TResNet: High Performance GPU-Dedicated Architecture

Packaged TResNet based on Official PyTorch Implementation [paper] [github]

Installation

Install with pip:

pip install torch_tresnet

or directly:

pip install git+https://github.com/tczhangzhi/torch-tresnet

Use

Follow the grammatical conventions of torchvision

from torch_tresnet import tresnet_m, tresnet_l, tresnet_xl, tresnet_m_448, tresnet_l_448, tresnet_xl_448

# pretrianed on 224*224
model = tresnet_m(pretrained=True)
model = tresnet_m(pretrained=True, num_classes=10)
model = tresnet_m(pretrained=True, num_classes=10, in_chans=3)

# pretrianed on 448*448
model = tresnet_m_448(pretrained=True)
model = tresnet_m_448(pretrained=True, num_classes=10)
model = tresnet_m_448(pretrained=True, num_classes=10, in_chans=3)

For a demonstration of training and testing, please see the jupyter notebook, which is an unofficial implementation, please do not use this implementation as the final evaluation standard

Main Results

TResNet Models

TResNet models accuracy and GPU throughput on ImageNet, compared to ResNet50. All measurements were done on Nvidia V100 GPU, with mixed precision. All models are trained on input resolution of 224.

Models	Top Training Speed (img/sec)	Top Inference Speed (img/sec)	Max Train Batch Size	Top-1 Acc.
ResNet50	805	2830	288	79.0
EfficientNetB1	440	2740	196	79.2
TResNet-M	730	2930	512	80.7
TResNet-L	345	1390	316	81.4
TResNet-XL	250	1060	240	82.0

Comparison To Other Networks

Comparison of ResNet50 to top modern networks, with similar top-1 ImageNet accuracy. All measurements were done on Nvidia V100 GPU with mixed precision. For gaining optimal speeds, training and inference were measured on 90% of maximal possible batch size. Except TResNet-M, all the models' ImageNet scores were taken from the public repository, which specialized in providing top implementations for modern networks. Except EfficientNet-B1, which has input resolution of 240, all other models have input resolution of 224.

Model	Top Training Speed (img/sec)	Top Inference Speed (img/sec)	Top-1 Acc.	Flops[G]
ResNet50	805	2830	79.0	4.1
ResNet50-D	600	2670	79.3	4.4
ResNeXt50	490	1940	78.5	4.3
EfficientNetB1	440	2740	79.2	0.6
SEResNeXt50	400	1770	79.0	4.3
MixNet-L	400	1400	79.0	0.5
TResNet-M	730	2930	80.7	5.5

Transfer Learning SotA Results

Comparison of TResNet to state-of-the-art models on transfer learning datasets (only ImageNet-based transfer learning results). Models inference speed is measured on a mixed precision V100 GPU. Since no official implementation of Gpipe was provided, its inference speed is unknown.

Dataset	Model	Top-1 Acc.	Speed img/sec	Input
CIFAR-10	Gpipe	99.0	-	480
CIFAR-10	TResNet-XL	99.0	1060	224
CIFAR-100	EfficientNet-B7	91.7	70	600
CIFAR-100	TResNet-XL	91.5	1060	224
Stanford Cars	EfficientNet-B7	94.7	70	600
Stanford Cars	TResNet-L	96.0	500	368
Oxford-Flowers	EfficientNet-B7	98.8	70	600
Oxford-Flowers	TResNet-L	99.1	500	368