FastViT

Category: Efficient/Mobile · Code: timm/models/fastvit.py · Architectures: 12 · Pretrained weights: 21 (14 with ImageNet-1k results)

FastViT is a hybrid convolutional / transformer network that uses structural reparameterization to remove skip connections at inference. The module also includes the MobileCLIP image encoders (fastvit_mci*).

Papers

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('fastvit_sa24.apple_dist_in1k', pretrained=True)
>>> data_config = timm.data.resolve_model_data_config(model)

Model API support

Feature	`FastVit`
`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 (FastVit with fastvit_t8 at 256), 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
`fastvit_ma36`	43.98	7.82	256	2
`fastvit_mci0`	11.36	2.39	256	2
`fastvit_mci1`	21.46	4.67	256	1
`fastvit_mci2`	35.70	7.85	256	2
`fastvit_mci3`	125.07	14.82	256	1
`fastvit_mci4`	321.57	27.78	256	1
`fastvit_s12`	9.43	1.80	256	2
`fastvit_sa12`	11.55	1.94	256	2
`fastvit_sa24`	21.50	3.77	256	2
`fastvit_sa36`	31.46	5.59	256	2
`fastvit_t8`	4.00	0.69	256	2
`fastvit_t12`	7.51	1.39	256	2

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.

Weights are grouped by source: the paper or origin given in the pretrained config, else a Hub organization other than timm, else the first part of the pretrained tag (architecture.tag). Sources with a single weight are listed under other.

apple tags

Weight	Input size	Top-1 (%)	Eval size	Classes	License	Hub
`fastvit_ma36.apple_dist_in1k`	256	84.608	256	1000	fastvit-license	timm
`fastvit_ma36.apple_in1k`	256	83.884	256	1000	fastvit-license	timm
`fastvit_mci0.apple_mclip2_dfndr2b`	256	-	-	512	apple-amlr	timm
`fastvit_mci2.apple_mclip2_dfndr2b`	256	-	-	512	apple-amlr	timm
`fastvit_mci3.apple_mclip2_dfndr2b`	256	-	-	768	apple-amlr	timm
`fastvit_mci4.apple_mclip2_dfndr2b`	256	-	-	768	apple-amlr	timm
`fastvit_s12.apple_dist_in1k`	256	81.074	256	1000	fastvit-license	timm
`fastvit_s12.apple_in1k`	256	79.882	256	1000	fastvit-license	timm
`fastvit_sa12.apple_dist_in1k`	256	81.838	256	1000	fastvit-license	timm
`fastvit_sa12.apple_in1k`	256	80.846	256	1000	fastvit-license	timm
`fastvit_sa24.apple_dist_in1k`	256	83.368	256	1000	fastvit-license	timm
`fastvit_sa24.apple_in1k`	256	82.706	256	1000	fastvit-license	timm
`fastvit_sa36.apple_dist_in1k`	256	84.028	256	1000	fastvit-license	timm
`fastvit_sa36.apple_in1k`	256	83.502	256	1000	fastvit-license	timm
`fastvit_t8.apple_dist_in1k`	256	77.170	256	1000	fastvit-license	timm
`fastvit_t8.apple_in1k`	256	76.184	256	1000	fastvit-license	timm
`fastvit_t12.apple_dist_in1k`	256	80.364	256	1000	fastvit-license	timm
`fastvit_t12.apple_in1k`	256	79.276	256	1000	fastvit-license	timm

apple

Hosted by apple on the Hugging Face Hub.

Weight	Input size	Top-1 (%)	Eval size	Classes	License	Hub
`fastvit_mci0.apple_mclip`	256	-	-	512	apple-amlr	apple/mobileclip_s0_timm
`fastvit_mci1.apple_mclip`	256	-	-	512	apple-amlr	apple/mobileclip_s1_timm
`fastvit_mci2.apple_mclip`	256	-	-	512	apple-amlr	apple/mobileclip_s2_timm

Update on GitHub