Available architectures

Here is the complete list of all the neural network architectures available in Studio. When available, links to the research papers are provided.

Pre-training

All our architectures are pre-trained on real-world, open datasets. Training them from scratch would take days or weeks to have correct predictions. Thanks to the pre-training, all you are doing is fine-tuning the model to your specific use-case. You can see the pre-training dataset in the table below.

Classification and tagging backbone architectures

Backbone	Inference time on Nvidia T4 (ms) (Run v1.0)	Inference time on Nvidia T4 (ms) (Run v1.1)	Input size	Training batch size	Trained on	Research paper
EfficientNet B0	10.4	10.7	224x224	32	ImageNet	arXiv
EfficientNet B1	11.4	11.4	240x240	32	ImageNet	arXiv
EfficientNet B2	12.5	12.2	260x260	32	ImageNet	arXiv
EfficientNet B3	16.7	13.5	300x300	32	ImageNet	arXiv
EfficientNet B4	24.8	19.8	380x380	16	ImageNet	arXiv
EfficientNet B5	36.4	34.9	456x456	8	ImageNet	arXiv
EfficientNet B6	61.0	58.4	528x528	4	ImageNet	arXiv
Inception-Resnet v2	33.1	26.7	299x299	32	ImageNet	arXiv
ResNet-50	11.2	11.6	224x224	32	ImageNet	arXiv
ResNet-101	16.8	N/A	224x224	32	ImageNet	arXiv
ResNet-152	23.2	21.7	224x224	32	ImageNet	arXiv
Inception v1	7.5	10.8	224x224	32	ImageNet	arXiv
Inception v2	10.5	10.9	224x224	32	ImageNet	arXiv
Inception v3	15.4	12.7	299x299	32	ImageNet	arXiv
Inception v4	22.8	21.0	299x299	32	ImageNet	arXiv
VGG 16	63.1	62.8	224x224	32	ImageNet	arXiv
VGG 19	64.7	N/A	224x224	32	ImageNet	arXiv

Detection meta-architectures

Architecture	Backbone	Inference time on Nvidia T4 (ms) (Run v1.0)	Inference time on Nvidia T4 (ms) (Run v1.1)	Input size	Training batch size	Trained on	Research paper
EfficientDet	Eff. Net B0	26.5	42.2	512x512	16	COCO 2018	arXiv
	Eff. Net B1	37.9	49.0	640x640	8	COCO 2018	arXiv
	Eff. Net B2	57.2	55.3	768x768	4	COCO 2018	arXiv
	Eff. Net B3	99.1	92.5	896x896	2	COCO 2018	arXiv
	Eff. Net B4	174	167	1024x1024	1	COCO 2018	arXiv
	Eff. Net B5	N/A	N/A	1280x1280	1	COCO 2018	arXiv
Yolo v2	Darknet 19	21.3	51.8	416x416	64	VOC 2007+2012	arXiv
Yolo v3	Darknet 53	33.6	53.8	416x416	64	ImageNet 2012	arXiv
Yolo v3	Keras - Darknet 53	33.6	53.8	416x416	16	ImageNet 2012	arXiv
Yolo v3 SPP	Darknet 53	58.9	56.5	608x608	64	COCO 2018
Faster-RCNN	Resnet-50	153	N/A	1024x1024*	1	COCO 2018	arXiv
	Resnet-101	195	N/A	1024x1024*	1	COCO 2018	arXiv
RFCN	Resnet-101	N/A	N/A	1024x1024*	1	COCO 2018	arXiv
SSD	Inception v2	12.1	46.6	300x300	24	COCO 2018	arXiv
	MobileNet v1	9.9	43.8	300x300	24	COCO 2018	arXiv
	MobileNet v2	9.6	50.6	300x300	24	COCO 2018	arXiv
SSDLite	MobileNet v2	10.2	50.4	300x300	24	COCO 2018	arXiv

(*) Faster-RCNN and RFCN do not require a fixed image input size. As such they can accept images from 600 to 1024 pixel.