HiViT¶

class mmpretrain.models.backbones.HiViT(arch='base', img_size=224, patch_size=16, inner_patches=4, in_chans=3, stem_mlp_ratio=3.0, mlp_ratio=4.0, qkv_bias=True, qk_scale=None, drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.0, norm_cfg={'type': 'LN'}, out_indices=[23], ape=True, rpe=False, patch_norm=True, frozen_stages=-1, kernel_size=None, pad_size=None, layer_scale_init_value=0.0, init_cfg=None)[source]¶

HiViT.

A PyTorch implement of: HiViT: A Simple and More Efficient Design of Hierarchical Vision Transformer.

Parameters:

arch (str | dict) –
Swin Transformer architecture. If use string, choose from ‘tiny’, ‘small’, and’base’. If use dict, it should have below keys:
- embed_dims (int): The dimensions of embedding.
- depths (List[int]): The number of blocks in each stage.
- num_heads (int): The number of heads in attention modules of each stage.
'tiny'. (Defaults to) –
img_size (int) – Input image size.
patch_size (int) – Patch size. Defaults to 16.
inner_patches (int) – Inner patch. Defaults to 4.
in_chans (int) – Number of image input channels.
embed_dim (int) – Transformer embedding dimension.
depths (list[int]) – Number of successive HiViT blocks.
num_heads (int) – Number of attention heads.
stem_mlp_ratio (int) – Ratio of MLP hidden dim to embedding dim in the first two stages.
mlp_ratio (int) – Ratio of MLP hidden dim to embedding dim in the last stage.
qkv_bias (bool) – Enable bias for qkv projections if True.
qk_scale (float) – The number of divider after q@k. Default to None.
drop_rate (float) – Probability of an element to be zeroed after the feed forward layer. Defaults to 0.
attn_drop_rate (float) – The drop out rate for attention output weights. Defaults to 0.
drop_path_rate (float) – Stochastic depth rate. Defaults to 0.
norm_cfg (dict) – Config dict for normalization layer. Defaults to dict(type='LN').
ape (bool) – If True, add absolute position embedding to the patch embedding.
rpe (bool) – If True, add relative position embedding to the patch embedding.
patch_norm (bool) – If True, use norm_cfg for normalization layer.
frozen_stages (int) – Stages to be frozen (stop grad and set eval mode). -1 means not freezing any parameters. Defaults to -1.
kernel_size (int) – Kernel size.
pad_size (int) – Pad size.
layer_scale_init_value (float) – Layer-scale init values. Defaults to 0.
init_cfg (dict, optional) – The extra config for initialization. Defaults to None.

get_layer_depth(param_name, prefix='')[source]¶

Get the layer-wise depth of a parameter.

Parameters:

param_name (str) – The name of the parameter.
prefix (str) – The prefix for the parameter. Defaults to an empty string.

Returns:

The layer-wise depth and the num of layers.

Return type:

Tuple[int, int]

Note

The first depth is the stem module (layer_depth=0), and the last depth is the subsequent module (layer_depth=num_layers-1)