towhee.models.multiscale_vision_transformers.mvit

Functions

align_scale
compute_rollout_attention
drop_path_func	Stochastic Depth per sample.
find_most_h_w
resize_last_dim_to_most
round_width

Classes

AttentionPool	A MLP block that contains two linear layers with a normalization layer. the MLP block is used in a transformer model after the attention block. :: Input ↓ Reshape ↓ Pool ↓ Reshape ↓ norm :param thw_shape: the shape of the input tensor (before flattening). :type thw_shape: List[int] :param pool: Pool operation that is applied to the input tensor. If pool is None, return the input tensor. :type pool: Callable :param has_cls_embed: whether the input tensor contains cls token. Pool operation excludes cls token. :type has_cls_embed: bool :param norm: Optional normalization operation applied to tensor after pool. :type norm: Callable.
DropPath	Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
MViT	Multiscale Vision Transformers Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, Christoph Feichtenhofer https://arxiv.org/abs/2104.11227
Mlp	Multi-layer perception module.
MultiScaleAttention
MultiScaleBlock	MultiScaleBlock for mvit
PatchEmbed	PatchEmbed.
Permute
TransformerBasicHead	BasicHead.