File size: 8,356 Bytes
211ae03 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 | ## OpenPI $\pi_{0.5}$
### Set-up Envrionment
> ℹ️ The following commands assume the variables eg., `PSI_HOME` are loaded by running `cd /path/to/psi0 && source .env`.
Set up seperate environment for $\pi_{0.5}$:
> ℹ️ We manage the $\Psi_0$ environment and all the baselines through `uv` and they all share the same `src/` code. See [Environment Management](../README.md) for more details.
```
uv venv .venv-openpi --python 3.10
source .venv-openpi/bin/activate
VIRTUAL_ENV=.venv-openpi uv pip install -e .
VIRTUAL_ENV=.venv-openpi uv pip install -e src/openpi/openpi-client
VIRTUAL_ENV=.venv-openpi GIT_LFS_SKIP_SMUDGE=1 uv pip install -r baselines/pi05/requirements-openpi.txt
```
Apply the `transformers` library patches:
> See also the official OpenPI [README.md](https://github.com/Physical-Intelligence/openpi?tab=readme-ov-file#setup)
```
cp -r src/openpi/models_pytorch/transformers_replace/* .venv-openpi/lib/python3.10/site-packages/transformers/
```
### Download Pre-Trained Weights
Download pretrained `pi05_droid` model of `torch` variant:
```
hf download USC-PSI-Lab/psi-model \
--local-dir=$PSI_HOME/cache/checkpoints \
--include="openpi/pi05_droid/*" \
--repo-type=model
```
<details>
<summary>[Optional] Expand to see how we produced the `pi05_droid` pytorch checkpoint.</summary>
Download `pi05-droid` checkpoints
```
python src/openpi/shared/download.py
```
Convet from `jax` to `pytorch` checkpoint
```
uv run examples/convert_jax_model_to_pytorch.py \
--checkpoint_dir=/home/user/.cache/openpi/openpi-assets/checkpoints/pi05_droid \
--config_name=pi05_droid \
--output_path $PSI_HOME/cache/checkpoints/openpi/pi05_droid
```
</details>
### Download Psi0 Task Data
Download the task data, for example
```
export task=G1WholebodyBendPick-v1
```
```
hf download songlinwei/hfm simple/$task.zip --local-dir=$PSI_HOME/data --repo-type=dataset
unzip "$PSI_HOME/data/simple/$task.zip" -d "$PSI_HOME/data/simple"
```
Create a new `TrainConfig` for the task in `src/openpi/training/config.py`:
```
vim src/openpi/training/config.py
```
for example
```
# ...
TrainConfig(
name="simple_bend_pick_v1",
project_name="psi",
num_workers=8,
model=pi0_config.Pi0Config(
pi05=True,
action_dim=36,
action_horizon=30,
max_token_len=250,
),
data=LeRobotHFMDataConfig(
repo_id=f"{os.environ['PSI_HOME']}/data/simple/G1WholebodyBendPick-v1",
base_config=DataConfig(prompt_from_task=True),
),
weight_loader=weight_loaders.CheckpointWeightLoader("gs://openpi-assets/checkpoints/pi05_droid/params"),
num_train_steps=40_000,
batch_size=128,
lr_schedule=_optimizer.CosineDecaySchedule(
warmup_steps=1_000,
peak_lr=1e-4,
decay_steps=40_000,
decay_lr=1e-8,
),
pytorch_weight_path=f"{os.environ['PSI_HOME']}/cache/checkpoints/pi05_droid",
policy_metadata={"dataset": "G1WholebodyBendPick-v1"},
checkpoint_base_dir=f".runs/openpi-05"
),
# ...
```
Compute stats for the task using official script (which is slow)
```
python src/openpi/compute_norm_stats.py --config-name simple_bend_pick_v1
```
> Or you can rewrite our precomputed stats to `openpi` format:
> `python src/openpi/rewrite_norm_stats.py --task_path=$PSI_HOME/data/simple/G1WholebodyBendPick-v1`.
> The calculations are slightly different but it's ok.
