metadata/generation_config.yaml

# Reconstructed from the retained mPLUG-2 generation scripts and configs.
# This documents the settings used for the released semantic prompts; it is
# not a guarantee of bit-for-bit regeneration because the original processed
# frame snapshot and filesystem ordering are not redistributed.
caption_model:
  name: mPLUG-2
  upstream_repository: https://github.com/X-PLUG/mPLUG-2
  task: video_captioning
  checkpoint: mPLUG2_MSVD_Caption.pth
  checkpoint_description: MSVD-finetuned video-captioning checkpoint
  visual_encoder: ViT-L-14
  text_encoder: bert-large-uncased

visual_input:
  modality: RGB_frames_only
  frames_per_clip: 16
  sampling: uniformly_spaced_over_available_stored_frame_list
  frame_filename_order_explicitly_sorted: false
  resolution: 224
  normalization:
    mean: [0.48145466, 0.4578275, 0.40821073]
    std: [0.26862954, 0.26130258, 0.27577711]

decoding:
  method: deterministic_beam_search
  beam_size: 5
  minimum_length: 4
  maximum_length: 20
  length_penalty_alpha: 0.6
  random_seed: 42

postprocessing:
  configured_prompt: "a video of"
  note: >-
    A leading "a video of" phrase was removed from decoded output when it was
    present. The semantic_prompt column contains the resulting caption.

retained_generation_entry_points:
  AVE:
    script: video_caption_mplug2_AVE.py
    config: configs_video/VideoCaption_AVE.yaml
  Kinetics-Sounds:
    script: video_caption_mplug2_VGGsound.py
    config: configs_video/VideoCaption_Ksound.yaml
  VGGSound100:
    script: visualencoder_mplug2_VGGSound.py
    config: configs_video/VideoCaption_VGGsound.yaml