Fix feature_info.reduction

[brianhou0208]

Hi @rwightman ,

This PR fix feature_info.reduction for:

• MetaFormer (CaFormer, ConvFormer, PoolFormer, PoolFormerV2)
• DaViT
• EfficientFormer

Test code

import timm
import torch

model_list = [
    ["caformer_s18", 224],
    ["convformer_s18", 224],
    ["poolformer_s12", 224],
    ["poolformerv2_s12", 224],
    ["davit_tiny", 224],
    ["efficientformer_l1", 224]
]

if __name__ == "__main__":
    for model_name, img_size in model_list:
        x = torch.rand(1, 3, img_size, img_size)
        model = timm.create_model(f"{model_name}", features_only=True).eval()
        y = model(x)
        print(f"timm-{model_name}-(C, H, W) = {(3, img_size, img_size)}")
        print(f"    Feature shape: {[f.detach().numpy().shape[1:] for f in y]}")
        print(f"    Feature channels: {model.feature_info.channels()}")
        print(f"    Feature reduction: {model.feature_info.reduction()}")

Before this PR

pytorch-image-models/timm/models/metaformer.py

Line 544 in ea23107

self.feature_info += [dict(num_chs=dims[i], reduction=2, module=f'stages.{i}')]

Output

timm-caformer_s18-(C, H, W) = (3, 224, 224)
    Feature shape: [(64, 56, 56), (128, 28, 28), (320, 14, 14), (512, 7, 7)]
    Feature channels: [64, 128, 320, 512]
    Feature reduction: [2, 2, 2, 2]

After this PR

self.feature_info += [dict(num_chs=dims[i], reduction=2**(i+2), module=f'stages.{i}')]

Output

timm-caformer_s18-(C, H, W) = (3, 224, 224)
    Feature shape: [(64, 56, 56), (128, 28, 28), (320, 14, 14), (512, 7, 7)]
    Feature channels: [64, 128, 320, 512]
    Feature reduction: [4, 8, 16, 32]

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.