https://github.com/rmsnow/amnet-rumor
https://github.com/rmsnow/amnet-rumor
Last synced: over 1 year ago
JSON representation
- Host: GitHub
- URL: https://github.com/rmsnow/amnet-rumor
- Owner: RMSnow
- License: mit
- Created: 2018-10-23T11:56:55.000Z (over 7 years ago)
- Default Branch: master
- Last Pushed: 2018-10-31T06:23:30.000Z (over 7 years ago)
- Last Synced: 2025-01-30T10:30:45.040Z (over 1 year ago)
- Language: Jupyter Notebook
- Size: 5.31 MB
- Stars: 1
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# AMNet: Memorability Estimation with Attention
A PyTorch implementation of our paper [AMNet: Memorability Estimation with Attention](https://arxiv.org/abs/1804.03115)
by Jiri Fajtl, Vasileios Argyriou, Dorothy Monekosso and Paolo Remagnino. This paper will be presented
at [CVPR 2018](http://cvpr2018.thecvf.com/).
==> [AMNet](https://github.com/ok1zjf/AMNet)
## 实验流程
1. Train
开始之前：
- `amnet.py`：（1）修改实验序号（2）修改AMNet-Train-Output文件夹
- `config.py`：修改GPU设备
- 服务器：（1）同步代码（2）创建AMNet-Train-Output文件夹
- 修改Cmd命令的参数，如`--lstm-steps`等
训练完成：
- 保存可视化结果
- 保存log日志文件
- 保存模型文件`.pkl`、训练日志`.csv`
2. Predict
开始之前：
- 服务器：创建AMNet-Predict文件夹
- 修改Cmd命令的参数，如`--lstm-steps`等
预测完成：
- 保存部分att_maps图
- 保存预测结果：`mem`文件
`scp qipeng@10.25.0.232:/home/qipeng/PicMemorability/AMNet-Rumor-Baseline/*expt4* /Users/snow/snow_学习/4-研究生/实验室/AMNet/AMNet/_expt`
3. Eval
- 记录实验日志：
- Accuracy：sklearn分类报告
- AUC
- `README.md`：可视化图片添加
- 分析`eval-expt*.csv`，保存相应的图片
4. 每日结束
- git同步代码
## Path & Cmd
### 服务器文件路径
`/media/Data/qipeng/modified_complete_images`
`/home/qipeng/PicMemorability/AMNet-Rumor`
### 本机路径
`/Users/snow/snow_学习/4-研究生/实验室/AMNet/AMNet/_expt`
### Train Cmd
`python3.5 main.py --train-batch-size 256 --test-batch-size 256 --cnn ResNet50FC --dataset lamem --dataset-root /media/Data/qipeng/modified_complete_images/AMNet-Rumor/lamem/ --train-split train_0 --val-split val_0 --lstm-steps 6`
### Predict Cmd - Rumor / Nonrumor
`python3.5 main.py --cnn ResNet50FC --model-weights /media/Data/qipeng/modified_complete_images/AMNet-Rumor/AMNet-Train-Output/expt5/lamem_ResNet50FC_lstm5_train_0/weights_35.pkl --eval-images /media/Data/qipeng/modified_complete_images/AMNet-Rumor/AMNet-Predict/images/rumor --csv-out memorabilities-expt5-rumor.txt --att-maps-out /media/Data/qipeng/modified_complete_images/AMNet-Rumor/AMNet-Predict/expt5/att_maps/rumor --lstm-steps 5 --gpu 1`
`python3.5 main.py --cnn ResNet50FC --model-weights /media/Data/qipeng/modified_complete_images/AMNet-Rumor/AMNet-Train-Output/expt5/lamem_ResNet50FC_lstm5_train_0/weights_35.pkl --eval-images /media/Data/qipeng/modified_complete_images/AMNet-Rumor/AMNet-Predict/images/nonrumor --csv-out memorabilities-expt5-nonrumor.txt --att-maps-out /media/Data/qipeng/modified_complete_images/AMNet-Rumor/AMNet-Predict/expt5/att_maps/nonrumor --lstm-steps 5 --gpu 3`
`python3.5 main.py --cnn ResNet50FC --model-weights /media/Data/qipeng/modified_complete_images/AMNet-Rumor/AMNet-Train-Output/expt6/lamem_ResNet50FC_lstm6_train_0/weights_54.pkl --eval-images /media/Data/qipeng/modified_complete_images/AMNet-Rumor/AMNet-Predict/images/rumor --csv-out memorabilities-expt6-rumor.txt --att-maps-out /media/Data/qipeng/modified_complete_images/AMNet-Rumor/AMNet-Predict/expt6/att_maps/rumor --lstm-steps 6 --gpu 1`
`python3.5 main.py --cnn ResNet50FC --model-weights /media/Data/qipeng/modified_complete_images/AMNet-Rumor/AMNet-Train-Output/expt6/lamem_ResNet50FC_lstm6_train_0/weights_54.pkl --eval-images /media/Data/qipeng/modified_complete_images/AMNet-Rumor/AMNet-Predict/images/nonrumor --csv-out memorabilities-expt6-nonrumor.txt --att-maps-out /media/Data/qipeng/modified_complete_images/AMNet-Rumor/AMNet-Predict/expt6/att_maps/nonrumor --lstm-steps 6 --gpu 3`
