Fixed a bug

Author: Haozhe Qi

llava/train/train.py

@@ -1293,10 +1293,9 @@ def _get_item(self, i) -> Dict[str, torch.Tensor]:
                 meta_data = None
 
                 if self.eval_args.learn_neighbor_actions and 'EK100' in video_file:
-                    vid = video_info
-                    
-                    start_timestamp = round(self.list_data_dict[i]['start_timestamp'], 2)
-                    end_timestamp = round(self.list_data_dict[i]['end_timestamp'], 2)
+                    vid = video_info                  
+                    start_timestamp = round(float(self.list_data_dict[i]['start_timestamp']), 2)
+                    end_timestamp = round(float(self.list_data_dict[i]['end_timestamp']), 2)
                     uid = f"{vid}_{start_timestamp}_{end_timestamp}"
                     meta_data = self.train_triple_lookup.get(uid, None)
 

run_llmseval_clariden.sbatch

@@ -0,0 +1,104 @@
+#!/bin/bash
+#SBATCH --job-name multinode
+#SBATCH -A a-a03
+#SBATCH --hint nomultithread    
+#SBATCH --cpus-per-task 288
+#SBATCH --no-requeue
+#SBATCH --nodes 1                   # number of Nodes
+#SBATCH --ntasks-per-node 1         # number of MP tasks. IMPORTANT: torchrun represents just 1 Slurm task
+#SBATCH --gres gpu:4                # Number of GPUs
+#SBATCH --time 23:00:00             # maximum execution time (DD-HH:MM:SS). Mandatory field in MN5
+#SBATCH --output logs/R-%x.%j-lmmseval-dev_7b_4f_llavavideo_test_haozhe.out
+#SBATCH --error logs/R-%x.%j-lmmseval-dev_7b_4f_llavavideo_test_haozhe.err
+
+mkdir -p logs
+
+echo "START TIME: $(date)"
+
+# auto-fail on any errors in this script
+# set -eo pipefail
+
+# logging script's variables/commands for future debug needs
+set -x
+
+######################
+### Set enviroment ###
+######################
+# module purge
+# module load singularity
+
+GPUS_PER_NODE=4
+echo "NODES: $SLURM_NNODES"
+######################
+
+######################
+#### Set network #####
+######################
+MASTER_ADDR=$(scontrol show hostnames $SLURM_JOB_NODELIST | head -n 1)
+MASTER_PORT=6000
+######################
+
+# note that we don't want to interpolate `\$SLURM_PROCID` till `srun` since otherwise all nodes will get
+# 0 and the launcher will hang
+#
+# same goes for `\$(hostname -s|tr -dc '0-9')` - we want it to interpolate at `srun` time
+LAUNCHER="accelerate launch \
+    --num_processes=$GPUS_PER_NODE \
+    --rdzv_backend c10d \
+    --max_restarts 0 \
+    --tee 3 \
+    "
+
+PYTHON_FILE="-m lmms_eval"
+# PYTHON_ARGS=" \
+#     --model llava_onevision \
+#     --model_args pretrained=lmms-lab/llava-onevision-qwen2-0.5b-ov,conv_template=qwen_1_5,model_name=llava_qwen \
+#     --tasks video_dc499 \
+#     --batch_size 1 \
+#     --log_samples_suffix llava_onevision \
+#     --output_path ./logs/ \
+#     --verbosity=DEBUG \
+#     "
+
+PYTHON_ARGS=" \
+    --model llava_vid \
+    --model_args pretrained=lmms-lab/LLaVA-Video-7B-Qwen2,conv_template=qwen_1_5,max_frames_num=64,mm_spatial_pool_mode=average \
+    --tasks ai2d \
+    --batch_size 1 \
+    --log_samples \
+    --log_samples_suffix llava_vid \
+    --output_path ./logs/
+    --verbosity=DEBUG \
+    "
+
+export CMD="$LAUNCHER $PYTHON_FILE $PYTHON_ARGS"
+export HF_HOME=$SCRATCH/huggingface
+export OMP_NUM_THREADS="8"
+export ACCELERATE_CPU_AFFINITY="1"
+export WANDB_API_KEY="65aeda82a75f1eed29c8e9250b175fcc73dca0d7"
+export OPENAI_API_KEY=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
+
+echo $CMD
+
+# srun error handling:
+# --wait=60: wait 60 sec after the first task terminates before terminating all remaining tasks
+SRUN_ARGS=" \
+    -ul \
+    --cpus-per-task $SLURM_CPUS_PER_TASK \
+    --jobid $SLURM_JOB_ID \
+    --wait 60 \
+    --environment=llava-env \
+    --container-workdir=$PWD \
+    "
+# SINGULARITY_CONTAINER=/path/to/singularity/.sif/file
+# SINGULARITY_ARGS=" \
+#     --bind /path/to/bind/folder \
+#     $SINGULARITY_CONTAINER \
+#     "  
+
+# bash -c is needed for the delayed interpolation of env vars to work
+srun $SRUN_ARGS numactl --membind=0-3 bash -c "
+    source /iopsstor/scratch/cscs/hqi/VFM/llava_dependency/llava-venv/bin/activate
+    $CMD"
+
+echo "END TIME: $(date)"