dodox/dppo
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

v0.7 (#26)

Browse Source 
* 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
This commit is contained in:
Allen Z. Ren
2024-11-20 15:56:23 -05:00
committed by allenzren
parent d2929f65e1
commit 1d04211666
158 changed files with 3350 additions and 410 deletions
Download Patch File Download Diff File Expand all files Collapse all files
cfg/robomimic/pretrain/lift/pre_diffusion_mlp.yaml
+2 -2
View File
@@ -24,12 +24,12 @@ wandb:
run: ${now:%H-%M-%S}_${name}
train:
n_epochs: 8000
n_epochs: 3000
batch_size: 256
learning_rate: 1e-4
weight_decay: 1e-6
lr_scheduler:
first_cycle_steps: 10000
first_cycle_steps: 3000
warmup_steps: 100
min_lr: 1e-5
save_model_freq: 500
cfg/robomimic/pretrain/lift/pre_diffusion_mlp_img.yaml
+3 -3
View File
@@ -34,12 +34,12 @@ shape_meta:
shape: [7]
train:
n_epochs: 2500
n_epochs: 2000
batch_size: 256
learning_rate: 1e-4
weight_decay: 1e-6
lr_scheduler:
first_cycle_steps: 8000
first_cycle_steps: 2000
warmup_steps: 100
min_lr: 1e-5
save_model_freq: 500
@@ -53,7 +53,7 @@ model:
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
num_channel: ${eval:'3 * ${img_cond_steps}'} # each image patch is history concatenated
cfg:
patch_size: 8
depth: 1
cfg/robomimic/pretrain/lift/pre_diffusion_unet.yaml
+2 -2
View File
@@ -24,12 +24,12 @@ wandb:
run: ${now:%H-%M-%S}_${name}
train:
n_epochs: 8000
n_epochs: 3000
batch_size: 256
learning_rate: 1e-4
weight_decay: 1e-6
lr_scheduler:
first_cycle_steps: 10000
first_cycle_steps: 3000
warmup_steps: 100
min_lr: 1e-5
save_model_freq: 500
cfg/robomimic/pretrain/lift/pre_diffusion_unet_img.yaml
+94
View File
@@ -0,0 +1,94 @@
defaults:
- _self_
hydra:
run:
dir: ${logdir}
_target_: agent.pretrain.train_diffusion_agent.TrainDiffusionAgent
name: ${env}_pre_diffusion_unet_img_ta${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: 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.unet.VisionUnet1D
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
img_cond_steps: ${img_cond_steps}
augment: True
spatial_emb: 128
#
diffusion_step_embed_dim: 32
dim: 40
dim_mults: [1, 2]
kernel_size: 5
n_groups: 8
smaller_encoder: False
cond_predict_scale: True
action_dim: ${action_dim}
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
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}
cfg/robomimic/pretrain/lift/pre_gaussian_mlp.yaml
+2 -2
View File
@@ -23,12 +23,12 @@ wandb:
run: ${now:%H-%M-%S}_${name}
train:
n_epochs: 5000
n_epochs: 3000
batch_size: 256
learning_rate: 1e-4
weight_decay: 1e-6
lr_scheduler:
first_cycle_steps: 5000
first_cycle_steps: 3000
warmup_steps: 100
min_lr: 1e-5
save_model_freq: 500
cfg/robomimic/pretrain/lift/pre_gaussian_mlp_img.yaml
+3 -3
View File
@@ -33,12 +33,12 @@ shape_meta:
shape: [7]
train:
n_epochs: 500
n_epochs: 2000
batch_size: 256
learning_rate: 1e-4
weight_decay: 1e-6
lr_scheduler:
first_cycle_steps: 3000
first_cycle_steps: 2000
warmup_steps: 100
min_lr: 1e-5
save_model_freq: 500
@@ -50,7 +50,7 @@ model:
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
num_channel: ${eval:'3 * ${img_cond_steps}'} # each image patch is history concatenated
cfg:
patch_size: 8
depth: 1
cfg/robomimic/pretrain/lift/pre_gaussian_transformer.yaml
+2 -2
View File
@@ -23,12 +23,12 @@ wandb:
run: ${now:%H-%M-%S}_${name}
train:
n_epochs: 5000
n_epochs: 3000
batch_size: 256
learning_rate: 1e-4
weight_decay: 1e-6
lr_scheduler:
first_cycle_steps: 5000
first_cycle_steps: 3000
warmup_steps: 100
min_lr: 1e-5
save_model_freq: 500
cfg/robomimic/pretrain/lift/pre_gmm_mlp.yaml
+2 -2
View File
@@ -24,12 +24,12 @@ wandb:
run: ${now:%H-%M-%S}_${name}
train:
n_epochs: 5000
n_epochs: 3000
batch_size: 256
learning_rate: 1e-4
weight_decay: 1e-6
lr_scheduler:
first_cycle_steps: 5000
first_cycle_steps: 3000
warmup_steps: 100
min_lr: 1e-5
save_model_freq: 500
cfg/robomimic/pretrain/lift/pre_gmm_transformer.yaml
+2 -2
View File
@@ -24,12 +24,12 @@ wandb:
run: ${now:%H-%M-%S}_${name}
train:
n_epochs: 5000
n_epochs: 3000
batch_size: 256
learning_rate: 1e-4
weight_decay: 1e-6
lr_scheduler:
first_cycle_steps: 5000
first_cycle_steps: 3000
warmup_steps: 100
min_lr: 1e-5
save_model_freq: 500