dppo/cfg/robomimic/pretrain/lift/pre_diffusion_mlp_img.yaml
Allen Z. Ren e0842e71dc
v0.5 to main (#10)
* v0.5 (#9)

* update idql configs

* update awr configs

* update dipo configs

* update qsm configs

* update dqm configs

* update project version to 0.5.0
2024-10-07 16:35:13 -04:00

92 lines
2.2 KiB
YAML

defaults:
- _self_
hydra:
run:
dir: ${logdir}
_target_: agent.pretrain.train_diffusion_agent.TrainDiffusionAgent
name: ${env}_pre_diffusion_mlp_img_ta${horizon_steps}_td${denoising_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env}-img/train.npz
seed: 42
device: cuda:0
env: lift
obs_dim: 9 # proprioception only
action_dim: 7
denoising_steps: 100
horizon_steps: 4
cond_steps: 1
img_cond_steps: 1
wandb:
entity: ${oc.env:DPPO_WANDB_ENTITY}
project: robomimic-${env}-pretrain
run: ${now:%H-%M-%S}_${name}
shape_meta:
obs:
rgb:
shape: [3, 96, 96] # not counting img_cond_steps
state:
shape: [9]
action:
shape: [7]
train:
n_epochs: 2500
batch_size: 256
learning_rate: 1e-4
weight_decay: 1e-6
lr_scheduler:
first_cycle_steps: 8000
warmup_steps: 100
min_lr: 1e-5
epoch_start_ema: 20
update_ema_freq: 10
save_model_freq: 500
model:
_target_: model.diffusion.diffusion.DiffusionModel
predict_epsilon: True
denoised_clip_value: 1.0
network:
_target_: model.diffusion.mlp_diffusion.VisionDiffusionMLP
backbone:
_target_: model.common.vit.VitEncoder
obs_shape: ${shape_meta.obs.rgb.shape}
num_channel: ${eval:'${shape_meta.obs.rgb.shape[0]} * ${img_cond_steps}'} # each image patch is history concatenated
cfg:
patch_size: 8
depth: 1
embed_dim: 128
num_heads: 4
embed_style: embed2
embed_norm: 0
img_cond_steps: ${img_cond_steps}
augment: True
spatial_emb: 128
time_dim: 32
mlp_dims: [512, 512, 512]
residual_style: True
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
horizon_steps: ${horizon_steps}
action_dim: ${action_dim}
horizon_steps: ${horizon_steps}
obs_dim: ${obs_dim}
action_dim: ${action_dim}
denoising_steps: ${denoising_steps}
device: ${device}
ema:
decay: 0.995
train_dataset:
_target_: agent.dataset.sequence.StitchedSequenceDataset
use_img: True
dataset_path: ${train_dataset_path}
horizon_steps: ${horizon_steps}
max_n_episodes: 100
cond_steps: ${cond_steps}
img_cond_steps: ${img_cond_steps}
device: ${device}