1d04211666
* update from scratch configs * update gym pretraining configs - use fewer epochs * update robomimic pretraining configs - use fewer epochs * allow trajectory plotting in eval agent * add simple vit unet * update avoid pretraining configs - use fewer epochs * update furniture pretraining configs - use same amount of epochs as before * add robomimic diffusion unet pretraining configs * update robomimic finetuning configs - higher lr * add vit unet checkpoint urls * update pretraining and finetuning instructions as configs are updated
91 lines
2.2 KiB
YAML
91 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: transport
|
|
obs_dim: 18 # proprioception only
|
|
action_dim: 14
|
|
denoising_steps: 100
|
|
horizon_steps: 8
|
|
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: [6, 96, 96] # not counting img_cond_steps
|
|
state:
|
|
shape: [9]
|
|
action:
|
|
shape: [7]
|
|
|
|
train:
|
|
n_epochs: 2000
|
|
batch_size: 256
|
|
learning_rate: 1e-4
|
|
weight_decay: 1e-6
|
|
lr_scheduler:
|
|
first_cycle_steps: 2000
|
|
warmup_steps: 100
|
|
min_lr: 1e-5
|
|
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:'3 * ${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
|
|
augment: True
|
|
num_img: 2
|
|
spatial_emb: 128
|
|
time_dim: 32
|
|
mlp_dims: [768, 768, 768]
|
|
residual_style: True
|
|
img_cond_steps: ${img_cond_steps}
|
|
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} |