data

PVTConfig

extendsModelConfig

PVTConfig(num_classes: int = 1000, in_channels: int = 3, variant: str = 'pvt_tiny', embed_dims: tuple[int, ...] = (64, 128, 320, 512), depths: tuple[int, ...] = (2, 2, 2, 2), num_heads: tuple[int, ...] = (1, 2, 5, 8), sr_ratios: tuple[int, ...] = (8, 4, 2, 1), mlp_ratios: tuple[float, ...] = (8.0, 8.0, 4.0, 4.0))

source edit

Configuration dataclass for every PVT (v2) variant.

PVTConfig is an immutable container that fully specifies the architecture of a Pyramid Vision Transformer v2 (Wang et al., 2021 and 2022). The implementation in lucid.models.vision.pvt builds PVT v2 by default: overlapping convolutional patch embedding at each stage, a depthwise-conv-augmented MLP, and the Spatial-Reduction Attention (SRA) from PVT v1.

Parameters

num_classesint= 1000

Number of output classes for the classification head. Defaults to 1000 (ImageNet-1k).

in_channelsint= 3

Number of input image channels. Defaults to 3 (RGB).

variantstr= 'pvt_tiny'

Informational label identifying the canonical variant ("pvt_v2_b0" through "pvt_v2_b5"). Defaults to "pvt_tiny".

embed_dimstuple of int= (64, 128, 320, 512)

Output channel width per stage (4 stages). Defaults to (64, 128, 320, 512) — PVT v2-B1.

depthstuple of int= (2, 2, 2, 2)

Number of transformer blocks per stage. Defaults to (2, 2, 2, 2).

num_headstuple of int= (1, 2, 5, 8)

Number of attention heads per stage. Defaults to (1, 2, 5, 8).

sr_ratiostuple of int= (8, 4, 2, 1)

Spatial-reduction ratio

R_i

per stage; 1 disables SR for the deepest stage. Defaults to (8, 4, 2, 1).

mlp_ratiostuple of float= (8.0, 8.0, 4.0, 4.0)

Per-stage MLP expansion ratio. Defaults to (8.0, 8.0, 4.0, 4.0) — early stages widen more than late stages, following PVT v2-B1.

Attributes

model_typeClassVar[str]

Constant string "pvt" used by the model registry.

Notes

Spatial-Reduction Attention (SRA) reduces the per-stage attention cost from $\mathcal{O}((HW)^2)$ to $\mathcal{O}((HW)^2 / R_i^2)$ by downsampling the K / V feature map by stride $R_i$ before computing attention.

Reference (v1): Wenhai Wang et al., "Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions", ICCV 2021, arXiv:2102.12122. Reference (v2): Wenhai Wang et al., "PVT v2: Improved Baselines with Pyramid Vision Transformer", CVMJ 2022, arXiv:2106.13797.

Examples

Build a PVT v2-B1 configuration with a 100-class head:
>>> from lucid.models.vision.pvt import PVTConfig
>>> cfg = PVTConfig(num_classes=100)
>>> cfg.embed_dims, cfg.depths, cfg.sr_ratios, cfg.num_classes
((64, 128, 320, 512), (2, 2, 2, 2), (8, 4, 2, 1), 100)

Used by 3

Constructors

dunder

init

→None

__init__(num_classes: int = 1000, in_channels: int = 3, variant: str = 'pvt_tiny', embed_dims: tuple[int, ...] = (64, 128, 320, 512), depths: tuple[int, ...] = (2, 2, 2, 2), num_heads: tuple[int, ...] = (1, 2, 5, 8), sr_ratios: tuple[int, ...] = (8, 4, 2, 1), mlp_ratios: tuple[float, ...] = (8.0, 8.0, 4.0, 4.0))

source edit

PVTConfig(num_classes: int = 1000, in_channels: int = 3, variant: str = 'pvt_tiny', embed_dims: tuple[int, ...] = (64, 128, 320, 512), depths: tuple[int, ...] = (2, 2, 2, 2), num_heads: tuple[int, ...] = (1, 2, 5, 8), sr_ratios: tuple[int, ...] = (8, 4, 2, 1), mlp_ratios: tuple[float, ...] = (8.0, 8.0, 4.0, 4.0))

Build a PVT v2-B1 configuration with a 100-class head: >>> from lucid.models.vision.pvt import PVTConfig >>> cfg = PVTConfig(num_classes=100) >>> cfg.embed_dims, cfg.depths, cfg.sr_ratios, cfg.num_classes ((64, 128, 320, 512), (2, 2, 2, 2), (8, 4, 2, 1), 100)

__init__(num_classes: int = 1000, in_channels: int = 3, variant: str = 'pvt_tiny', embed_dims: tuple[int, ...] = (64, 128, 320, 512), depths: tuple[int, ...] = (2, 2, 2, 2), num_heads: tuple[int, ...] = (1, 2, 5, 8), sr_ratios: tuple[int, ...] = (8, 4, 2, 1), mlp_ratios: tuple[float, ...] = (8.0, 8.0, 4.0, 4.0))