As you can see, I am currently facing a problem. I modified the Transformer part of my code using the code you provided, but it shows that the dimensions of my matrix multiplication do not match. I hope you can solve my doubts.There is a dimension mismatch in the multiplication of x1_ssm and x2_ssm with z. At the same time, I am still confused. Doesn't vim say that mamba blocks are of two types, front and back? Why didn't I see this in the code you provided?
def forward_temporal(self, x,F):
B, J, C = x.shape
# Skip connection
skip = x
# Normalization
x = self.norm(x)
# Split x into x1 and x2 with linears
z1 = self.proj_x(x)
x1 = self.proj_z(x)
# forward
x1 = x1.reshape(B,C,J)
x1_rearranged = self.softplus(x1)
forward_conv_output = self.forward_conv1d(x1_rearranged)
forward_conv_output = forward_conv_output.reshape(B,J,C)
x1_ssm = self.forward_ssm(forward_conv_output)
# backward
x2 = x1.reshape(B,C,J)
x2_rearranged = self.softplus(x2)
backward_conv_output = self.backward_conv1d(x2_rearranged)
backward_conv_output = backward_conv_output.reshape(B,J,C)
x2_ssm = self.backward_ssm(backward_conv_output)
# Activation
z = self.activation(z1)
# matmul with z + backward ssm
x2 = x2_ssm * z
# Matmul with z and x1
x1 = x1_ssm * z
# Add both matmuls
x = x1 + x2
# Add skip connection
return x + skip
kyegomez / VisionMamba
Implementation of Vision Mamba from the paper: "Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model" It's 2.8x faster than DeiT and saves 86.8% GPU memory when performing batch inference to extract features on high-res images
385
Contributors get 40% of received funds after fees
How does funding with Polar work?
1
Pay now to fund the work behind this issue.
2
Get updates on progress being made.
3
Maintainer is rewarded once the issue is completed.
FAQ
Backer
You're funding impactful open source efforts
Contributor
You want to contribute to this effort
Maintainer
You want to get funding like this too