108 lines
2.2 KiB
YAML
108 lines
2.2 KiB
YAML
# @package _global_
|
|
|
|
shape_meta:
|
|
# acceptable types: rgb, low_dim
|
|
obs:
|
|
image:
|
|
shape: [3, 96, 96]
|
|
type: rgb
|
|
agent_pos:
|
|
shape: [2]
|
|
type: low_dim
|
|
action:
|
|
shape: [2]
|
|
|
|
seed: 100000
|
|
horizon: 16
|
|
n_obs_steps: 2
|
|
n_action_steps: 8
|
|
dataset_obs_steps: ${n_obs_steps}
|
|
past_action_visible: False
|
|
keypoint_visible_rate: 1.0
|
|
|
|
eval_episodes: 50
|
|
eval_freq: 5000
|
|
save_freq: 5000
|
|
log_freq: 250
|
|
|
|
offline_steps: 200000
|
|
online_steps: 0
|
|
|
|
offline_prioritized_sampler: true
|
|
|
|
policy:
|
|
name: diffusion
|
|
|
|
shape_meta: ${shape_meta}
|
|
|
|
horizon: ${horizon}
|
|
n_obs_steps: ${n_obs_steps}
|
|
n_action_steps: ${n_action_steps}
|
|
num_inference_steps: 100
|
|
# crop_shape: null
|
|
diffusion_step_embed_dim: 128
|
|
down_dims: [512, 1024, 2048]
|
|
kernel_size: 5
|
|
n_groups: 8
|
|
film_scale_modulation: True
|
|
|
|
pretrained_model_path:
|
|
|
|
batch_size: 64
|
|
|
|
per_alpha: 0.6
|
|
per_beta: 0.4
|
|
|
|
balanced_sampling: false
|
|
utd: 1
|
|
offline_steps: ${offline_steps}
|
|
use_ema: true
|
|
lr_scheduler: cosine
|
|
lr_warmup_steps: 500
|
|
grad_clip_norm: 10
|
|
|
|
delta_timestamps:
|
|
observation.image: [-0.1, 0]
|
|
observation.state: [-0.1, 0]
|
|
action: [-0.1, 0, .1, .2, .3, .4, .5, .6, .7, .8, .9, 1.0, 1.1, 1.2, 1.3, 1.4]
|
|
|
|
rgb_encoder:
|
|
backbone_name: resnet18
|
|
pretrained_backbone: false
|
|
use_group_norm: True
|
|
num_keypoints: 32
|
|
relu: true
|
|
norm_mean_std: [0.5, 0.5] # for PushT the original impl normalizes to [-1, 1] (maybe not the case for robomimic envs)
|
|
crop_shape: [84, 84]
|
|
random_crop: True
|
|
|
|
noise_scheduler:
|
|
_target_: diffusers.schedulers.scheduling_ddpm.DDPMScheduler
|
|
num_train_timesteps: 100
|
|
beta_start: 0.0001
|
|
beta_end: 0.02
|
|
beta_schedule: squaredcos_cap_v2
|
|
variance_type: fixed_small # Yilun's paper uses fixed_small_log instead, but easy to cause Nan
|
|
clip_sample: True # required when predict_epsilon=False
|
|
prediction_type: epsilon # or sample
|
|
|
|
rgb_model:
|
|
pretrained: false
|
|
num_keypoints: 32
|
|
relu: true
|
|
|
|
ema:
|
|
_target_: lerobot.common.policies.diffusion.model.ema_model.EMAModel
|
|
update_after_step: 0
|
|
inv_gamma: 1.0
|
|
power: 0.75
|
|
min_value: 0.0
|
|
max_value: 0.9999
|
|
|
|
optimizer:
|
|
_target_: torch.optim.AdamW
|
|
lr: 1.0e-4
|
|
betas: [0.95, 0.999]
|
|
eps: 1.0e-8
|
|
weight_decay: 1.0e-6
|