feat: add vision transfer backbones and IMF variants

roboimi/vla/modules/projectors.py

@@ -0,0 +1,17 @@
+from __future__ import annotations
+
+import torch
+from torch import nn
+
+
+class LinearConditionProjector(nn.Module):
+    """Projects per-step visual+state conditioning to the head conditioning width."""
+
+    def __init__(self, input_dim: int, output_dim: int, bias: bool = True):
+        super().__init__()
+        self.input_dim = int(input_dim)
+        self.output_dim = int(output_dim)
+        self.linear = nn.Linear(self.input_dim, self.output_dim, bias=bias)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.linear(x)