Tolerate Linear layers without bias parameters in profiler (#369)

Summary: Pull Request resolved: #369 Handle `bias = None` as a registered parameter, as used in `Lienar(..., bias=False)`. Reviewed By: mannatsingh Differential Revision: D19700470 fbshipit-source-id: 286c1d40f1ead3d235ed5414aa72c7c8ff959e16
facebookresearch · Feb 4, 2020 · dce1ef9 · dce1ef9
1 parent a9a0855
commit dce1ef9
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/classy_vision/generic/profiler.py b/classy_vision/generic/profiler.py
@@ -166,7 +166,7 @@ def _layer_flops(layer, x, _):
     # linear layer:
     elif layer_type in ["Linear"]:
         weight_ops = layer.weight.numel()
-        bias_ops = layer.bias.numel()
+        bias_ops = layer.bias.numel() if layer.bias is not None else 0
         return x.size()[0] * (weight_ops + bias_ops)
 
     # 2D/3D batch normalization: