FocalNet

Category: Hybrid · Code: timm/models/focalnet.py · Architectures: 12 · Pretrained weights: 12 (6 with ImageNet-1k results)

Focal Modulation Networks (FocalNet) replace self-attention with focal modulation, which aggregates context at several granularities with depthwise convolutions and gates it into each query token.

Papers

Focal Modulation Networks

Usage

See the Quickstart for loading pretrained weights, image preprocessing and fine-tuning, and Feature Extraction for using the models as backbones.

>>> import timm
>>> model = timm.create_model('focalnet_tiny_srf.ms_in1k', pretrained=True)
>>> data_config = timm.data.resolve_model_data_config(model)

Model API support

Feature	`FocalNet`
`features_only=True` feature maps	✓ (strides 4, 8, 16, 32)
`forward_intermediates()`	✓
`prune_intermediate_layers()`	✓
Any input size, no changes needed	✓
`set_input_size()`	–
`dynamic_img_size=True`	–
`set_grad_checkpointing()`	✓
Layer-wise LR decay (`group_matcher()`)	✓
`reset_classifier()`	✓

✓ supported, ✓ (square) works for square sizes only, ✗ not supported, – not needed (the model already accepts any input size). Checked by running each feature on a randomly initialized model (FocalNet with focalnet_tiny_srf at 224), other architectures of the same class are expected to behave the same. See the Model API page for details.

Architectures

Architecture	Params (M)	GMACs	Input size	Pretrained weights
`focalnet_base_lrf`	88.75	15.43	224	1
`focalnet_base_srf`	88.15	15.28	224	1
`focalnet_huge_fl3`	745.28	118.26	224	1
`focalnet_huge_fl4`	686.46	118.90	224	1
`focalnet_large_fl3`	239.13	105.06	384	1
`focalnet_large_fl4`	239.32	105.20	384	1
`focalnet_small_lrf`	50.34	8.74	224	1
`focalnet_small_srf`	49.89	8.62	224	1
`focalnet_tiny_lrf`	28.65	4.49	224	1
`focalnet_tiny_srf`	28.43	4.42	224	1
`focalnet_xlarge_fl3`	408.79	185.61	384	1
`focalnet_xlarge_fl4`	409.03	185.79	384	1

GMACs and parameter counts are for the default input size and classifier of each architecture, from the benchmark results.

Pretrained weights

Top-1 is ImageNet-1k validation accuracy from results-imagenet.csv at the listed eval size, - if the weight has no ImageNet-1k result (e.g. pretraining or image-text weights). Classes is the size of the pretrained classifier, none for weights without one.

Weight	Input size	Top-1 (%)	Eval size	Classes	License	Hub
`focalnet_base_lrf.ms_in1k`	224	83.828	224	1000	mit	timm
`focalnet_base_srf.ms_in1k`	224	83.808	224	1000	mit	timm
`focalnet_huge_fl3.ms_in22k`	224	-	-	21842	mit	timm
`focalnet_huge_fl4.ms_in22k`	224	-	-	none	mit	timm
`focalnet_large_fl3.ms_in22k`	384	-	-	21842	mit	timm
`focalnet_large_fl4.ms_in22k`	384	-	-	21842	mit	timm
`focalnet_small_lrf.ms_in1k`	224	83.508	224	1000	mit	timm
`focalnet_small_srf.ms_in1k`	224	83.428	224	1000	mit	timm
`focalnet_tiny_lrf.ms_in1k`	224	82.174	224	1000	mit	timm
`focalnet_tiny_srf.ms_in1k`	224	82.146	224	1000	mit	timm
`focalnet_xlarge_fl3.ms_in22k`	384	-	-	21842	mit	timm
`focalnet_xlarge_fl4.ms_in22k`	384	-	-	21842	mit	timm

Update on GitHub