提交 ab8c592a 编写于 作者: littletomatodonkey's avatar littletomatodonkey

polish mainpage readme

上级 162281e7
此差异已折叠。
...@@ -57,7 +57,7 @@ It takes 196 times for grid search, and takes 10 times less for Bayesian search. ...@@ -57,7 +57,7 @@ It takes 196 times for grid search, and takes 10 times less for Bayesian search.
## Large-scale image classification ## Large-scale image classification
In practical applications, due to the lack of training data, the classification model trained on the ImageNet1k data set is often used as the pretrained model for other image classification tasks. In order to further help solve practical problems, based on ResNet50_vd, Baidu open sourced a self-developed large-scale classification pretrained model, in which the training data contains 100,000 categories and 43 million pictures. In practical applications, due to the lack of training data, the classification model trained on the ImageNet1k data set is often used as the pretrained model for other image classification tasks. In order to further help solve practical problems, based on ResNet50_vd, Baidu open sourced a self-developed large-scale classification pretrained model, in which the training data contains 100,000 categories and 43 million pictures. The pretrained model can be downloaded as follows:[**download link**](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_10w_pretrained.tar)
We conducted transfer learning experiments on 6 self-collected datasets, We conducted transfer learning experiments on 6 self-collected datasets,
......
...@@ -11,3 +11,12 @@ RegNet was proposed in 2020 by Facebook to deepen the concept of design space. B ...@@ -11,3 +11,12 @@ RegNet was proposed in 2020 by Facebook to deepen the concept of design space. B
| ResNeSt50_fast_1s1x64d | 0.8035 | 0.9528| 0.8035 | -| 8.68 | 26.3 | | ResNeSt50_fast_1s1x64d | 0.8035 | 0.9528| 0.8035 | -| 8.68 | 26.3 |
| ResNeSt50 | 0.8102 | 0.9542| 0.8113 | -| 10.78 | 27.5 | | ResNeSt50 | 0.8102 | 0.9542| 0.8113 | -| 10.78 | 27.5 |
| RegNetX_4GF | 0.7850 | 0.9416| 0.7860 | -| 8.0 | 22.1 | | RegNetX_4GF | 0.7850 | 0.9416| 0.7860 | -| 8.0 | 22.1 |
## Inference speed based on T4 GPU
| Models | Crop Size | Resize Short Size | FP16<br>Batch Size=1<br>(ms) | FP16<br>Batch Size=4<br>(ms) | FP16<br>Batch Size=8<br>(ms) | FP32<br>Batch Size=1<br>(ms) | FP32<br>Batch Size=4<br>(ms) | FP32<br>Batch Size=8<br>(ms) |
|--------------------|-----------|-------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------|
| ResNeSt50_fast_1s1x64d | 224 | 256 | - | - | - | - | - | - |
| ResNeSt50 | 224 | 256 | - | - | - | - | - | - |
| RegNetX_4GF | 224 | 256 | 6.69042 | 8.01664 | 11.60608 | 6.46478 | 11.19862 | 16.89089 |
...@@ -57,7 +57,7 @@ Mixup: [False, True] ...@@ -57,7 +57,7 @@ Mixup: [False, True]
## 二、 大规模分类模型 ## 二、 大规模分类模型
在实际应用中,由于训练数据的匮乏,往往将ImageNet1k数据集训练的分类模型作为预训练模型,进行图像分类的迁移学习。为了进一步助力解决实际问题,基于ResNet50_vd, 百度开源了自研的大规模分类预训练模型,其中训练数据为10万个类别,4300万张图片。 在实际应用中,由于训练数据的匮乏,往往将ImageNet1k数据集训练的分类模型作为预训练模型,进行图像分类的迁移学习。为了进一步助力解决实际问题,基于ResNet50_vd, 百度开源了自研的大规模分类预训练模型,其中训练数据为10万个类别,4300万张图片。10万类预训练模型的下载地址:[**下载地址**](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_10w_pretrained.tar)
我们在6个自有采集的数据集上进行迁移学习实验,采用一组固定参数以及网格搜索方式,其中训练轮数设置为20epochs,选用ResNet50_vd模型,ImageNet预训练精度为79.12%。实验数据集参数以及模型精度的对比结果如下: 我们在6个自有采集的数据集上进行迁移学习实验,采用一组固定参数以及网格搜索方式,其中训练轮数设置为20epochs,选用ResNet50_vd模型,ImageNet预训练精度为79.12%。实验数据集参数以及模型精度的对比结果如下:
......
# ResNeSt以及RegNet网络 # ResNeSt与RegNet系列
## 概述 ## 概述
...@@ -14,3 +14,12 @@ RegNet是由facebook于2020年提出,旨在深化设计空间理念的概念 ...@@ -14,3 +14,12 @@ RegNet是由facebook于2020年提出,旨在深化设计空间理念的概念
| ResNeSt50_fast_1s1x64d | 0.8035 | 0.9528| 0.8035 | -| 8.68 | 26.3 | | ResNeSt50_fast_1s1x64d | 0.8035 | 0.9528| 0.8035 | -| 8.68 | 26.3 |
| ResNeSt50 | 0.8102 | 0.9542| 0.8113 | -| 10.78 | 27.5 | | ResNeSt50 | 0.8102 | 0.9542| 0.8113 | -| 10.78 | 27.5 |
| RegNetX_4GF | 0.7850 | 0.9416| 0.7860 | -| 8.0 | 22.1 | | RegNetX_4GF | 0.7850 | 0.9416| 0.7860 | -| 8.0 | 22.1 |
## 基于T4 GPU的预测速度
| Models | Crop Size | Resize Short Size | FP16<br>Batch Size=1<br>(ms) | FP16<br>Batch Size=4<br>(ms) | FP16<br>Batch Size=8<br>(ms) | FP32<br>Batch Size=1<br>(ms) | FP32<br>Batch Size=4<br>(ms) | FP32<br>Batch Size=8<br>(ms) |
|--------------------|-----------|-------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------|
| ResNeSt50_fast_1s1x64d | 224 | 256 | - | - | - | - | - | - |
| ResNeSt50 | 224 | 256 | - | - | - | - | - | - |
| RegNetX_4GF | 224 | 256 | 6.69042 | 8.01664 | 11.60608 | 6.46478 | 11.19862 | 16.89089 |
...@@ -16,9 +16,6 @@ import logging ...@@ -16,9 +16,6 @@ import logging
import os import os
import datetime import datetime
from imp import reload
reload(logging)
logging.basicConfig( logging.basicConfig(
level=logging.INFO, level=logging.INFO,
format="%(asctime)s %(levelname)s: %(message)s", format="%(asctime)s %(levelname)s: %(message)s",
...@@ -26,7 +23,7 @@ logging.basicConfig( ...@@ -26,7 +23,7 @@ logging.basicConfig(
def time_zone(sec, fmt): def time_zone(sec, fmt):
real_time = datetime.datetime.now() + datetime.timedelta(hours=8) real_time = datetime.datetime.now()
return real_time.timetuple() return real_time.timetuple()
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册