAxis_grad_rule(ApplyContext& ctx, CustomBackward::Maker& maker) {
+    auto&& op = ctx.op->cast_final_safe<RemoveAxis>();
+    mgb_assert(ctx.nargs == 1);
+    auto&& grad_op = AddAxis::make(op.axis);
+    std::sort(grad_op->axis.begin(), grad_op->axis.end());
     maker.output_size(1).output_captured(0, false);
     maker.backward([grad_op_=std::move(grad_op)](BackwardContext&, Tensor*const* grads, size_t ngrads) {
         mgb_assert(ngrads == 1);
@@ -201,8 +217,8 @@ struct Init {
         reg.emplace(Subtensor::typeinfo(), subtensor_grad_rule);
         reg.emplace(IndexingMultiAxisVec::typeinfo(), indexingMultiAxisVec_grad_rule);
         reg.emplace(Reduce::typeinfo(), reduce_grad_rule);
-        reg.emplace(AddAxis::typeinfo(), axisAddRemove_grad_rule<AddAxis, RemoveAxis>);
-        reg.emplace(RemoveAxis::typeinfo(), axisAddRemove_grad_rule<RemoveAxis, AddAxis>);
+        reg.emplace(AddAxis::typeinfo(), addAxis_grad_rule);
+        reg.emplace(RemoveAxis::typeinfo(), removeAxis_grad_rule);
     }
 } _;
 
diff --git a/imperative/python/test/unit/core/test_autodiff.py b/imperative/python/test/unit/core/test_autodiff.py
index 9c1efdf7..2e7fc29b 100644
--- a/imperative/python/test/unit/core/test_autodiff.py
+++ b/imperative/python/test/unit/core/test_autodiff.py
@@ -335,3 +335,25 @@ def test_Reduce_mean():
 
     grad(y, F.ones_like(y))
     np.testing.assert_equal(np.ones((3, 3), dtype=np.float32) / 3, x.grad.numpy())
+
+
+def test_addAxis():
+    x_np = np.random.rand(3, 3).astype("float32")
+    x = mge.Tensor(x_np)
+
+    grad = Grad().wrt(x, callback=save_to(x))
+    y = F.expand_dims(x, [2, 3])
+
+    grad(y, F.ones_like(y))
+    np.testing.assert_equal(np.ones((3, 3), dtype=np.float32), x.grad.numpy())
+
+
+def test_removeAxis():
+    x_np = np.random.rand(3, 3, 1, 1).astype("float32")
+    x = mge.Tensor(x_np)
+
+    grad = Grad().wrt(x, callback=save_to(x))
+    y = F.squeeze(x, [2, 3])
+
+    grad(y, F.ones_like(y))
+    np.testing.assert_equal(np.ones((3, 3, 1, 1), dtype=np.float32), x.grad.numpy())
-- 
GitLab