From cc8e0d09359e7e76b0cc506d02670442b09f07e3 Mon Sep 17 00:00:00 2001 From: cuicheng01 <45199522+cuicheng01@users.noreply.github.com> Date: Mon, 23 Sep 2019 15:48:32 +0800 Subject: [PATCH] Update readme about efficientnet (#3389) * update TensorRT 1.5.2 inference time * add EfficientNet configuration in README --- PaddleCV/image_classification/README.md | 7 ++++--- PaddleCV/image_classification/README_en.md | 7 ++++--- 2 files changed, 8 insertions(+), 6 deletions(-) diff --git a/PaddleCV/image_classification/README.md b/PaddleCV/image_classification/README.md index d2ccf79e..402e0efd 100644 --- a/PaddleCV/image_classification/README.md +++ b/PaddleCV/image_classification/README.md @@ -241,8 +241,9 @@ PaddlePaddle/Models ImageClassification 支持自定义数据 - 注意 - 1：ResNet50_vd_v2是ResNet50_vd蒸馏版本。 - - 2：InceptionV4和Xception采用的输入图像的分辨率为299x299，DarkNet53为256x256，Fix_ResNeXt101_32x48d_wsl为320x320，其余模型使用的分辨率均为224x224。在预测时，DarkNet53与Fix_ResNeXt101_32x48d_wsl系列网络resize_short_size与输入的图像分辨率的宽或高相同，InceptionV4和Xception网络resize_short_size为320，其余网络resize_short_size均为256。 - - 3：调用动态链接库预测时需要将训练模型转换为二进制模型 + - 2：除EfficientNet外，InceptionV4和Xception采用的输入图像的分辨率为299x299，DarkNet53为256x256，Fix_ResNeXt101_32x48d_wsl为320x320，其余模型使用的分辨率均为224x224。在预测时，DarkNet53与Fix_ResNeXt101_32x48d_wsl系列网络resize_short_size与输入的图像分辨率的宽或高相同，InceptionV4和Xception网络resize_short_size为320，其余网络resize_short_size均为256。 + - 3: EfficientNetB0~B7的分辨率大小分别为224x224，240x240，260x260，300x300，380x380，456x456，528x528，600x600，预测时的resize_short_size在其分辨率的长或高的基础上加32，如EfficientNetB1的resize_short_size为272，在该系列模型训练和预测的过程中，图片resize参数interpolation的值设置为2（cubic插值方式），该模型在训练过程中使用了指数滑动平均策略，具体请参考[指数滑动平均](https://www.paddlepaddle.org.cn/documentation/docs/zh/1.5/api_cn/optimizer_cn.html#exponentialmovingaverage)。 + - 4：调用动态链接库预测时需要将训练模型转换为二进制模型。 ```bash python infer.py \ @@ -251,7 +252,7 @@ PaddlePaddle/Models ImageClassification 支持自定义数据 --save_inference=True ``` - - 4: ResNeXt101_wsl系列的预训练模型转自pytorch模型，详情见[ResNeXt wsl](https://pytorch.org/hub/facebookresearch_WSL-Images_resnext/)。 + - 5: ResNeXt101_wsl系列的预训练模型转自pytorch模型，详情见[ResNeXt wsl](https://pytorch.org/hub/facebookresearch_WSL-Images_resnext/)。 ### AlexNet diff --git a/PaddleCV/image_classification/README_en.md b/PaddleCV/image_classification/README_en.md index 0a8b6115..5511f142 100644 --- a/PaddleCV/image_classification/README_en.md +++ b/PaddleCV/image_classification/README_en.md @@ -226,15 +226,16 @@ Pretrained models can be downloaded by clicking related model names. - Note - 1: ResNet50_vd_v2 is the distilled version of ResNet50_vd. - - 2: The image resolution feeded in InceptionV4 and Xception net is ```299x299```, Fix_ResNeXt101_32x48d_wsl is ```320x320```, DarkNet is ```256x256```, others are ```224x224```.In test time, the resize_short_size of the DarkNet53 and Fix_ResNeXt101_32x48d_wsl series networks is the same as the width or height of the input image resolution, the InceptionV4 and Xception network resize_short_size is 320, and the other networks resize_short_size are 256. - - 3: It's necessary to convert the train model to a binary model when appling dynamic link library to infer, One can do it by running following command: + - 2: In addition to EfficientNet, the image resolution feeded in InceptionV4 and Xception net is ```299x299```, Fix_ResNeXt101_32x48d_wsl is ```320x320```, DarkNet is ```256x256```, others are ```224x224```.In test time, the resize_short_size of the DarkNet53 and Fix_ResNeXt101_32x48d_wsl series networks is the same as the width or height of the input image resolution, the InceptionV4 and Xception network resize_short_size is 320, and the other networks resize_short_size are 256. + - 3: The resolutions of EfficientNetB0~B7 are ```224x224```,```240x240```,```260x260```,```300x300```,```380x380```,```456x456```,```528x528```,```600x600``` respectively, the resize_short_size in the inference phase is increased by 32 on the basis of the length or height of the resolution, for example, the resize_short_size of EfficientNetB1 is 272.In the process of training and inference phase of these series of models, the value of the resize parameter interpolation is set to 2 (cubic interpolation mode). Besides, the model uses ExponentialMovingAverage during the training process, this trick please refer to [ExponentialMovingAverage](https://www.paddlepaddle.org.cn/documentation/docs/en/1.5/api/optimizer.html#exponentialmovingaverage). + - 4: It's necessary to convert the train model to a binary model when appling dynamic link library to infer, One can do it by running following command: ```bash python infer.py\ --model=model_name \ --pretrained_model=${path_to_pretrained_model} \ --save_inference=True ``` - - 4: The pretrained model of the ResNeXt101_wsl series network is converted from the pytorch model. Please refer to [RESNEXT WSL](https://pytorch.org/hub/facebookresearch_WSL-Images_resnext/) for details. + - 5: The pretrained model of the ResNeXt101_wsl series network is converted from the pytorch model. Please refer to [RESNEXT WSL](https://pytorch.org/hub/facebookresearch_WSL-Images_resnext/) for details. ### AlexNet |Model | Top-1 | Top-5 | Paddle Fluid inference time(ms) | Paddle TensorRT inference time(ms) | -- GitLab