@package global

defaults:

• /benchmark/rearrange/skills: place # fix to point towards habitat-ext
• /habitat_baselines: habitat_baselines_rl_config_base
• /policy/monolithic@habitat_baselines.rl.policy.main_agent
• /habitat/simulator/sim_sensors@habitat_baselines.eval.extra_sim_sensors.third_rgb_sensor: third_rgb_sensor
• self

habitat_baselines: verbose: False trainer_name: "ver" torch_gpu_id: 2 tensorboard_dir: "tb_eval/place" checkpoint_folder: "data/final_checkpoints/place" video_dir: "place_video_dir" video_fps: 30 test_episode_count: -1 eval_ckpt_path_dir: "data/final_checkpoints/place/latest.pth"

26 environments will just barely be below 16gb.

num_environments: 26

18 environments will just barely be below 11gb.

num_environments: 18 num_updates: -1 total_num_steps: 1.0e9 log_interval: 10 num_checkpoints: 20

Force PyTorch to be single threaded as

this improves performance considerably

force_torch_single_threaded: True eval_keys_to_include_in_name: ["reward", "force", "success"]

eval: video_option: ["disk"]

rl: policy: main_agent: action_dist: clamp_std: True std_init: -1.0 use_std_param: True ppo: # ppo params clip_param: 0.2 ppo_epoch: 2 num_mini_batch: 2 value_loss_coef: 0.5 entropy_coef: 0.001 lr: 3e-4 eps: 1e-5 max_grad_norm: 0.2 num_steps: 128 use_gae: True gamma: 0.99 tau: 0.95 use_linear_clip_decay: False use_linear_lr_decay: False reward_window_size: 50

  use_normalized_advantage: False

  hidden_size: 512

  # Use double buffered sampling, typically helps
  # when environment time is similar or larger than
  # policy inference time during rollout generation
  use_double_buffered_sampler: False

ddppo:
  sync_frac: 0.6
  # The PyTorch distributed backend to use
  distrib_backend: NCCL
  # Visual encoder backbone
  pretrained_weights: data/ddppo-models/gibson-2plus-resnet50.pth
  # Initialize with pretrained weights
  pretrained: False
  # Initialize just the visual encoder backbone with pretrained weights
  pretrained_encoder: False
  # Whether the visual encoder backbone will be trained.
  train_encoder: True
  # Whether to reset the critic linear layer
  reset_critic: True

  # Model parameters
  # resnet50_clip_avgpool
  backbone: resnet18
  rnn_type: LSTM
  num_recurrent_layers: 2

habitat: dataset: data_path: data/datasets/replica_cad/rearrange/v2/{split}/rearrange_easy_language_clip.json.gz simulator: habitat_sim_v0: gpu_device_id: 2