Replace deprecated v2.ToTensor() in Matrix-Game-2/inference.py with v2.ToImage() + v2.ToDtype(scale=True)

Start at Matrix-Game-2/inference.py:43 and inspect the frame_process pipeline around the deprecated transform. Validate the inference preprocessing with torchvision 0.26.0+cpu and warnings treated as errors; done means the warning is gone and the replacement produces equivalent output.

Summary

Matrix-Game-2/inference.py:43 instantiates torchvision.transforms.v2.ToTensor(), which is a deprecated alias: constructing it emits a UserWarning on every instantiation:

self.frame_process = v2.Compose([
    v2.Resize(size=(352, 640), antialias=True),
    v2.ToTensor(),          # ← emits UserWarning on construction
    v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
])

Warning text (verified on torchvision 0.26.0):

The transform ToTensor() is deprecated and will be removed in a future release. Instead, please use v2.Compose([v2.ToImage(), v2.ToDtype(torch.float32, scale=True)]).

This warning has been present since torchvision 0.15 (when the v2 namespace was introduced) — torchvision/transforms/v2/_deprecated.py has called warnings.warn in ToTensor.__init__ since the first release.

Proposed fix

Replace v2.ToTensor() with the officially recommended equivalent inside the same v2.Compose:

self.frame_process = v2.Compose([
    v2.Resize(size=(352, 640), antialias=True),
    v2.ToImage(),
    v2.ToDtype(torch.float32, scale=True),
    v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
])

Validation (torchvision 0.26.0+cpu)

• v2.ToTensor()(img) and v2.Compose([v2.ToImage(), v2.ToDtype(torch.float32, scale=True)])(img) produce identical output up to float precision: max abs diff = 5.96e-8 (a single last-bit rounding difference on uint8→float32 scaling; the official deprecation message itself states "Output is equivalent up to float precision").
• With warnings.simplefilter("error", UserWarning): the current code raises the deprecation UserWarning; the proposed pipeline constructs and runs cleanly.

Impact

• No deprecation UserWarning during inference; no behavior change to frame preprocessing.

References

• torchvision docs: v2.ToTensor (deprecated) — "[DEPRECATED] Use v2.Compose([v2.ToImage(), v2.ToDtype(torch.float32, scale=True)]) instead."
• torchvision source torchvision/transforms/v2/_deprecated.py — warnings.warn in ToTensor.__init__, present since 0.15.