本文介绍Recurrent Attention Model (RAM)的复现情况。RAM通过循环神经网络处理图像子区域信息，自主选择子区域，降低复杂度。其含glimpse sensor等五部分结构。复现用MNIST数据集，验证误差1.18%（290epoch），测试误差1.17%~1.28%，还提及复现中rsample方法和索引操作的问题及解决，提升了训练速度。
☞☞☞AI 智能聊天, 问答助手, AI 智能搜索, 免费无限量使用 DeepSeek R1 模型☜☜☜
一、论文简介
1.1 简介
Recurrent Attention Model (RAM)，它能顺序处理输入信息，在每个时间步关注图像内不同的子区域，然后增量式的结合来自这些固定位置的信息，并建立图像的动态内部表示。
RAM的优点在于能自主选择图像的子区域进行处理，而不像传统的卷积模型一样复杂度随着输入图像像素变大而线性增长。
论文地址
1.2 网络结构
本文将注意力问题建模为目标导向的agent与视觉环境交互的序列决策过程，agent的核心是一个循环神经网络，它在每一时间步处理来自sensor收集的子图信息，并随着时间推移集成图像信息，并选择如何行动和部署下一个时间步的sensor。
RAM模型结构如上图所示，其中包含如下五个部分：
-
glimpse sensor:glimpse sensor受到视网膜注意力机制的启发，即人往往能清晰的看见所关注对象的细节（内容少，高分辨率），同时保留对背景的模糊感受（内容多，低分辨率）。于是设计的glimpse sensor能从图像 x中提取漏斗状的一瞥（glimpse）phi，sensor首先编码靠近位置l的一块高像素的小区域，然后再渐进的从l附近取更大且像素更低的子区域（所有的截取的子区域需要缩放到同一大小，所以大图像素低），从而得到原始图像 x的压缩表示；
下面第一张图是截取位置l附近不同尺度的区域，然后第二章是将他们缩放到同一尺度，使得细节部分有高分辨率，背景低分辨率。
glimpse network: 该网络将sensor得到的压缩表示"what" (phi)和位置信息"where" (l)结合起来，得到这一瞥的特征向量g_t；
core network: 核心网络是个循环神经网络，该网络维持一个内部状态 h_t ，代表从过去观测历史中提取的整合信息。它通过状态向量 h_t 编码angent对环境的知识，并且在每个时间步 t都会更新。时间步t时的输入为上一个时刻glimpse向量g_(t-1)和状态向量h_(t-1)；
location network:位置网络，使用rnn状态向量h_t，在时间步t时产生shape为[bsz,2]的位置坐标l_t，再同输入图像 x送入glimpse得到输入向量g_(t+1),同状态向量h_t作为t+1时刻rnn的输入；
action network: 在固定数的时间步之后，使用rnn的内部状态‘h_t’生成最终的分类输出 y。
总的来说，RAM是围绕rnn展开的，输入是glimpse向量和t时刻状态向量，输出是t+1时刻状态向量，代表集成的图像信息。利用状态向量输入两个子网络location和action 可以得到两个输出：l_t和a_t，l_t用于指导sensor截取子图并编码为输入向量，a_t用来完成分类任务。
二、复现结果
2.1 实验结果
本项目使用28x28的MNIST数据集来复现，RAM模型包含6个glimpses，patch_size为8x8，缩放因子scale为1，论文中指标为：
I-Shop购物系统
下载
部分功能简介：商品收藏夹功能热门商品最新商品分级价格功能自选风格打印结算页面内部短信箱商品评论增加上一商品，下一商品功能增强商家提示功能友情链接用户在线统计用户来访统计用户来访信息用户积分功能广告设置用户组分类邮件系统后台实现更新用户数据系统图片设置模板管理CSS风格管理申诉内容过滤功能用户注册过滤特征字符IP库管理及来访限制及管理压缩，恢复，备份数据库功能上传文件管理商品类别管理商品添加/修改/
本项目的验证误差为1.18%（290epoch），原文和本项目在MNIST测试集上的误差为：
| Task | Paper | Me |
|---|---|---|
| 28x28 MNIST | 1.29% | 1.17%~1.28% |
本项目的模型权重ram_6_8x8_1_model_best.pdparams（aistudio上zip里面有）已经上传到百度网盘:链接 ,提取码：v6d3
2.2 实验环境以及超参
| NO. | Paddle Version | Memory | Card | Batch Size | Learning Rate | LR Factor | LR Patience | Epoch | Training time | val err | test err |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 01 | 2.1.2 | 16G | V100*1 | 128 | 3e-4 | 0.8 | 20 | 290 | ~2h | 1.15% | 1.17% |
| 02 | 2.1.2 | 16G | V100*1 | 128 | 3e-4 | 0.8 | 20 | 315 | ~3h | 1.033 % | 1.28% |
注：
第一次是先用factor=0.1,patience=20训练了200轮，发现142轮达到最优，未达到指定精度，且后面学习率过小为0了。于是从142轮开始恢复训练，初始学习率仍为3e-4，然后factor=0.8，patience=10, 继续训练到290轮，详细见logs里RAM_local290.log日志。且该指标是在本地3060环境达到精度，3060一轮约45s，v100一轮约30s。
第二次是在aistudio上，初始学习率3e-4，然后factor=0.8，patience=10,训练到200轮，发现第192轮best，test acc为1.68，然后恢复训练，到315轮时验证误差最小为1.033%，于是停止训练，评估得到1.28%
三、准备工作
In [1]
# 解压代码!unzip RAM.zip
Archive: RAM.zip replace RAM/align.py? [y]es, [n]o, [A]ll, [N]one, [r]ename: ^C
In [ ]
# 进入目录，并安装依赖%cd RAM/ !pip install tensorboard_logger
/home/aistudio/RAM Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple Requirement already satisfied: tensorboard_logger in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (0.1.0) Requirement already satisfied: six in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from tensorboard_logger) (1.15.0) Requirement already satisfied: scipy>=0.19.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from tensorboard_logger) (1.6.3) Requirement already satisfied: protobuf in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from tensorboard_logger) (3.14.0) Requirement already satisfied: numpy in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from tensorboard_logger) (1.20.3) Requirement already satisfied: pillow>=4.1.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from tensorboard_logger) (7.1.2)
目录结构
