Closed DrewdropLife closed 4 years ago
@changxinC Hi. Thanks for raising the issue. Can you please share the code to reproduce the results?
Thank you for your reply!
class Mish(nn.Module):
def __init__(self):
super(Mish,self).__init__()
def forward(self,x):
return x*torch.tanh(F.softplus(x))
def mish(x):
M = Mish()
return M(x)
I used the above code to replace all F.relu_() and leaklyrelu(0.2, inplace = True).
@changxinC can you please post the complete program along with the data loaders and architecture.
I had found the problem, it might be that I used a pretrained backbone network using relu, and then replace the relu with mish, instead of initializing the backbone network after replacing the mish. Now I haven't modified the backbone network, only changed the relu in other places to mish, and the AP is improved by 0.3. I'd like to ask if you have a resnet-50 pretrained model using mish. Thank you very much!
@changxinC Ah, got it. Thanks for the response. Glad to know that the AP improved. I am working on ImageNet training as of now so soon will upload a ResNet-50 pretrained model. Will take some time though. Closing this issue since it seems to be resolved.
Hi, I tried to use mish instead of relu on mask r-cnn, including res-net and three head, but the AP dropped 0.3 in the end. What is the possible problem? Thank you!