lerobot/lerobot/configs/policy/sac_maniskill.yaml

117 lines
3.2 KiB
YAML

# @package _global_
# Train with:
#
# python lerobot/scripts/train.py \
# +dataset=lerobot/pusht_keypoints
# env=pusht \
# env.gym.obs_type=environment_state_agent_pos \
seed: 1
# dataset_repo_id: null
dataset_repo_id: "AdilZtn/Maniskill-Pushcube-demonstration-medium"
training:
# Offline training dataloader
num_workers: 4
# batch_size: 256
batch_size: 512
grad_clip_norm: 10.0
lr: 3e-4
storage_device: "cpu"
eval_freq: 2500
log_freq: 10
save_freq: 2000000
online_steps: 1000000
online_rollout_n_episodes: 10
online_rollout_batch_size: 10
online_steps_between_rollouts: 1000
online_sampling_ratio: 1.0
online_env_seed: 10000
online_buffer_capacity: 200000
online_buffer_seed_size: 0
online_step_before_learning: 500
do_online_rollout_async: false
policy_update_freq: 1
# delta_timestamps:
# observation.environment_state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
# observation.state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
# action: "[i / ${fps} for i in range(${policy.horizon})]"
# next.reward: "[i / ${fps} for i in range(${policy.horizon})]"
policy:
name: sac
pretrained_model_path:
# Input / output structure.
n_action_repeats: 1
horizon: 1
n_action_steps: 1
shared_encoder: true
# vision_encoder_name: "helper2424/resnet10"
vision_encoder_name: null
# freeze_vision_encoder: true
freeze_vision_encoder: false
input_shapes:
# # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
observation.state: ["${env.state_dim}"]
observation.image: [3, 64, 64]
output_shapes:
action: [7]
camera_number: 1
# Normalization / Unnormalization
input_normalization_modes: null
# input_normalization_modes:
# observation.state: min_max
input_normalization_params: null
# observation.state:
# min: [-1.9361e+00, -7.7640e-01, -7.7094e-01, -2.9709e+00, -8.5656e-01,
# 1.0764e+00, -1.2680e+00, 0.0000e+00, 0.0000e+00, -9.3448e+00,
# -3.3828e+00, -3.8420e+00, -5.2553e+00, -3.4154e+00, -6.5082e+00,
# -6.0500e+00, -8.7193e+00, -8.2337e+00, -3.4650e-01, -4.9441e-01,
# 8.3516e-03, -3.1114e-01, -9.9700e-01, -2.3471e-01, -2.7137e-01]
# max: [ 0.8644, 1.4306, 1.8520, -0.7578, 0.9508, 3.4901, 1.9381, 0.0400,
# 0.0400, 5.0885, 4.7156, 7.9393, 7.9100, 2.9796, 5.7720, 4.7163,
# 7.8145, 9.7415, 0.2422, 0.4505, 0.6306, 0.2622, 1.0000, 0.5135,
# 0.4001]
output_normalization_modes:
action: min_max
output_normalization_params:
action:
min: [-1.0, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0]
max: [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
output_normalization_shapes:
action: [7]
# Architecture / modeling.
# Neural networks.
image_encoder_hidden_dim: 32
# discount: 0.99
discount: 0.80
temperature_init: 1.0
num_critics: 10 #10
num_subsample_critics: 2
critic_lr: 3e-4
actor_lr: 3e-4
temperature_lr: 3e-4
# critic_target_update_weight: 0.005
critic_target_update_weight: 0.01
utd_ratio: 2 # 10
actor_learner_config:
learner_host: "127.0.0.1"
learner_port: 50051
policy_parameters_push_frequency: 15