admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-05-23 12:40:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

more changes

Browse Source
This commit is contained in:
Jade Choghari
2025-12-27 20:26:23 +00:00
parent 4434c863b4
commit 7556c7fd70
7 changed files with 3536 additions and 2138 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fast_tokenizer_local/metadata.json
+6 -6
View File
@@ -14,14 +14,14 @@
"delta_dim_list": null, "delta_dim_list": null,
"use_delta_transform": false, "use_delta_transform": false,
"state_key": "observation.state", "state_key": "observation.state",
"normalization_mode": "MEAN_STD", "normalization_mode": "QUANTILES",
"action_horizon": 10, "action_horizon": 10,
"num_training_chunks": 25065, "num_training_chunks": 25065,
"compression_stats": { "compression_stats": {
"compression_ratio": 2.8901734104046244, "compression_ratio": 3.464660463274599,
"mean_token_length": 24.22, "mean_token_length": 20.204,
"p99_token_length": 40.0, "p99_token_length": 36.00999999999999,
"min_token_length": 9.0, "min_token_length": 5.0,
"max_token_length": 46.0 "max_token_length": 38.0
} }
} }
fast_tokenizer_local/processor_config.json
+1 -1
View File
@@ -3,7 +3,7 @@
"auto_map": { "auto_map": {
"AutoProcessor": "processing_action_tokenizer.UniversalActionProcessor" "AutoProcessor": "processing_action_tokenizer.UniversalActionProcessor"
}, },
"min_token": -203, "min_token": -32,
"processor_class": "UniversalActionProcessor", "processor_class": "UniversalActionProcessor",
"scale": 10.0, "scale": 10.0,
"time_horizon": 10, "time_horizon": 10,
fast_tokenizer_local/tokenizer.json
+3507 -2115
View File
File diff suppressed because it is too large Load Diff
src/lerobot/policies/pi05/modeling_pi05.py
+1 -1
View File
@@ -2183,7 +2183,7 @@ class PI05Policy(PreTrainedPolicy):
# Some checkpoints might have this, but current model expects different structure # Some checkpoints might have this, but current model expects different structure
logging.warning(f"Vision embedding key might need handling: {key}") logging.warning(f"Vision embedding key might need handling: {key}")
if key == "model.paligemma_with_expert.paligemma.lm_head.weight": if key == "model.paligemma_with_expert.paligemma.lm_head.weight" or key == "paligemma_with_expert.paligemma.lm_head.weight":
fixed_state_dict["model.paligemma_with_expert.paligemma.model.language_model.embed_tokens.weight"] = value.clone() fixed_state_dict["model.paligemma_with_expert.paligemma.model.language_model.embed_tokens.weight"] = value.clone()
fixed_state_dict[new_key] = value fixed_state_dict[new_key] = value
src/lerobot/policies/pi05/train_fast.sh
+2 -2
View File
@@ -5,8 +5,8 @@ python src/lerobot/policies/pi05/train_fast_tokenizer.py \
--encoded_dims "0:7" \ --encoded_dims "0:7" \
--vocab_size 1024 \ --vocab_size 1024 \
--push_to_hub \ --push_to_hub \
--hub_repo_id jadechoghari/fast-libero-tokenizer-mean-std \ --hub_repo_id jadechoghari/fast-libero-tokenizer-quantiles \
--normalization_mode MEAN_STD \ --normalization_mode QUANTILES \
# python train_fast_tokenizer.py --repo_id my_dataset # python train_fast_tokenizer.py --repo_id my_dataset
src/lerobot/policies/pi05/train_libero.sh
+6 -4
View File
@@ -5,7 +5,7 @@ lerobot-train \
--output_dir=/fsx/jade_choghari/outputs/libero_training_fast_4 \ --output_dir=/fsx/jade_choghari/outputs/libero_training_fast_4 \
--job_name=libero_training_fast \ --job_name=libero_training_fast \
--policy.repo_id=jade_choghari/pi05-fast-libero \ --policy.repo_id=jade_choghari/pi05-fast-libero \
--policy.path=/fsx/jade_choghari/models/libero-pi-fast \ --policy.path=/fsx/jade_choghari/models/pi05-base \
--policy.dtype=bfloat16 \ --policy.dtype=bfloat16 \
--steps=100000 \ --steps=100000 \
--save_freq=20000 \ --save_freq=20000 \
@@ -16,7 +16,9 @@ lerobot-train \
--policy.scheduler_decay_steps=30000 \ --policy.scheduler_decay_steps=30000 \
--policy.scheduler_decay_lr=1e-5 \ --policy.scheduler_decay_lr=1e-5 \
--policy.gradient_checkpointing=true \ --policy.gradient_checkpointing=true \
# --wandb.enable=true \ --rename_map='{
# --wandb.disable_artifact=true \ "observation.images.image1": "observation.images.base_0_rgb",
# --wandb.project=pi05-libero-training \ "observation.images.image2": "observation.images.left_wrist_0_rgb",
}' \
--policy.empty_cameras=1 \
# /fsx/jade_choghari/.cache/huggingface/lerobot/jadechoghari/collect-data # /fsx/jade_choghari/.cache/huggingface/lerobot/jadechoghari/collect-data
src/lerobot/policies/pi05/train_multi.sh
+13 -9
View File
@@ -9,21 +9,25 @@ accelerate launch --mixed_precision=bf16 --multi_gpu --num_processes=8 \
$(which lerobot-train) \ $(which lerobot-train) \
--dataset.repo_id=local \ --dataset.repo_id=local \
--dataset.root=/fsx/jade_choghari/data/libero \ --dataset.root=/fsx/jade_choghari/data/libero \
--output_dir=/fsx/jade_choghari/outputs/libero_training_fast_5 \ --output_dir=/fsx/jade_choghari/outputs/libero_training_fast_mean \
--job_name=libero_training_fast \ --job_name=libero_training_fast \
--policy.repo_id=jade_choghari/pi05-fast-libero-8 \ --policy.repo_id=jade_choghari/pi05-fast-libero \
--policy.path=/fsx/jade_choghari/models/libero-pi-fast \ --policy.path=/fsx/jade_choghari/models/pi05-base \
--policy.dtype=bfloat16 \ --policy.dtype=bfloat16 \
--steps=120000 \ --steps=100000 \
--save_freq=12000 \ --save_freq=20000 \
--batch_size=8 \ --batch_size=4 \
--policy.compile_model=false \
--policy.device=cuda \ --policy.device=cuda \
--policy.fast_only=true \ --policy.fast_only=true \
--policy.scheduler_warmup_steps=4000 \ --policy.scheduler_warmup_steps=4000 \
--policy.scheduler_decay_steps=120000 \ --policy.scheduler_decay_steps=100000 \
--policy.scheduler_decay_lr=1e-5 \ --policy.scheduler_decay_lr=1e-5 \
--policy.gradient_checkpointing=false \ --policy.gradient_checkpointing=true \
--rename_map='{
"observation.images.image1": "observation.images.base_0_rgb",
"observation.images.image2": "observation.images.left_wrist_0_rgb",
}' \
--policy.empty_cameras=1 \
--wandb.enable=true \ --wandb.enable=true \
--wandb.disable_artifact=true \ --wandb.disable_artifact=true \
--wandb.project=pi05-libero-training \ --wandb.project=pi05-libero-training \