Implement a proper batch inference for the depth estimator

[integraledelebesgue]

The obvious implementation

    def predict_batch(
        self,
        frames: list[Frame],
        properties: Properties,
    ) -> list[FloatArray2]:
        return [self.predict(frame, properties) for frame in frames]

        input = torch.stack([torch.from_numpy(frame) for frame in frames])
        input = input.to(self.device)
        input = torch.permute(input, (0, 3, 1, 2))
        input = self.to_model(input)

        result = self.model.predict(input)
        depth: FloatArray3 = (
            Resize((properties.height, properties.width))  # type: ignore[no-untyped-call]
            .forward(result.depth)
            .to(torch.float32)
            .mul(1000.0)
            .cpu()
            .numpy()
        )
        del result

        return list(depth)

results in the error:

RuntimeError: NVML_SUCCESS == DriverAPI::get()->nvmlInit_v2_() INTERNAL ASSERT FAILED at "../c10/cuda/CUDACachingAllocator.cpp":963, please report a bug to PyTorch.