small change

Author: Haozhe Qi

.vscode/launch.json

@@ -103,16 +103,19 @@
                 "WANDB_API_KEY": "65aeda82a75f1eed29c8e9250b175fcc73dca0d7",
                 "CUDA_LAUNCH_BLOCKING": "1",
                 "HF_HOME": "/media/data/haozhe/VFM/huggingface",
+                "OPENAI_API_KEY": "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"
             },
             "args": [
                 "--num_processes", "4",
                 "-m", "lmms_eval",
-                "--model", "llava_vid",
-                "--model_args", "pretrained=experiments/dev_LLaVA-Video-7B-Qwen2_4f_test_haozhe,conv_template=qwen_1_5,max_frames_num=64,mm_spatial_pool_mode=average",
-                "--tasks", "videomme",
+                // "--model", "llava_vid",
+                "--model", "llava_onevision",
+                // "--model_args", "pretrained=experiments/dev_LLaVA-Video-7B-Qwen2_4f_test_haozhe,conv_template=qwen_1_5,max_frames_num=64,mm_spatial_pool_mode=average",
+                "--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",
-                "--log_samples_suffix", "llava_vid_retrained",
+                "--log_samples_suffix", "llava_onevision",
                 "--output_path", "./logs/"
             ],
             "console": "integratedTerminal",

run_clariden.sbatch

@@ -75,10 +75,10 @@ PYTHON_ARGS=" \
     --output_dir experiments/dev_7b_4f_llavavideo_test_haozhe \
     --num_train_epochs 1 \
     --per_device_train_batch_size 1 \
-    --per_device_eval_batch_size 1 \
+    --per_device_eval_batch_size 4 \
     --gradient_accumulation_steps 2 \
-    --evaluation_strategy no \
-    --eval_steps 2000 \
+    --evaluation_strategy epoch \
+    --eval_steps 1 \
     --save_strategy steps \
     --save_steps 2000 \
     --learning_rate 1e-5 \
@@ -95,7 +95,11 @@ PYTHON_ARGS=" \
     --torch_compile True \
     --torch_compile_backend inductor \
     --dataloader_drop_last True \
-    --frames_upbound 4 \
+    --frames_upbound 64 \
+    --mm_newline_position grid \
+    --add_time_instruction True \
+    --force_sample True \
+    --mm_spatial_pool_stride 2
     --root /iopsstor/scratch/cscs/hqi/VFM/onevision/llava_video/EK100 \
     --action_predictions /iopsstor/scratch/cscs/hqi/VFM/llava_data/TIM_PREDS/tim_pred_ids_val.json \
     --val_metadata /iopsstor/scratch/cscs/hqi/VFM/EK100/epic-kitchens-100-annotations/EPIC_100_validation.csv \