leonnnop
performance advice, find_unused_parameters
hello, thanks for this code,
FYI this setting induce a big performance hit, leading to slow training,
configs/_gmmseg/models/base_model_config.py:1
find_unused_parameters = True
You probably added it because part of your network is not used during training (and will therefore generate an error, which you manage to workaround by using find_unused_parameters = True).
I think I found the unused part of the net.
You inherit BaseDecodeHead
mmseg/models/decode_heads/gmmseg_head.py:23
class GMMSegHead(BaseDecodeHead):
Which define an unused conv2d:
mmseg/models/decode_heads/decode_head.py:104
self.conv_seg = nn.Conv2d(channels, num_classes, kernel_size=1)
You can remote it. (del self.conv_seg in class GMMSegHead ) which would allow you to set find_unused_parameters = False and enjoy a nice speedup during training.
Thank you for bringing up the performance issue and investing your diligent effort in suggesting a potential solution!
I encourage future users to check for more details, and I have pinned this issue at the top. Hope this helps for future users.
Your valuable input is highly appreciated. Thank you once again!
