model:
  args:
    net:
      args:
        active_reward_monitors:
          env_reward: 1
        attention_heads: 4
        attention_mask_style: clipped_causal
        attention_memory_size: 128
        diff_mlp_embedding: false
        hidsize: 128
        img_shape:
        - 144
        - 144
        - 3
        impala_chans:
        - 16
        - 32
        - 32
        impala_kwargs:
          post_pool_groups: 1
        impala_width: 1
        init_norm_kwargs:
          batch_norm: false
          group_norm_groups: 1
        n_recurrence_layers: 2
        only_img_input: true
        pointwise_ratio: 4
        pointwise_use_activation: false
        recurrence_is_residual: true
        recurrence_type: transformer
        timesteps: 64
        use_pointwise_layer: true
        use_pre_lstm_ln: false
    pi_head_opts:
      temperature: 1.0