training the model

examples/train_smolvla.sh

@@ -0,0 +1,58 @@
+#!/bin/bash
+
+
+cd ~/lerobot_pi
+
+source ~/miniconda3/bin/activate
+conda activate lerobot
+
+
+
+export WORK=/home/mustafa_shukor
+# export TRANSFORMERS_CACHE=$WORK/.cache/huggingface/transformers
+# export HF_HOME=$WORK/.cache/huggingface
+# export DATA_DIR=$WORK/.cache/huggingface/datasets
+# export HF_LEROBOT_HOME=$WORK/.cache/huggingface/lerobot
+
+# export HF_DATASETS_OFFLINE=1 
+# export HF_HUB_OFFLINE=1
+
+# export WANDB_CACHE_DIR=/lustre/fsn1/projects/rech/dyf/ugz83ue/wandb
+# export WANDB_MODE=offline
+
+
+## then later
+## wandb sync wandb/offline-run-*
+
+
+
+# V3 So100
+REPO_ID=danaaubakirova/svla_so100_task1_v3
+DATASET_NAME=so100_v3_task_1
+
+
+POLICY=smolvla
+POLICY_NAME=smolvla
+
+
+
+
+OFFLINE_STEPS=200000
+BATCH_SIZE=64
+
+
+TASK_NAME=lerobot_${DATASET_NAME}_${POLICY_NAME}
+
+
+
+
+TRAIN_DIR=$WORK/logs/lerobot/$TASK_NAME
+echo $TRAIN_DIR
+
+
+rm -r $TRAIN_DIR
+CUDA_VISIBLE_DEVICES=2 python lerobot/scripts/train.py \
+     --policy.type=$POLICY  \
+     --dataset.repo_id=$REPO_ID \
+     --output_dir=$TRAIN_DIR 
+

lerobot/common/policies/smolvla/modeling_smolvla.py

@@ -335,7 +335,7 @@ class SmolVLAPolicy(PreTrainedPolicy):
         loss = losses.mean()
         # For backward pass
         loss_dict["loss"] = loss
-        return loss_dict
+        return loss, loss_dict
 
     def prepare_images(self, batch):
         """Apply SmolVLA preprocessing to the images, like resizing to 224x224 and padding to keep aspect ratio, and