### Fine-Tune $\pi_{0.5}$
Launch the training script
```
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 torchrun --standalone --nnodes=1 --nproc_per_node=8 src/openpi/train_pytorch.py \
simple_bend_pick_v1 \
--exp_name=simple_bend_pick_v1 \
--save_interval=10000 \
--checkpoint_base_dir=.runs/openpi-05
```
### Eval $\pi_{0.5}$
```
export port=9000
bash scripts/deploy/serve_pi05.sh $task $step $port
```
Open-loop evaluation
```
export PYTORCH_WEIGHT_PATH=/hfm/songlin/we_learn/.cache/checkpoints/pi05_droid
export DATA_HOME=/hfm/songlin/we_learn/.data/real_teleop_g1/lerobot
export task=Pick_bottle_and_turn_and_pour_into_cup
export port=9000
python scripts/train/openpi/eval_openloop.py --port=$port --task=$task
```
<details>
<summary>[Optional] Expand to see more implementation details.</summary>
Below are the gist of how to adpat OpenPI on humaonoid loco-manipulation tasks:
1. change action dim from `32` to `36`
```
# vim src/openpi/models_pytorch/pi0_pytorch.py
self.action_in_proj = nn.Linear(36, action_expert_config.width)
self.action_out_proj = nn.Linear(action_expert_config.width, 36)
```
2. adapt the loading of pretrained `pi05_torch`
```
model_path = os.path.join(config.pytorch_weight_path, "model.safetensors")
# Psi-0: adapt action dim to 36
from safetensors.torch import load_file
state_dict = load_file(model_path)
pad_dim = config.model.action_dim - state_dict["action_in_proj.weight"].shape[1]
if pad_dim > 0:
# eg., torch.Size([1024, 32]) -> torch.Size([1024, 36])
# Replicate the last 4 columns instead of padding with zeros
w = state_dict["action_in_proj.weight"]
to_pad = w[:, -pad_dim:]
state_dict["action_in_proj.weight"] = torch.cat([w, to_pad], dim=1)
b = state_dict["action_out_proj.bias"]
state_dict["action_out_proj.bias"] = torch.cat([b, b[-pad_dim:]], dim=0)
w = state_dict["action_out_proj.weight"]
to_pad = w[-pad_dim:, :]
state_dict["action_out_proj.weight"] = torch.cat([w, to_pad], dim=0)
# https://github.com/Physical-Intelligence/openpi/issues/669
state_dict["paligemma_with_expert.paligemma.model.language_model.embed_tokens.weight"] = \
state_dict["paligemma_with_expert.paligemma.lm_head.weight"]
_model = model.module if isinstance(model, torch.nn.parallel.DistributedDataParallel) else model
missing_keys, unexpected_keys = _model.load_state_dict(state_dict, strict=False)
```
3. create TrainConfig in config.py
```
TrainConfig(
name="Pick_toys_into_box_and_lift_and_turn_and_put_on_the_chair_new_target_yaw",
project_name="hfm",
num_workers=8,
model=pi0_config.Pi0Config(
pi05=True,
action_dim=36,
action_horizon=16,
max_token_len=250,
),
data=LeRobotHFMDataConfig( # FIXME
repo_id= f"{os.environ['DATA_HOME']}/Pick_toys_into_box_and_lift_and_turn_and_put_on_the_chair_new_target_yaw",
base_config=DataConfig(prompt_from_task=True),
),
weight_loader=weight_loaders.CheckpointWeightLoader("gs://openpi-assets/checkpoints/pi05_droid/params"),
num_train_steps=40_000,
batch_size=128,
lr_schedule=_optimizer.CosineDecaySchedule(
warmup_steps=1_000,
peak_lr=1e-4,
decay_steps=40_000,
decay_lr=1e-8,
),
pytorch_weight_path=os.environ["PYTORCH_WEIGHT_PATH"],
policy_metadata={"dataset": "Pick_toys_into_box_and_lift_and_turn_and_put_on_the_chair_new_target_yaw"},
),
```
4. launch training
```
bash scripts/train/openpi/benchmark_pi05_nv_slurm.sh \
Pick_toys_into_box_and_lift_and_turn_and_put_on_the_chair_new_target_yaw
```
5. [Optional] upload model weights
```
#export task=Hold_lunch_bag_with_both_hands_and_squat_to_put_on_the_coffee_table
#export task=Pick_toys_into_box_and_lift_and_turn_and_put_on_the_chair_new_target_yaw
export task=Pull_the_tray_out_of_chips_can_and_throw_the_can_into_trash_bin
export step=40000
hf upload USC-PSI-Lab/psi-model \
.runs/openpi-05/$task/$task/$step/model.safetensors \
benchmarks/openpi-05/$task/$step/model.safetensors \
--repo-type=model
```
6. Serve
Download:
```
export task=Remove_the_cap_turn_on_the_faucet_and_fill_the_bottle_with_water
export step=40000
python scripts/data/download.py \
--repo-id=USC-PSI-Lab/psi-data-models \
--remote-dir=benchmarks/openpi-05/$task/$step \
--repo-type=model \
--local-dir=.runs/openpi-05/$task/$task/$step
```
and Serve:
```
export port=9000
bash scripts/deploy/serve_pi05.sh $task $step $port
```
Open-loop evaluation
```
export PYTORCH_WEIGHT_PATH=/hfm/songlin/we_learn/.cache/checkpoints/pi05_droid
export DATA_HOME=/hfm/songlin/we_learn/.data/real_teleop_g1/lerobot
export task=Pick_bottle_and_turn_and_pour_into_cup
export port=9000
python scripts/train/openpi/eval_openloop.py --port=$port --task=$task
```
</details> |