pvt_v2_b4

lucid.models.pvt_v2_b4(img_size: int = 224, num_classes: int = 1000, in_channels: int = 3, **kwargs) → PVT_V2

The pvt_v2_b4 function instantiates the PVTv2-B4 model, a high-capacity variant of the Pyramid Vision Transformer V2 family. With a significantly deeper second and third stage, it provides rich hierarchical representations and is designed for demanding image understanding tasks where precision is critical.

Total Parameters: 62,556,072

Function Signature

@register_model
def pvt_v2_b4(img_size: int = 224, num_classes: int = 1000, in_channels: int = 3, **kwargs) -> PVT_V2

Parameters

  • img_size (int, optional): The input image size. Default is 224.

  • num_classes (int, optional): The number of output classes for classification. Default is 1000.

  • in_channels (int, optional): Number of input channels. Default is 3.

  • kwargs (dict, optional): Additional parameters for customization, including:

  • embed_dims (list[int]): Embedding dimensions across four stages. Default is [64, 128, 320, 512].

  • depths (list[int]): Number of transformer blocks per stage. Default is [3, 8, 27, 3].

  • num_heads (list[int]): Multi-head self-attention head counts per stage. Default is [1, 2, 5, 8].

  • mlp_ratios (list[int]): MLP expansion ratios per stage. Default is [8, 8, 4, 4].

  • sr_ratios (list[int]): Spatial reduction ratios for attention projection. Default is [8, 4, 2, 1].

  • qkv_bias (bool): Whether to include bias terms in query/key/value. Default is True.

  • norm_layer (Callable): Normalization layer used in the model. Default is nn.LayerNorm(eps=1e-6).

Returns

  • PVT_V2: An instance of the PVT_V2 class configured as a PVTv2-B4 vision transformer.

Examples

>>> import lucid.models as models
>>> model = models.pvt_v2_b4()
>>> print(model)
PVT_V2(img_size=224, num_classes=1000, patch_size=7, embed_dims=[64, 128, 320, 512],
       num_heads=[1, 2, 5, 8], depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1], linear=False, ...)