train_online.yaml

defaults:
  # - agent: android_tavi #tavi # Optimizer will be initialized inside the agent
  # - base_policy: human_retargeter # human_retargeter
  - agent: tavi
  - rl_learner: drqv2
  - base_policy: openloop
  - explorer: ou_noise
  - rewarder: sinkhorn_cosine # Type of the reward 
  - task: tea_picking
  - suite: gym

data_path: ${task.data_path}
image_out_dir: ${task.image_out_dir} # resnet - plier picking 
image_model_type: ${task.image_model_type} # TODO: Add this information to the config - or get it from the learner type and etc

vinn_cont_steps: 10
vinn_no_change_in_demo: False # If set to true demo will not change at each time step - it will only change in the beginning

# tactile_out_dir: /home/irmak/Workspace/tactile-learning/franka_allegro/out/2023.10.10/21-43_curved_tactile_byol 
# tactile_model_type: byol

tactile_out_dir: ${task.tactile_out_dir}
tactile_model_type: ${task.tactile_model_type}



reward_representations: ['image']
policy_representations: ['image', 'tactile', 'features']
data_representations: ['image', 'allegro', 'tactile', 'franka'] # features = allegro + kinova, image and tactile is added

seed: 42 
device: cuda:0
frequency: 5

num_train_frames: 10000 # Total training numbers
num_seed_frames: 800 # Frames until training starts 
num_expl_steps: 800 # ${num_seed_frames} # Frames of full exploration
eval_every_frames: 600  # Evaluate in each every 600 frames
num_eval_episodes: 20
evaluate: True
max_steps: 100

buffer_path: null

# FISH and environment params
expert_id: 0
expert_demo_nums: ${task.expert_demo_nums}
episode_frame_matches: 10
expert_frame_matches: 1
camera_num: ${task.camera_num}

# Agent params 
bc_regularize: False
features_repeat: 5 # Number to how many times to repeat the features as the input to the model

# Replay buffer params
replay_buffer_size: 150000
replay_buffer_num_workers: 2
nstep: 3
batch_size: 256

# Recorder
save_eval_video: False
save_train_video: False 
save_train_cost_matrices: True

# Snapshot loading
load_snapshot: True
snapshot_weight: /home/aadhithya/Workspace/tactile-learning/weights/snapshot_tea_picking_thumb_index.pt


log: True
experiment: ${now:%Y.%m.%d}T${now:%H-%M}_${task.name}_${rl_learner.name}_${base_policy.name}

# hydra configuration - should be received separately
hydra:
    run:
        dir: /home/aadhithya/Workspace/tactile-learning/franka_allegro/out/${now:%Y.%m.%d}/${now:%H-%M}_${experiment}