Skip to content

Latest commit

 

History

History
84 lines (59 loc) · 4.13 KB

attention_lstm.md

File metadata and controls

84 lines (59 loc) · 4.13 KB
Raw

简体中文 | English

AttentionLSTM

content

Introduction

Recurrent Neural Networks (RNN) are often used in the processing of sequence data, which can model the sequence information of multiple consecutive frames of video, and are commonly used methods in the field of video classification. This model uses a two-way long and short-term memory network (LSTM) to encode all the frame features of the video in sequence. Unlike the traditional method that directly uses the output of the last moment of LSTM, this model adds an Attention layer, and the hidden state output at each moment has an adaptive weight, and then linearly weights the final feature vector. The reference paper implements a two-layer LSTM structure, while this model implements a two-way LSTM with Attention.

The Attention layer can refer to the paper AttentionCluster

Data

PaddleVide provides training and testing scripts on the Youtube-8M dataset. Youtube-8M data download and preparation please refer to YouTube-8M data preparation

Train

Youtube-8M data set training

Start training

  • The Youtube-8M data set uses 8 cards for training. In the feature format, video and audio features will be used as input. The training start command of the data is as follows

    python3.7 -B -m paddle.distributed.launch --gpus="0,1,2,3,4,5,6,7" --log_dir=log_attetion_lstm main.py --validate -c configs/recognition/attention_lstm/attention_lstm_youtube8m.yaml

Test

The command is as follows:

python3.7 -B -m paddle.distributed.launch --gpus="0,1,2,3,4,5,6,7" --log_dir=log_attetion_lstm main.py --test -c configs/recognition/attention_lstm/attention_lstm_youtube8m.yaml -w "output/AttentionLSTM/AttentionLSTM_best.pdparams"

When the test configuration uses the following parameters, the test indicators on the validation data set of Youtube-8M are as follows:

Hit@1 PERR GAP checkpoints
89.05 80.49 86.30 AttentionLSTM_yt8.pdparams

Inference

Export inference model

python3.7 tools/export_model.py -c configs/recognition/attention_lstm/attention_lstm_youtube8m.yaml \
                                -p data/AttentionLSTM_yt8.pdparams \
                                -o inference/AttentionLSTM

The above command will generate the model structure file AttentionLSTM.pdmodel and the model weight file AttentionLSTM.pdiparams required for prediction.

For the meaning of each parameter, please refer to Model Reasoning Method

Use prediction engine inference

python3.7 tools/predict.py --input_file data/example.pkl \
                           --config configs/recognition/attention_lstm/attention_lstm_youtube8m.yaml \
                           --model_file inference/AttentionLSTM/AttentionLSTM.pdmodel \
                           --params_file inference/AttentionLSTM/AttentionLSTM.pdiparams \
                           --use_gpu=True \
                           --use_tensorrt=False

An example of the output is as follows:

Current video file: data/example.pkl
         top-1 class: 11
         top-1 score: 0.9841002225875854

It can be seen that using the AttentionLSTM model trained on Youtube-8M to predict data/example.pkl, the output top1 category id is 11, and the confidence is 0.98.

Reference paper