132 lines
2.7 KiB
YAML
132 lines
2.7 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]
|
|
|
|
horizon: 16
|
|
n_obs_steps: 2
|
|
n_action_steps: 8
|
|
n_latency_steps: 0
|
|
dataset_obs_steps: ${n_obs_steps}
|
|
past_action_visible: False
|
|
keypoint_visible_rate: 1.0
|
|
obs_as_global_cond: True
|
|
|
|
eval_episodes: 1
|
|
eval_freq: 10000
|
|
save_freq: 100000
|
|
log_freq: 250
|
|
|
|
offline_steps: 1344000
|
|
online_steps: 0
|
|
|
|
offline_prioritized_sampler: true
|
|
|
|
policy:
|
|
name: diffusion
|
|
|
|
shape_meta: ${shape_meta}
|
|
|
|
horizon: ${horizon}
|
|
# n_action_steps: ${eval:'${n_action_steps}+${n_latency_steps}'}
|
|
n_obs_steps: ${n_obs_steps}
|
|
num_inference_steps: 100
|
|
obs_as_global_cond: ${obs_as_global_cond}
|
|
# crop_shape: null
|
|
diffusion_step_embed_dim: 256 # before 128
|
|
down_dims: [256, 512, 1024] # before [512, 1024, 2048]
|
|
kernel_size: 5
|
|
n_groups: 8
|
|
cond_predict_scale: 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
|
|
|
|
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
|
|
|
|
obs_encoder:
|
|
# _target_: diffusion_policy.model.vision.multi_image_obs_encoder.MultiImageObsEncoder
|
|
shape_meta: ${shape_meta}
|
|
# resize_shape: null
|
|
# crop_shape: [76, 76]
|
|
# constant center crop
|
|
# random_crop: True
|
|
use_group_norm: True
|
|
share_rgb_model: False
|
|
imagenet_norm: True
|
|
|
|
rgb_model:
|
|
_target_: diffusion_policy.model.vision.model_getter.get_resnet
|
|
name: resnet18
|
|
weights: null
|
|
|
|
ema:
|
|
_target_: diffusion_policy.model.diffusion.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
|
|
|
|
training:
|
|
device: "cuda:0"
|
|
seed: 42
|
|
debug: False
|
|
resume: True
|
|
# optimization
|
|
# lr_scheduler: cosine
|
|
# lr_warmup_steps: 500
|
|
num_epochs: 8000
|
|
# gradient_accumulate_every: 1
|
|
# EMA destroys performance when used with BatchNorm
|
|
# replace BatchNorm with GroupNorm.
|
|
# use_ema: True
|
|
freeze_encoder: False
|
|
# training loop control
|
|
# in epochs
|
|
rollout_every: 50
|
|
checkpoint_every: 50
|
|
val_every: 1
|
|
sample_every: 5
|
|
# steps per epoch
|
|
max_train_steps: null
|
|
max_val_steps: null
|
|
# misc
|
|
tqdm_interval_sec: 1.0
|