wkcn commented on a change in pull request #18492:
URL: https://github.com/apache/incubator-mxnet/pull/18492#discussion_r436195297
##########
File path: src/operator/nn/cudnn/cudnn_batch_norm-inl.h
##########
@@ -228,7 +228,7 @@ class CuDNNBatchNormOp {
&a,
&b,
&a,
- req[cudnnbatchnorm::kGamma] == kWriteTo ? &b: &b_add,
+ req[cudnnbatchnorm::kGamma] == kAddTo ? &b_add : &b,
Review comment:
The gradient of gamma on CPU is wrong when `grad_req` is True, since it
is not assigned to any value if `grad_req = True`.
```python
import mxnet as mx
from mxnet.gluon import nn
N = 1
C = 3
H = W = 2
block = nn.BatchNorm()
block.collect_params().initialize()
block.collect_params().setattr('grad_req', 'add')
for i in range(1):
with mx.autograd.record():
x = mx.nd.arange(N*C*H*W).reshape((N, C, H, W))
y = block(x)
loss = (y * y).sum()
loss.backward()
print(block.gamma.grad())
```
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
