Model Zoo

The following table provides an overview of the models available in DLFeat. Performance metrics (Acc., mAP, R@1, etc.) are typically reported on standard benchmarks (e.g., ImageNet, Kinetics-400, GLUE, MSCOCO, MSR-VTT). FLOPS and Speed are indicative and can vary significantly based on hardware, batch size, input resolution, and specific implementation. “SSL” denotes Self-Supervised Learning. “Multimodal” models are trained on multiple data types.

DLFeat Model Zoo

Modality	Model Name (Identifier)	Feat. Dim	Performance (Benchmark)	FLOPS (G)	Speed	Supervision	Source
Image
Image	resnet18	512	69.76% (ImageNet Top-1)	1.8	Fast	Supervised	torchvision
Image	resnet34	512	73.30% (ImageNet Top-1)	3.6	Fast	Supervised	torchvision
Image	resnet50	2048	80.86% (ImageNet Top-1, tv)	4.1	Medium	Supervised	torchvision_or_timm
Image	resnet101	2048	81.88% (ImageNet Top-1, tv)	7.8	Medium	Supervised	torchvision_or_timm
Image	resnet152	2048	82.28% (ImageNet Top-1, tv)	11.5	Slower	Supervised	torchvision_or_timm
Image	efficientnet_b0	1280	77.69% (ImageNet Top-1)	0.39	Very Fast	Supervised	timm
Image	efficientnet_b2	1408	80.51% (ImageNet Top-1)	1.0	Fast	Supervised	timm
Image	efficientnet_b4	1792	83.37% (ImageNet Top-1)	4.4	Medium	Supervised	timm
Image	mobilenet_v2	1280	71.88% (ImageNet Top-1)	0.3	Very Fast	Supervised	torchvision
Image	mobilenet_v3_small	576	67.67% (ImageNet Top-1)	0.06	Very Fast	Supervised	torchvision
Image	mobilenet_v3_large	960	74.04% (ImageNet Top-1)	0.22	Very Fast	Supervised	torchvision
Image	vit_tiny_patch16_224	192	75.4% (ImageNet Top-1, DeiT)	1.3	Fast	Supervised (DeiT)	timm
Image	vit_small_patch16_224	384	81.2% (ImageNet Top-1, DeiT)	4.6	Medium	Supervised (DeiT)	timm
Image	vit_base_patch16_224	768	85.2% (ImageNet Top-1, MAE FT)	17.6	Medium	SSL (MAE)	timm
Image	dinov2_base	768	82.8% (ImageNet k-NN, ViT-B/14)	~33 (ViT-B/14)	Medium	SSL (DINOv2)	Transformers
Video
Video	r2plus1d_18	512	65.2% (K400 Top-1)	31.6 (16f)	Medium	Supervised	torchvision
Video	video_swin_t	768	78.8% (K400 Top-1)	48 (32x224^2)	Medium	Supervised	torchvision
Video	video_swin_s	768	81.6% (K400 Top-1)	92 (32x224^2)	Slower	Supervised	torchvision
Video	video_swin_b	1024	82.7% (K400 Top-1)	199 (32x224^2)	Slower	Supervised	torchvision
Video	videomae_base_k400_pt	768	81.2% (K400 Top-1, ViT-B)	~168 (16x224^2)	Medium	Supervised (PT+FT)	transformers
Audio
Audio	wav2vec2_base	768	~6.9% (LibriSpeech WER, no LM)	94.5M Params	Fast	SSL (Wav2Vec2)	Transformers
Audio	ast_vit_base_patch16_224	768	0.459 (AudioSet mAP)	87M Params	Medium	Supervised	Transformers
Text
Text	sentence-bert	384	85.3 (STS-B Spearman)	N/A	Very Fast	SSL (SBERT)	sentence-transformers
Text	bert_base_uncased	768	79.6 (GLUE Avg.)	110M Params	Medium	SSL (BERT)	Transformers
Multimodal
Image-Text	clip_vit_b32	512	63.3% (ImageNet zero-shot)	N/A	Fast	Multimodal SSL	Transformers
Video-Text	xclip_base_patch16	512	46.7% (MSR-VTT R@1)	N/A	Medium	Multimodal SSL	Transformers

Note on FLOPS/Speed: These are highly approximate and depend on input size, hardware, and batching. “Fast” might mean >100 FPS for images on a modern GPU. “N/A” indicates data not readily found or highly variable. For parameter counts, “M Params” refers to millions of parameters.