vlfm/config/experiments/ver_pointnav.yaml at c97aae2c9c5b1c64a87591a7701c8ae2d99e0d64 · rai-opensource/vlfm · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
# @package _global_

# Copyright (c) 2023 Robotics and AI Institute LLC dba RAI Institute. All rights reserved.

defaults:
  - /tasks: pointnav_depth_hm3d
  - /habitat_baselines: habitat_baselines_rl_config_base
  - /habitat_baselines/rl/policy/obs_transforms:
    - resize
  - _self_

habitat_baselines:
  verbose: False
  trainer_name: "ver"
  torch_gpu_id: 0
  tensorboard_dir: "tb"
  video_dir: "video_dir"
  test_episode_count: -1
  eval_ckpt_path_dir: "data/new_checkpoints"
  num_environments: 80
  checkpoint_folder: "data/new_checkpoints"
  num_updates: -1
  total_num_steps: 2.5e9
  log_interval: 10
  num_checkpoints: 100
  # Force PyTorch to be single threaded as
  # this improves performance considerably
  force_torch_single_threaded: True

  rl:
    policy:
      name: "PointNavResNetPolicy"

    ppo:
      # ppo params
      clip_param: 0.2
      ppo_epoch: 2
      num_mini_batch: 2
      value_loss_coef: 0.5
      entropy_coef: 0.01
      lr: 2.5e-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
      backbone: resnet18
      rnn_type: LSTM
      num_recurrent_layers: 2