English | 简体中文
A PaddlePaddle version image model zoo.
-
Install by pip：
$ pip install ppim==1.0.5 -i https://pypi.python.org/pypi
-
Install by wheel package：【Releases Packages】
-
Quick Start
import paddle from ppim import rednet26 # Load the model model, val_transforms = rednet26(pretrained=True) # Model summary paddle.summary(model, input_size=(1, 3, 224, 224)) # Random a input x = paddle.randn(shape=(1, 3, 224, 224)) # Model forword out = model(x)
-
Finetune
import paddle import paddle.nn as nn import paddle.vision.transforms as T from paddle.vision import Cifar100 from ppim import rexnet_100 # Load the model model, val_transforms = rexnet_100(pretrained=True) # Use the PaddleHapi Model model = paddle.Model(model) # Set the optimizer opt = paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters()) # Set the loss function loss = nn.CrossEntropyLoss() # Set the evaluate metric metric = paddle.metric.Accuracy(topk=(1, 5)) # Prepare the model model.prepare(optimizer=opt, loss=loss, metrics=metric) # Set the data preprocess train_transforms = T.Compose([ T.Resize(256, interpolation='bicubic'), T.RandomCrop(224), T.ToTensor(), T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) # Load the Cifar100 dataset train_dataset = Cifar100(mode='train', transform=train_transforms, backend='pil') val_dataset = Cifar100(mode='test', transform=val_transforms, backend='pil') # Finetune the model model.fit( train_data=train_dataset, eval_data=val_dataset, batch_size=256, epochs=2, eval_freq=1, log_freq=1, save_dir='save_models', save_freq=1, verbose=1, drop_last=False, shuffle=True, num_workers=0 )
-
Paper：ReXNet: Diminishing Representational Bottleneck on Convolutional Neural Network
-
Origin Repo：clovaai/rexnet
-
Evaluate Transforms：
# backend: pil # input_size: 224x224 transforms = T.Compose([ T.Resize(256, interpolation='bicubic'), T.CenterCrop(224), T.ToTensor(), T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ])
-
Model Details：
Model Params(M) FLOPs(G) Top-1 (%) Top-5 (%) ReXNet-1.0 4.8 0.40 77.9 93.9 ReXNet-1.3 7.6 0.66 79.5 94.7 ReXNet-1.5 7.6 0.66 80.3 95.2 ReXNet-2.0 16 1.5 81.6 95.7 ReXNet-3.0 34 3.4 82.8 96.2
-
Paper：Involution: Inverting the Inherence of Convolution for Visual Recognition
-
Origin Repo：d-li14/involution
-
Evaluate Transforms：
# backend: cv2 # input_size: 224x224 transforms = T.Compose([ T.Resize(256), T.CenterCrop(224), T.Normalize( mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True, data_format='HWC' ), T.ToTensor(), ])
-
Model Details：
Model Params(M) FLOPs(G) Top-1 (%) Top-5 (%) RedNet-26 9.23 1.73 75.96 93.19 RedNet-38 12.39 2.22 77.48 93.57 RedNet-50 15.54 2.71 78.35 94.13 RedNet-101 25.65 4.74 78.92 94.35 RedNet-152 33.99 6.79 79.12 94.38
-
Origin Repo：DingXiaoH/RepVGG
-
Evaluate Transforms：
# backend: pil # input_size: 224x224 transforms = T.Compose([ T.Resize(256), T.CenterCrop(224), T.ToTensor(), T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ])
-
Model Details：
Model Params(M) FLOPs(G) Top-1 (%) Top-5 (%) RepVGG-A0 8.30 1.4 72.41 RepVGG-A1 12.78 2.4 74.46 RepVGG-A2 25.49 5.1 76.48 RepVGG-B0 14.33 3.1 75.14 RepVGG-B1 51.82 11.8 78.37 RepVGG-B2 80.31 18.4 78.78 RepVGG-B3 110.96 26.2 80.52 RepVGG-B1g2 41.36 8.8 77.78 RepVGG-B1g4 36.12 7.3 77.58 RepVGG-B2g4 55.77 11.3 79.38 RepVGG-B3g4 75.62 16.1 80.21
-
Origin Repo：naver-ai/pit
-
Evaluate Transforms：
# backend: pil # input_size: 224x224 transforms = T.Compose([ T.Resize(248, interpolation='bicubic'), T.CenterCrop(224), T.ToTensor(), T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ])
-
Model Details：
Model Params(M) FLOPs(G) Top-1 (%) Top-5 (%) PiT-Ti 4.9 0.71 73.0 PiT-XS 10.6 1.4 78.1 PiT-S 23.5 2.9 80.9 PiT-B 73.8 12.5 82.0 PiT-Ti distilled 4.9 0.71 74.6 PiT-XS distilled 10.6 1.4 79.1 PiT-S distilled 23.5 2.9 81.9 PiT-B distilled 73.8 12.5 84.0
-
Paper：Training data-efficient image transformers & distillation through attention
-
Origin Repo：facebookresearch/deit
-
Evaluate Transforms：
# backend: pil # input_size: 224x224 transforms = T.Compose([ T.Resize(248, interpolation='bicubic'), T.CenterCrop(224), T.ToTensor(), T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) # backend: pil # input_size: 384x384 transforms = T.Compose([ T.Resize(384, interpolation='bicubic'), T.CenterCrop(384), T.ToTensor(), T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ])
-
Model Details：
Model Params(M) FLOPs(G) Top-1 (%) Top-5 (%) DeiT-tiny 5 72.2 91.1 DeiT-small 22 79.9 95.0 DeiT-base 86 81.8 95.6 DeiT-tiny distilled 6 74.5 91.9 DeiT-small distilled 22 81.2 95.4 DeiT-base distilled 87 83.4 96.5 DeiT-base 384 87 82.9 96.2 DeiT-base distilled 384 88 85.2 97.2
@article{han2020rexnet,
title = {ReXNet: Diminishing Representational Bottleneck on Convolutional Neural Network},
author = {Han, Dongyoon and Yun, Sangdoo and Heo, Byeongho and Yoo, YoungJoon},
journal = {arXiv preprint arXiv:2007.00992},
year = {2020},
}
@InProceedings{Li_2021_CVPR,
title = {Involution: Inverting the Inherence of Convolution for Visual Recognition},
author = {Li, Duo and Hu, Jie and Wang, Changhu and Li, Xiangtai and She, Qi and Zhu, Lei and Zhang, Tong and Chen, Qifeng},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2021}
}
@article{ding2021repvgg,
title={RepVGG: Making VGG-style ConvNets Great Again},
author={Ding, Xiaohan and Zhang, Xiangyu and Ma, Ningning and Han, Jungong and Ding, Guiguang and Sun, Jian},
journal={arXiv preprint arXiv:2101.03697},
year={2021}
}
@article{heo2021pit,
title={Rethinking Spatial Dimensions of Vision Transformers},
author={Byeongho Heo and Sangdoo Yun and Dongyoon Han and Sanghyuk Chun and Junsuk Choe and Seong Joon Oh},
journal={arXiv: 2103.16302},
year={2021},
}
@article{touvron2020deit,
title = {Training data-efficient image transformers & distillation through attention},
author = {Hugo Touvron and Matthieu Cord and Matthijs Douze and Francisco Massa and Alexandre Sablayrolles and Herv'e J'egou},
journal = {arXiv preprint arXiv:2012.12877},
year = {2020}
}