latent_motion_tokenizer_trained_on_oxe/config.yaml

_target_: latent_motion_tokenizer.src.models.latent_motion_tokenizer.LatentMotionTokenizer
codebook_dim: 32
commit_loss_w: 1.0
recon_loss_w: 1.0
perceptual_loss_w: 1.0

image_encoder:
  _target_: transformers.ViTMAEModel.from_pretrained
  pretrained_model_name_or_path: "facebook/vit-mae-large"

m_former:
  _target_: latent_motion_tokenizer.src.models.m_former.MFormer
  add_pooling_layer: false
  config:
    _target_: transformers.ViTConfig
    query_num: 8
    input_hidden_size: 1024
    num_patches: 197 # include the [CLS] token
    attention_probs_dropout_prob: 0.0
    hidden_act: "gelu"
    hidden_dropout_prob: 0.0
    hidden_size: 768 # the hidden size of MAE decoder is 512
    initializer_range: 0.02
    intermediate_size: 3072
    layer_norm_eps: 1e-12
    model_type: "vit"
    num_attention_heads: 12
    num_hidden_layers: 4
    qkv_bias: true

vector_quantizer:
  _target_: latent_motion_tokenizer.src.models.vector_quantizer.VectorQuantizer2
  n_e: 128
  e_dim: 32
  beta: 0.25
  remap: null
  sane_index_shape: true

decoder:
  _target_: latent_motion_tokenizer.src.models.latent_motion_decoder.LatentMotionDecoder
  config:
    _target_: transformers.ViTConfig
    query_num: 8
    attention_probs_dropout_prob: 0.0
    hidden_act: "gelu"
    hidden_dropout_prob: 0.0
    hidden_size: 768
    image_size: 224
    initializer_range: 0.02
    intermediate_size: 3072
    layer_norm_eps: 1e-12
    model_type: "vit"
    num_attention_heads: 12
    num_channels: 3
    num_hidden_layers: 12
    patch_size: 16
    qkv_bias: true
    encoder_stride: 16
    num_patches: 196