Skip to content

Commit

Permalink
modify the name of dataset and add release note in README.md (PaddleP…
Browse files Browse the repository at this point in the history
…addle#14)

* add update notes

* update dataset name
  • Loading branch information
lilong12 authored Dec 23, 2019
1 parent 5910453 commit 8abf442
Show file tree
Hide file tree
Showing 2 changed files with 13 additions and 4 deletions.
13 changes: 12 additions & 1 deletion README.md
Original file line number Diff line number Diff line change
Expand Up @@ -57,6 +57,10 @@ softmax的计算公示如下图所示:

飞桨是由百度研发的一款源于产业实践的开源深度学习平台,致力于让深度学习技术的创新与应用更简单。PLSC基于飞桨平台研发,实现与飞桨平台的无缝链接,可以更好地服务产业实践。

- 支持大规模分类

单机8张V100 GPU配置下,支持的分类类别数增大了2.52倍;

- 包含多种预训练模型

除了PLSC库源码,我们还发布了基于ResNet50模型、ResNet101模型、ResNet152模型的大规模分类模型在多种数据集上的预训练模型,方便用户基于这些预训练模型进行下游任务的fine-tuning。
Expand All @@ -73,7 +77,7 @@ softmax的计算公示如下图所示:

| 模型 | 描述 |
| :--------------- | :------------- |
| [resnet50_distarcface_ms1m_v2](http://icm.baidu-int.com/user-center/account) | 该模型使用ResNet50网络训练,数据集为MS1M_v2,训练阶段使用的loss_type为'dist_arcface',预训练模型在lfw验证集上的验证精度为0.99817。 |
| [resnet50_distarcface_ms1m_arcface](https://plsc.bj.bcebos.com/pretrained_model/resnet50_distarcface_ms1mv2.tar.gz) | 该模型使用ResNet50网络训练,数据集为MS1M-ArcFace,训练阶段使用的loss_type为'dist_arcface',预训练模型在lfw验证集上的验证精度为0.99817。 |

### 训练性能

Expand Down Expand Up @@ -109,3 +113,10 @@ softmax的计算公示如下图所示:

* [分布式参数转换](docs/distributed_params.md)
* [Base64格式图像预处理](docs/base64_preprocessor.md)

* 2019.12.23
**`0.1.0`**
*PaddlePaddle大规模分类库(PLSC)发布,内建ResNet50、ResNet101和ResNet152三种模型,并支持自定义模型;
* 单机8张V100 GPU配置下,ResNet50模型一百万类别训练速度2,122.56 images/s, 并支持多机分布式训练;
* 发布模型在线预测库;
* 发布基于ResNet50网络和MS1M-ArcFace数据集的预训练模型。
4 changes: 1 addition & 3 deletions plsc/models/resnet.py
Original file line number Diff line number Diff line change
Expand Up @@ -41,13 +41,11 @@ def build_network(self,

if layers == 50:
depth = [3, 4, 14, 3]
num_filters = [64, 128, 256, 512]
elif layers == 101:
depth = [3, 4, 23, 3]
num_filters = [256, 512, 1024, 2048]
elif layers == 152:
depth = [3, 8, 36, 3]
num_filters = [256, 512, 1024, 2048]
num_filters = [64, 128, 256, 512]

conv = self.conv_bn_layer(
input=input, num_filters=64, filter_size=3, stride=1,
Expand Down

0 comments on commit 8abf442

Please sign in to comment.