# Incorrect gradient with \`relay.transform.gradient\`

**URL:** <https://discuss.tvm.apache.org/t/incorrect-gradient-with-relay-transform-gradient/8105>\
**Category:** Uncategorized\
**Created:** [October 7, 2020, 2:23pm UTC](https://discuss.tvm.apache.org/t/incorrect-gradient-with-relay-transform-gradient/8105 "2020-10-07T14:23:43Z")\
**Posts on this page:** 4\
**Page:** 1

<div class="post-metadata">

**Author:** ![aseyboldt](https://discuss.tvm.apache.org/letter_avatar_proxy/v4/letter/a/ed655f/32.png) [@aseyboldt](https://discuss.tvm.apache.org/u/aseyboldt)\
**Post date:** [October 7, 2020, 2:23pm UTC](https://discuss.tvm.apache.org/t/incorrect-gradient-with-relay-transform-gradient/8105/1 "2020-10-07T14:23:43Z")

</div>

relay seems to return an incorrect gradient if we roundtrip from ANormal to GraphNormal to ANormal forms. The error disappears if I change the order of the good and bad transforms, because the original model seems to be modified inplace. Maybe this here is related: [https://github.com/apache/incubator-tvm/issues/6624](https://github.com/apache/incubator-tvm/issues/6624)

```python
from tvm import relay
import tvm

N = 50
x = relay.var('x', shape=(N,))
z = relay.sum(x)
func = relay.Function([x], z)
mod = tvm.IRModule()
mod['func'] = func
grad = relay.transform.gradient(mod['func'], mod)
mod['main'] = grad

print("### Bad:")
mod_bad = tvm.transform.Sequential([
    relay.transform.ToANormalForm(),
    relay.transform.ToGraphNormalForm(),
    relay.transform.ToANormalForm(),
    relay.transform.PartialEvaluate(),
    relay.transform.DeadCodeElimination(),
    relay.transform.ToGraphNormalForm(),
])(mod)

with tvm.transform.PassContext(opt_level=3):
    mod_bad, params = relay.optimize(mod_bad, 'llvm')

print(mod_bad.astext())

print("### Good:")
mod_good = tvm.transform.Sequential([
    relay.transform.PartialEvaluate(),
    relay.transform.DeadCodeElimination(),
    relay.transform.ToGraphNormalForm(),
])(mod)

with tvm.transform.PassContext(opt_level=3):
    mod_good, params = relay.optimize(mod_good, 'llvm')

print(mod_good.astext())

```

The output is

```auto
### Bad:
#[version = "0.0.5"]
def @main(%x: Tensor[(50), float32]) -> (float32, (Tensor[(50), float32],)) {
  %0 = fn (%p0: Tensor[(50), float32], Primitive=1) -> float32 {
    sum(%p0) /* ty=float32 */
  };
  %1 = %0(%x) /* ty=float32 */;
  %2 = fn (%p01: Tensor[(50), float32], Primitive=1) -> Tensor[(50), float32] {
    zeros_like(%p01) /* ty=Tensor[(50), float32] */
  };
  %3 = %2(%x) /* ty=Tensor[(50), float32] */;
  %4 = (%3,);
  (%1, %4)
}

### Good:
#[version = "0.0.5"]
def @main(%x: Tensor[(50), float32]) -> (float32, (Tensor[(50), float32],)) {
  %0 = fn (%p0: Tensor[(50), float32], Primitive=1) -> float32 {
    sum(%p0) /* ty=float32 */
  };
  %1 = %0(%x) /* ty=float32 */;
  %6 = fn (%p01: Tensor[(50), float32], %p1: float32, Primitive=1) -> Tensor[(50), float32] {
    %2 = zeros_like(%p01) /* ty=Tensor[(50), float32] */;
    %3 = ones_like(%p1) /* ty=float32 */;
    %4 = expand_dims(%3, axis=0) /* ty=Tensor[(1), float32] */;
    %5 = broadcast_to_like(%4, %p01) /* ty=Tensor[(50), float32] */;
    add(%2, %5) /* ty=Tensor[(50), float32] */
  };
  %7 = %6(%x, %1) /* ty=Tensor[(50), float32] */;
  %8 = (%7,);
  (%1, %8)
}

```

I am using a very recent commit from master (21002cd094c34716b1e07a63ed76f53dadd60e23), on arch linux.

---

<div class="post-metadata">

**Author:** ![altanh](https://discuss.tvm.apache.org/user_avatar/discuss.tvm.apache.org/altanh/32/2082_2.png) [@altanh](https://discuss.tvm.apache.org/u/altanh)\
**Post date:** [October 30, 2020, 5:57pm UTC](https://discuss.tvm.apache.org/t/incorrect-gradient-with-relay-transform-gradient/8105/2 "2020-10-30T17:57:48Z")

</div>

Hi, can you try disabling the `DeadCodeElimination` pass in both cases? This may be an issue with references being removed, which is a current bug.

See [https://github.com/apache/incubator-tvm/issues/6803](https://github.com/apache/incubator-tvm/issues/6803) for info.

---

<div class="post-metadata">

**Author:** ![aseyboldt](https://discuss.tvm.apache.org/letter_avatar_proxy/v4/letter/a/ed655f/32.png) [@aseyboldt](https://discuss.tvm.apache.org/u/aseyboldt)\
**Post date:** [November 3, 2020, 12:35pm UTC](https://discuss.tvm.apache.org/t/incorrect-gradient-with-relay-transform-gradient/8105/3 "2020-11-03T12:35:23Z")

</div>

@altanhThat’s not it. The output doesn’t change if I remove the dead code elimination. And I just rebuilt tvm, the issue is still there on the current master. (Had to add a `mod = relay.transform.InferType()(mod)` before the gradient though)

---

<div class="post-metadata">

**Author:** ![altanh](https://discuss.tvm.apache.org/user_avatar/discuss.tvm.apache.org/altanh/32/2082_2.png) [@altanh](https://discuss.tvm.apache.org/u/altanh)\
**Post date:** [November 4, 2020, 9:52pm UTC](https://discuss.tvm.apache.org/t/incorrect-gradient-with-relay-transform-gradient/8105/4 "2020-11-04T21:52:31Z")

</div>

I looked into it some further and it seems that ToGraphNormalForm might also be unsound with references… Can you try disabling that also? (Although I’m guessing you want to keep that pass.) You can try using `gradient` with `mode="first_order"` if you want to keep the ToGNF pass, since first order will not create references. Hope this helps!