完善了跳连接，在上decode块后增加特征精炼层，未测效果

dist_temporal_train.sh

@@ -11,9 +11,10 @@ shift 2
 
 # Default parameters
 MODEL=${MODEL:-"SwiftFormerTemporal_XS"}
-BATCH_SIZE=${BATCH_SIZE:-256}
+BATCH_SIZE=${BATCH_SIZE:-128}
 EPOCHS=${EPOCHS:-100}
-LR=${LR:-1e-3}
+# LR=${LR:-1e-3}
+LR=${LR:-0.01}
 OUTPUT_DIR=${OUTPUT_DIR:-"./temporal_output"}
 
 echo "Starting distributed training with $NUM_GPUS GPUs"