test_configs/llama2_test_config.yaml

model:
  arch: mini_gpt4_llama_v2
  freeze_vit: True
  freeze_qformer: True
  max_txt_len: 512
  low_resource: True
  image_size: 224
  end_sym: "</s>"
  llama_model: "Qwen/Qwen2.5-7B-Instruct"
  ckpt: "checkpoints/video_llama_checkpoint_last.pth"
  use_grad_checkpoint: True
  chat_template: True
  lora_r: 64
  lora_alpha: 16
  length: 45
  use_grad_checkpoint_llm: True
  max_context_len: 3072
  architectures: [
    "MiniGPT4_Video"
  ]
  device: "cuda"
  drop_path_rate: 0.1
  img_size: 224
  model_type: "minigpt4_video"
  num_query_token: 32
  prompt: ""
  torch_dtype: "float16"
  transformers_version: "4.37.2"
  vit_precision: "fp16"
  vit_model: "eva_clip_g"
  token_pooling: true
  lora_target_modules: ["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj", "up_proj"]
  lora_dropout: 0.05
  remove_template: false
  prompt_path: ""
  minigpt4_gpu_id: 0
  whisper_gpu_id: 0
  answer_module_gpu_id: 0
  gradient_accumulation_steps: 1
  warmup_steps: 100
  save_steps: 1000
  logging_steps: 50
  eval_steps: 500
  max_new_tokens: 512
  temperature: 0.7
  top_p: 0.9
  do_sample: true
  num_beams: 1

datasets:
  video_chatgpt: #99378 row  - 13224 video
    batch_size: 2
    num_workers: 2
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 224
        mean: [0.48145466, 0.4578275, 0.40821073]
        std: [0.26862954, 0.26130258, 0.27577711]
      eval:
        name: "blip2_image_eval"
        image_size: 224
    text_processor:
      train:
        name: "blip_caption"
        max_words: 512
      eval:
        name: "blip_caption"
        max_words: 512
    sample_ratio: 100
run:
  seed: 42
  amp: true
  distributed: false
  gpu_id: 0
  world_size: 1
  rank: 0
  dataloader_num_workers: 2
  pin_memory: true
  persistent_workers: true
  prefetch_factor: 2
  clip_grad_norm: 1.0
  weight_decay: 0.01
  adam_epsilon: 1e-8
  adam_beta1: 0.9
  adam_beta2: 0.999

inference:
  batch_size: 1
  max_frames: 45
  frame_interval: 2
  subtitle_max_len: 400
  enable_subtitles: true
  whisper_model: "base"
  response_format: "detailed"
  include_timestamps: false