def create_multiscale_vision_transformers()

in pytorchvideo/models/vision_transformers.py [0:0]

• 33 lines of code
• 1 McCabe index (conditional complexity)

def create_multiscale_vision_transformers(
    *,
    spatial_size: _size_2_t,
    temporal_size: int,
    cls_embed_on: bool = True,
    sep_pos_embed: bool = True,
    depth: int = 16,
    norm: str = "layernorm",
    # Patch embed config.
    enable_patch_embed: bool = True,
    input_channels: int = 3,
    patch_embed_dim: int = 96,
    conv_patch_embed_kernel: Tuple[int] = (3, 7, 7),
    conv_patch_embed_stride: Tuple[int] = (2, 4, 4),
    conv_patch_embed_padding: Tuple[int] = (1, 3, 3),
    enable_patch_embed_norm: bool = False,
    use_2d_patch: bool = False,
    # Attention block config.
    num_heads: int = 1,
    mlp_ratio: float = 4.0,
    qkv_bias: bool = True,
    dropout_rate_block: float = 0.0,
    droppath_rate_block: float = 0.0,
    pooling_mode: str = "conv",
    pool_first: bool = False,
    embed_dim_mul: Optional[List[List[int]]] = None,
    atten_head_mul: Optional[List[List[int]]] = None,
    pool_q_stride_size: Optional[List[List[int]]] = None,
    pool_kv_stride_size: Optional[List[List[int]]] = None,
    pool_kv_stride_adaptive: Optional[_size_3_t] = None,
    pool_kvq_kernel: Optional[_size_3_t] = None,
    # Head config.
    head: Optional[Callable] = create_vit_basic_head,
    head_dropout_rate: float = 0.5,
    head_activation: Callable = None,
    head_num_classes: int = 400,