## Expt Log
| **实验序号** | 日期 | 主题 | 训练数据 | 超参数 | Accuracy | AUC | 备注 |
| :----------: | --------- | --------------- | :------------------- | --------------------------------------------------------- | -------- | ------- | --------------------------- |
| **0** | 1016-1017 | baseline | shuffle（不严格1:1） | lstm_steps: 3
batch_size: 222
epoch: 54 | 0.75963 | 0.84084 | |
| **1** | 1022 | baseline-可视化 | **train_0, val_0** | lstm_steps: 3
**batch_size: 256**
epoch: 54 | 0.75352 | 0.83668 | 召回率高 |
| **2** | 1023 | lstm_steps=1 | train_0, val_0 | **lstm_steps: 1**
batch_size: 256
epoch: 54 | 0.75537 | 0.84413 | 准确率高 |
| **3** | 1023 | lstm_steps=2 | train_0, val_0 | **lstm_steps: 2**
batch_size: 256
epoch: 54 | 0.755 | 0.83577 | 观察epoch图，**可能过拟合** |
| **4** | 1023 | lstm_steps=4 | train_0, val_0 | **lstm_steps: 4**
batch_size: 256
epoch: 54 | 0.73389 | 0.82929 | |
| **5** | 1024 | lstm_steps=5 | train_0, val_0 | **lstm_steps: 5**
batch_size: 256
**epoch: 35** | 0.75259 | 0.842 | |
| **6** | 1024 | lstm_steps=6 | train_0, val_0 | **lstm_steps: 6**
batch_size: 256
epoch: 54 | 0.764 | 0.845 | |
| **2.2** | 1024 | weights_35 | | epoch: 35 | 0.76111 | 0.838 | |
| **3.2** | 1024 | weights_35 | | epoch: 35 | 0.75685 | 0.843 | 减少epoch有效 |
| **1.2** | 1024 | weights_35 | | epoch: 35 | 0.74167 | 0.838 | |
| **4.2** | 1024 | weights_35 | | epoch: 35 | 0.74815 | 0.840 | 减少epoch有效 |
### 部分规律
| | AUC | 0-precision | 1-precision | 0-recall | 1-recall | 0-f1 | 1-f1 |
| :----------------: | :-------: | :---------: | :---------: | :------: | :------: | :------: | :------: |
| **Expt2: LSTM1** | **0.844** | **0.72** | **0.80** | **0.83** | **0.68** | **0.77** | **0.74** |
| Expt2.2: LSTM1 | 0.838 | 0.77 | 0.75 | 0.74 | 0.78 | 0.76 | 0.76 |
| Expt3: LSTM2 | 0.836 | 0.73 | 0.78 | 0.80 | 0.71 | 0.77 | 0.74 |
| **Expt3.2: LSTM2** | **0.843** | **0.73** | **0.79** | **0.82** | **0.69** | **0.77** | **0.74** |
| **Expt1: LSTM3** | **0.837** | **0.80** | **0.72** | **0.68** | **0.83** | **0.73** | **0.77** |
| Expt1.2: LSTM3 | 0.838 | 0.82 | 0.69 | 0.62 | 0.87 | 0.71 | 0.77 |
| Expt4: LSTM4 | 0.829 | 0.69 | 0.81 | 0.86 | 0.61 | 0.76 | 0.69 |
| **Expt4.2: LSTM4** | **0.840** | **0.71** | **0.80** | **0.83** | **0.67** | **0.77** | **0.73** |
| **Expt5: LSTM5** | **0.842** | **0.72** | **0.79** | **0.82** | **0.68** | **0.77** | **0.73** |
| **Expt6: LSTM6** | **0.845** | **0.77** | **0.76** | **0.75** | **0.78** | **0.76** | **0.77** |
比较Expt2，Expt3.2，Expt1，Expt4.2，Expt5，Expt6的预测值分布：
### Expt 0
**Accuracy**
```
precision recall f1-score support
0 0.76 0.75 0.76 2700
1 0.76 0.77 0.76 2700
avg / total 0.76 0.76 0.76 5400
```
### Expt 1: LSTM3
**Accuracy**
```
precision recall f1-score support
0 0.80 0.68 0.73 2700
1 0.72 0.83 0.77 2700
avg / total 0.76 0.75 0.75 5400
```
### Expt 2: LSTM1
**Accuracy**
```
precision recall f1-score support
0 0.72 0.83 0.77 2700
1 0.80 0.68 0.74 2700
avg / total 0.76 0.76 0.75 5400
```
### Expt 3: LSTM2
**Accuracy**
```
precision recall f1-score support
0 0.73 0.80 0.77 2700
1 0.78 0.71 0.74 2700
avg / total 0.76 0.76 0.75 5400
```
### Expt4: LSTM4
```
precision recall f1-score support
0 0.69 0.86 0.76 2700
1 0.81 0.61 0.69 2700
avg / total 0.75 0.73 0.73 5400
```
## TODO
### Debug
- 预测时有bug
### Learn
- pytorch
- 多块gpu尝试
- 看懂模型代码：如框架使用，nn.LSTM()，forward，参数设置，维度等
- Linux命令
- nohup, >>, &
- kill, ps
- top
- ls -lR|grep "^-"|wc -l