Avoid overwriting y and dy in the backward pass

Author: ducksoup

README.md

@@ -54,7 +54,7 @@ To install PyTorch, please refer to https://github.com/pytorch/pytorch#installat
 
 To install the package containing the iABN layers:
 ```bash
-pip install git+https://github.com/mapillary/[email protected].8
+pip install git+https://github.com/mapillary/[email protected].9
 ```
 Note that some parts of InPlace-ABN have native C++/CUDA implementations, meaning that the command above will need to
 compile them.

inplace_abn/functions.py

@@ -106,6 +106,9 @@ def forward(ctx, x, weight, bias, running_mean, running_var,
     def backward(ctx, dy_act):
         y_act, var, count, weight, bias = ctx.saved_tensors
 
+        # Create clones of y_act and dy_act, as the backend will modify them in-place
+        y_act, dy_act = y_act.clone(), dy_act.clone()
+
         # Call backward_reduce if we need to compute at least one of the gradients
         if any(ctx.needs_input_grad):
             sum_dy_local, sum_xhat_dy_local = _backend.backward_reduce(