mirror of
https://github.com/huggingface/lerobot.git
synced 2026-05-20 02:59:50 +00:00
Pepijn
1d24301b67
After _tool-good (2000 steps, 0.50/0.50/0.20 dropout) the LM head's
distribution at position 0 shifted from EOS to subtask-vocabulary
tokens but emitted bag-of-words ("cube arm and") rather than well-
formed sentences. That's the expected mid-fine-tuning phase: token-
level supervision has landed, sequence-level grammar hasn't.
Two changes for the next retrain:
* STEPS=15000 (from 2000) — chat-pretrained backbones need O(10k+)
steps to walk their pretraining priors down far enough to commit
to the fine-tuned distribution structurally, not just at the
token level. _tool-g2's bag-of-words output proves the model is
on the right path; it just needs more gradient signal.
* plan/memory dropout 0.50 -> 0.30 — 0.50 was probably too
aggressive for a small dataset. Half the training samples had
crucial context missing, which slows down learning the full
conditional structure. 0.30 still regularises against prompt
leakage but lets the model learn proper grammar first; the
higher dropout can be revisited once the head is solid.
Subtask dropout stays at 0.20 since subtask isn't in the high-level
prompt anyway (recipe fix removed the "Current subtask:" message).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
81 lines
3.1 KiB
Bash
81 lines
3.1 KiB
Bash
#!/bin/bash
|
|
#SBATCH --job-name=smolvla2-hirobot
|
|
#SBATCH --partition=hopper-prod
|
|
#SBATCH --qos=high
|
|
#SBATCH --time=48:00:00
|
|
#SBATCH --ntasks=1
|
|
#SBATCH --gpus-per-task=8
|
|
|
|
# SmolVLA2 training on an annotated dataset.
|
|
#
|
|
# The high_level_subtask recipe (recipes/smolvla2_hirobot.yaml) was
|
|
# fixed in PR3 to supervise the LM head with the *current* active
|
|
# subtask span at every frame, not the next-span target which is
|
|
# empty on stable phases. With the old recipe the head learned to
|
|
# emit ``\n`` on every chunk boundary; the new one supervises a
|
|
# real, scene-grounded string at every frame.
|
|
#
|
|
# Two regularisers are still on:
|
|
#
|
|
# * --dataset.image_transforms.enable=true: torchvision-v2
|
|
# ColorJitter + SharpnessJitter + RandomAffine per frame; default
|
|
# envelope (brightness ±20% etc).
|
|
# * --policy.{plan,memory,subtask}_dropout_prob: randomly drop the
|
|
# context messages carrying the named recipe binding so the model
|
|
# handles missing/stale context. Mirrors Pi0.7 §V.E.
|
|
|
|
set -euo pipefail
|
|
|
|
cd "${LEROBOT_ROOT:-$HOME/lerobot}"
|
|
|
|
export PATH="$HOME/miniconda3/bin:$HOME/.local/bin:$PATH"
|
|
export LD_LIBRARY_PATH="$HOME/miniconda3/lib:${LD_LIBRARY_PATH:-}"
|
|
export NCCL_TIMEOUT="${NCCL_TIMEOUT:-1800}"
|
|
export HF_HUB_DOWNLOAD_TIMEOUT="${HF_HUB_DOWNLOAD_TIMEOUT:-120}"
|
|
export WANDB_INIT_TIMEOUT="${WANDB_INIT_TIMEOUT:-300}"
|
|
|
|
DATASET="${DATASET:-pepijn223/super_poulain_full_tool3}"
|
|
POLICY_REPO_ID="${POLICY_REPO_ID:-pepijn223/smolvla2_hirobot_super_poulain_tool6}"
|
|
JOB_NAME="${JOB_NAME:-smolvla2-hirobot-super-poulain-tool6}"
|
|
NUM_PROCESSES="${NUM_PROCESSES:-8}"
|
|
BATCH_SIZE="${BATCH_SIZE:-32}"
|
|
STEPS="${STEPS:-15000}"
|
|
RUN_ID="${SLURM_JOB_ID:-$(date +%Y%m%d_%H%M%S)}"
|
|
OUTPUT_DIR="${OUTPUT_DIR:-/fsx/pepijn/outputs/train/smolvla2_hirobot_super_poulain_tool3_${STEPS}_${RUN_ID}}"
|
|
|
|
echo "Training smolvla2 on $DATASET"
|
|
echo " GPUs: $NUM_PROCESSES"
|
|
echo " batch: $BATCH_SIZE / GPU (global=$((NUM_PROCESSES * BATCH_SIZE)))"
|
|
echo " steps: $STEPS"
|
|
echo " output: $OUTPUT_DIR"
|
|
echo " augmentation: image_transforms ON, prompt dropout {plan:0.30 memory:0.30 subtask:0.20}"
|
|
|
|
accelerate launch --multi_gpu --num_processes="$NUM_PROCESSES" \
|
|
-m lerobot.scripts.lerobot_train \
|
|
--policy.type=smolvla2 \
|
|
--policy.recipe_path=recipes/smolvla2_hirobot.yaml \
|
|
--dataset.repo_id="$DATASET" \
|
|
--dataset.revision=main \
|
|
--dataset.video_backend=pyav \
|
|
--output_dir="$OUTPUT_DIR" \
|
|
--job_name="$JOB_NAME" \
|
|
--policy.repo_id="$POLICY_REPO_ID" \
|
|
--policy.compile_model=false \
|
|
--policy.device=cuda \
|
|
--policy.tokenizer_max_length=512 \
|
|
--steps="$STEPS" \
|
|
--policy.scheduler_decay_steps="$STEPS" \
|
|
--batch_size="$BATCH_SIZE" \
|
|
--wandb.enable=true \
|
|
--wandb.disable_artifact=true \
|
|
--wandb.project=hirobot \
|
|
--log_freq=100 \
|
|
--save_freq="$STEPS" \
|
|
--num_workers=0 \
|
|
--dataset.image_transforms.enable=true \
|
|
--dataset.image_transforms.max_num_transforms=3 \
|
|
--dataset.image_transforms.random_order=true \
|
|
--policy.plan_dropout_prob=0.30 \
|
|
--policy.memory_dropout_prob=0.30 \
|
|
--policy.subtask_dropout_prob=0.20
|