MinGRUPolicy.Config(
        embedding=NoEncoder.Config(observation_size=844),
        proj_in=Linear.Config(
            channels_in=0,
            channels_out=8,
            bias=False,
            depth_index=(),
            device=None,
            dtype=None,
            init_weight=priml.model.init.kaiming_uniform,
            init_bias=torch.nn.init.zeros_,
            linear_fn=<built-in function linear>,
            shard=None
        ),
        block=MinGRUBlock.Config(
            channels_hidden=8,
            proj_gates=Linear.Config(
                channels_in=8,
                channels_out=24,
                bias=False,
                depth_index=(),
                device=None,
                dtype=None,
                init_weight=priml.model.init.kaiming_uniform,
                init_bias=torch.nn.init.zeros_,
                linear_fn=<built-in function linear>,
                shard=None
            ),
            scan=TorchScan.Config(),
            dtype=None
        ),
        num_layers=2,
        channels_hidden=8,
        num_actions=43,
        proj_out=Linear.Config(
            channels_in=8,
            channels_out=44,
            bias=False,
            depth_index=(),
            device=None,
            dtype=None,
            init_weight=priml.model.init.kaiming_uniform,
            init_bias=torch.nn.init.zeros_,
            linear_fn=<built-in function linear>,
            shard=None
        ),
        dtype=torch.bfloat16,
        state_dtype=torch.float32,
        output_dtype=torch.float32,
        injection=None,
        auxiliary=FeasibilityLoss.Config(
            channels_in=8,
            channels_out=43,
            proj_out=Linear.Config(
                channels_in=8,
                channels_out=43,
                bias=False,
                depth_index=(),
                device=None,
                dtype=None,
                init_weight=priml.model.init.kaiming_uniform,
                init_bias=torch.nn.init.zeros_,
                linear_fn=<built-in function linear>,
                shard=None
            ),
            observation_size=843,
            coefficient=0.1
        ),
        proj_feature=Linear.Config(
            channels_in=5,
            channels_out=8,
            bias=False,
            depth_index=(),
            device=None,
            dtype=None,
            init_weight=priml.model.init.kaiming_uniform,
            init_bias=torch.nn.init.zeros_,
            linear_fn=<built-in function linear>,
            shard=None
        )
    )
