dppo/cfg/gym/pretrain/kitchen-complete-v0/pre_diffusion_mlp.yaml
Allen Z. Ren 1d04211666 v0.7 (#26)
* 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
2024-11-20 15:56:23 -05:00

64 lines
1.5 KiB
YAML

defaults:
- _self_
hydra:
run:
dir: ${logdir}
_target_: agent.pretrain.train_diffusion_agent.TrainDiffusionAgent
name: ${env}_pre_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/gym/${env}/train.npz
seed: 42
device: cuda:0
env: kitchen-complete-v0
obs_dim: 60
action_dim: 9
denoising_steps: 20
horizon_steps: 4
cond_steps: 1
wandb:
entity: ${oc.env:DPPO_WANDB_ENTITY}
project: gym-${env}-pretrain
run: ${now:%H-%M-%S}_${name}
train:
n_epochs: 3000
batch_size: 128
learning_rate: 1e-3
weight_decay: 1e-6
lr_scheduler:
first_cycle_steps: 3000
warmup_steps: 1
min_lr: 1e-4
save_model_freq: 500
model:
_target_: model.diffusion.diffusion.DiffusionModel
predict_epsilon: True
denoised_clip_value: 1.0
network:
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
time_dim: 16
mlp_dims: [256, 256, 256]
cond_mlp_dims: [128, 32]
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
dataset_path: ${train_dataset_path}
horizon_steps: ${horizon_steps}
cond_steps: ${cond_steps}
device: ${device}