openpi

openpi copied to clipboard

Is the precision of fine-tuning with Lora the same as that of full-parameter fine-tuning on Libero

1

I performed Lora fine-tuning according to the parameters provided in the code, but I couldn't get similar results, achieving only 90.6 on libero_spatial.

helloworld13138

When the training reaches 1000 steps, the training program automatically stops and exits

2

I encountered a strange problem where the training program automatically stopped and exited when the training reached 1000 steps, and the terminal did not print any errors. The printed information...

sunmoon2018

how to use a local dataset for LORA fine-tuning training

10

May I ask how to use a local dataset for LORA fine-tuning training? I don't seem to see an interface for local datasets in the config file. Is it only...

escapethetrap

Droid Dataset Finetuning action dimension mismatch

4

Hi, thanks for the great work! I'm trying to fine-tune the Pi0 base model on the Droid dataset. I defined a `TrainConfig` as follows: ``` TrainConfig( name="pi0_droid_finetune", model=pi0.Pi0Config( paligemma_variant="gemma_2b_lora", action_expert_variant="gemma_300m_lora",...

AsDeadAsADodo

HenryChen4

Out of Memory when fine-tune pi0.5 on the libero dataset

3

Hello, I am using two A100 80GB Gpus to finetune the pi0.5 on the Libero dataset. The training script is: XLA_PYTHON_CLIENT_MEM_FRACTION=0.95 uv run scripts/train.py pi05_libero --exp-name=my_experiment --overwrite --fsdp-devices 2 But...

swords123

env: pytorch 2.7.1. cuda 12.8，On a machine with 8 H200 or H100, single-card training to 7 cards training work without problems, but training with 8 cards results in an error....

Bobo-y

CUDA out of memory when using pytorch to finetune in RTX4090

10

I use the pytorch to finetune the PI0_base. the config I set shows below: TrainConfig( name="test_pi0", model=pi0_config.Pi0Config(paligemma_variant="gemma_2b_lora", action_expert_variant="gemma_300m_lora"), data=LeRobotR1ProDataConfig( repo_id="test/apple", base_config=DataConfig(prompt_from_task=True), extra_delta_transform=True, ), weight_loader=weight_loaders.CheckpointWeightLoader("gs://openpi-assets/checkpoints/pi0_base/params"), num_train_steps=20_000, freeze_filter=pi0_config.Pi0Config(paligemma_variant="gemma_2b_lora", action_expert_variant="gemma_300m_lora").get_freeze_filter(), ema_decay=None, num_workers=1, batch_size=1,...

ZijianWu1121

Fine-tuning (LoRA) using a custom LeRobot dataset

1

I currently have the **lerobot_dataset (v2.1)**, and I would like to perform **pi0 LoRA fine-tuning** using this dataset. My dataset is available at **Ksato5044/lerobot_dataset**. Could you please explain the detailed...

satokeeen

I was training on two graphics cards（RTX3090，24G； RTX3070， 8G） and encountered the following error： .... Loading dataset shards: 100%|████████████████████████████████████████████████████████████████████| 26/26 [00:00

sunmoon2018