pvt_huge¶
The pvt_huge function instantiates the PVT-Huge model, the largest and most powerful variant of the Pyramid Vision Transformer (PVT). This variant is designed for large-scale and highly complex vision tasks, providing the most detailed hierarchical feature representation.
The PVT-Huge model provides excellent performance on challenging tasks and is ideal for large-scale datasets.
Total Parameters: 286,706,920
Function Signature¶
@register_model
def pvt_huge(img_size: int = 224, num_classes: int = 1000, **kwargs) -> PVT
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.
kwargs (dict, optional): Additional parameters for customization, including:
embed_dims (list[int]): A list specifying the embedding dimensions at different hierarchical levels. The default configuration for PVT-Huge is [128, 256, 512, 768]. This allows for the finest resolution of feature extraction at each stage, making it suitable for large-scale datasets and complex tasks.
depths (list[int]): A list specifying the number of transformer blocks in each stage. The default values for PVT-Huge are [3, 10, 60, 3]. This provides very deep models in the intermediate layers for increased abstraction and representation learning.
num_heads (list[int]): The number of self-attention heads in each transformer stage. The default values are [2, 4, 8, 12]. The number of heads increases progressively to capture a richer set of dependencies across the hierarchical layers.
mlp_ratios (list[float]): The ratio of MLP hidden dimensions to the embedding dimension in each stage. Default values are [8.0, 8.0, 4.0, 4.0], balancing computational efficiency with expressiveness.
sr_ratios (list[float]): The spatial reduction ratios used in the self-attention mechanism. Default values are [8.0, 4.0, 2.0, 1.0]. These values control how much spatial information is preserved across different transformer blocks, optimizing both local and global contextual relationships.
Returns¶
PVT: An instance of the PVT class configured as a PVT-Huge vision transformer.
Examples¶
>>> import lucid.models as models
>>> model = models.pvt_huge()
>>> print(model)