From ffe11953ee4bab77a71f3c261ebb7680398be471 Mon Sep 17 00:00:00 2001 From: huangjun12 <2399845970@qq.com> Date: Thu, 15 Apr 2021 10:45:40 +0800 Subject: [PATCH] update DataParallel in slowfast (#5298) * update DP in slowfast and refine nextvlad doc * refine doc --- PaddleCV/video/models/nextvlad/README.md | 2 +- dygraph/slowfast/README.md | 2 +- dygraph/slowfast/eval.py | 3 ++- dygraph/slowfast/predict.py | 3 ++- dygraph/slowfast/train.py | 4 ++-- 5 files changed, 8 insertions(+), 6 deletions(-) diff --git a/PaddleCV/video/models/nextvlad/README.md b/PaddleCV/video/models/nextvlad/README.md index 8ed283ce..c9106286 100644 --- a/PaddleCV/video/models/nextvlad/README.md +++ b/PaddleCV/video/models/nextvlad/README.md @@ -12,7 +12,7 @@ ## 算法介绍 -NeXtVLAD模型是第二届Youtube-8M视频理解竞赛中效果最好的单模型，在参数量小于80M的情况下，能得到高于0.87的GAP指标。该模型提供了一种将桢级别的视频特征转化并压缩成特征向量，以适用于大尺寸视频文件的分类的方法。其基本出发点是在NetVLAD模型的基础上，将高维度的特征先进行分组，通过引入attention机制聚合提取时间维度的信息，这样既可以获得较高的准确率，又可以使用更少的参数量。详细内容请参考[NeXtVLAD: An Efficient Neural Network to Aggregate Frame-level Features for Large-scale Video Classification](https://arxiv.org/abs/1811.05014)。 +NeXtVLAD模型是第二届Youtube-8M视频理解竞赛中效果最好的单模型，在参数量小于80M的情况下，能得到高于0.87的GAP指标。该模型提供了一种将帧级别的视频特征转化并压缩成特征向量，以适用于大尺寸视频文件的分类的方法。其基本出发点是在NetVLAD模型的基础上，将高维度的特征先进行分组，通过引入attention机制聚合提取时间维度的信息，这样既可以获得较高的准确率，又可以使用更少的参数量。详细内容请参考[NeXtVLAD: An Efficient Neural Network to Aggregate Frame-level Features for Large-scale Video Classification](https://arxiv.org/abs/1811.05014)。 这里实现了论文中的单模型结构，使用2nd-Youtube-8M的train数据集作为训练集，在val数据集上做测试。 diff --git a/dygraph/slowfast/README.md b/dygraph/slowfast/README.md index b7ab9d4b..c09b4c13 100644 --- a/dygraph/slowfast/README.md +++ b/dygraph/slowfast/README.md @@ -68,7 +68,7 @@ SlowFast Overview ## 数据准备 -SlowFast模型的训练数据采用Kinetics400数据集，数据下载及准备请参考[数据说明](../PaddleCV/video/data/dataset/README.md) +SlowFast模型的训练数据采用Kinetics400数据集，数据下载及准备请参考[数据说明](../PaddleCV/video/data/dataset/README.md)。 ## 模型训练 diff --git a/dygraph/slowfast/eval.py b/dygraph/slowfast/eval.py index 1c229d6b..bc686010 100644 --- a/dygraph/slowfast/eval.py +++ b/dygraph/slowfast/eval.py @@ -101,7 +101,8 @@ def test_slowfast(args): if args.use_data_parallel: strategy = fluid.dygraph.parallel.prepare_context() - slowfast = fluid.dygraph.parallel.DataParallel(slowfast, strategy) + slowfast = fluid.dygraph.parallel.DataParallel( + slowfast, strategy, find_unused_parameters=False) #create reader test_data = KineticsDataset(mode="test", cfg=test_config) diff --git a/dygraph/slowfast/predict.py b/dygraph/slowfast/predict.py index 1a1c8b84..b9ac2b1d 100644 --- a/dygraph/slowfast/predict.py +++ b/dygraph/slowfast/predict.py @@ -110,7 +110,8 @@ def infer_slowfast(args): if args.use_data_parallel: strategy = fluid.dygraph.parallel.prepare_context() - slowfast = fluid.dygraph.parallel.DataParallel(slowfast, strategy) + slowfast = fluid.dygraph.parallel.DataParallel( + slowfast, strategy, find_unused_parameters=False) #create reader infer_data = KineticsDataset(mode="infer", cfg=infer_config) diff --git a/dygraph/slowfast/train.py b/dygraph/slowfast/train.py index c4e0bdd2..dea55318 100644 --- a/dygraph/slowfast/train.py +++ b/dygraph/slowfast/train.py @@ -277,8 +277,8 @@ def train(args): video_model = SlowFast(cfg=train_config, num_classes=400) if args.use_data_parallel: - video_model = fluid.dygraph.parallel.DataParallel(video_model, - strategy) + video_model = fluid.dygraph.parallel.DataParallel( + video_model, strategy, find_unused_parameters=False) bs_denominator = 1 if args.use_gpu: -- GitLab