bot commited on
Commit
44038e9
·
1 Parent(s): 3f121e9

Auto-detect resume from checkpoint, save_freq 5000

Browse files
Files changed (1) hide show
  1. train_cloud.sh +12 -1
train_cloud.sh CHANGED
@@ -29,10 +29,21 @@ if [ "$NUM_GPUS" -gt 1 ]; then
29
  ACCEL_FLAGS="--multi_gpu --num_processes $NUM_GPUS"
30
  fi
31
 
 
 
 
 
 
 
 
 
 
 
 
32
  python3.12 -m accelerate.commands.launch $ACCEL_FLAGS \
33
  -m lerobot.scripts.lerobot_train \
 
34
  --dataset.repo_id="so100:$DATASET_DIR:/workspace/pi05-so100-diverse/filtered_index.json:/workspace/pi05-so100-diverse/norm_stats.json" \
35
- --policy.path=lerobot/pi05_base \
36
  --policy.train_expert_only=true \
37
  --policy.dtype=bfloat16 \
38
  --policy.gradient_checkpointing=false \
 
29
  ACCEL_FLAGS="--multi_gpu --num_processes $NUM_GPUS"
30
  fi
31
 
32
+ # Check if we can resume from a checkpoint
33
+ RESUME_ARGS=""
34
+ LAST_CKPT="/ephemeral/production_run/checkpoints/last/pretrained_model/train_config.json"
35
+ if [ -f "$LAST_CKPT" ]; then
36
+ echo "Resuming from checkpoint: $LAST_CKPT" | tee -a "$LOG_FILE"
37
+ RESUME_ARGS="--resume=true --config_path=$LAST_CKPT"
38
+ else
39
+ echo "Starting fresh training" | tee -a "$LOG_FILE"
40
+ RESUME_ARGS="--policy.path=lerobot/pi05_base"
41
+ fi
42
+
43
  python3.12 -m accelerate.commands.launch $ACCEL_FLAGS \
44
  -m lerobot.scripts.lerobot_train \
45
+ $RESUME_ARGS \
46
  --dataset.repo_id="so100:$DATASET_DIR:/workspace/pi05-so100-diverse/filtered_index.json:/workspace/pi05-so100-diverse/norm_stats.json" \
 
47
  --policy.train_expert_only=true \
48
  --policy.dtype=bfloat16 \
49
  --policy.gradient_checkpointing=false \