Convmixer Versions Save

Implementation of ConvMixer for "Patches Are All You Need? 🤷"

timm-v1.0

2 years ago

These weights have slightly different parameter names and aren't compatible with this codebase.

v1.0

2 years ago

We provide weights for:

  • ConvMixer-1536/20 (k = 9, p = 7)
  • ConvMixer-768/32 (k = 7, p = 7)
    • IMPORTANT: This model used ReLU instead of GELU.
    • Currently, you would need to change nn.GELU() to nn.ReLU() in convmixer.py to use these weights; we will fix this later.
  • ConvMixer-1024/20 (k = 9, p = 14)