. ├── README.md ├── align.py # 转换权重├── ckpt # 权重│ └── ram_6_8x8_1_model_best.pdparams ├── config.py # 配置文件├── data # 数据│ ├── MNIST ├── data_loader.py # 加载数据├── logs # 日志├── main.py # 主函数├── model.py # RAM主体模型├── modules.py # RAM5个部分├── plot_glimpses.py # 画图├── plots # 图片├── requirements.txt ├── trainer.py # 训练、评估函数└── utils.py # 工具
四、模型训练
In [ ]
!python main.py
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/__init__.py:107: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working from collections import MutableMapping /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/rcsetup.py:20: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working from collections import Iterable, Mapping /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/colors.py:53: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working from collections import Sized W1123 12:01:46.425436 726 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.0, Runtime API Version: 10.1 W1123 12:01:46.429379 726 device_context.cc:465] device: 0, cuDNN Version: 7.6. [*] Model Checkpoint Dir: ./ckpt [*] Param Path: ./ckpt/ram_6_8x8_1_params.json 2021-11-23 12:01:49,497 | RAM: [*] Train on 54000 samples, validate on 6000 samples INFO:RAM: [*] Train on 54000 samples, validate on 6000 samples 2021-11-23 12:01:49,497 | RAM: Epoch: 1/200 - LR: 0.000300 INFO:RAM: Epoch: 1/200 - LR: 0.000300 0%| | 0/54000 [00:00
五、模型评估
In [ ]
## aistudio上训练的，315轮，验证误差1.033，测试误差1.28%!python main.py --is_train=False --best True --ckpt_dir ckpt_aistudio
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/__init__.py:107: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working from collections import MutableMapping /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/rcsetup.py:20: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working from collections import Iterable, Mapping /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/colors.py:53: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working from collections import Sized W1124124 09:46:43.853452 793 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1 W1124124 09:46:43.857133 793 device_context.cc:465] device: 0, cuDNN Version: 7.6. 2021-11-24 09:46:48,040 | RAM: [*] Loading model from ckpt_aistudio INFO:RAM:[*] Loading model from ckpt_aistudio 2021-11-24 09:46:48,050 | RAM: [*] Loaded ram_6_8x8_1_model_best.pdparams checkpoint @ epoch 315 with best valid acc of 98.917 INFO:RAM:[*] Loaded ram_6_8x8_1_model_best.pdparams checkpoint @ epoch 315 with best valid acc of 98.917 /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/tensor/creation.py:130: DeprecationWarning: `np.object` is a deprecated alias for the builtin `object`. To silence this warning, use `object` by itself. Doing this will not modify any behavior and is safe. Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations if data.dtype == np.object: 2021-11-24 09:46:53,835 | RAM: [*] Test Acc: 9872.0/10000 (98.72% - 1.28%) INFO:RAM:[*] Test Acc: 9872.0/10000 (98.72% - 1.28%)
In [ ]
## 本地3060训练的，见日志logs/RAM_local290.log，290轮，验证误差1.15，测试误差1.17%!python main.py --is_train=False --best True
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/__init__.py:107: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working from collections import MutableMapping /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/rcsetup.py:20: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working from collections import Iterable, Mapping /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/colors.py:53: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working from collections import Sized INFO:matplotlib.font_manager:font search path ['/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/mpl-data/fonts/ttf', '/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/mpl-data/fonts/afm', '/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/mpl-data/fonts/pdfcorefonts'] INFO:matplotlib.font_manager:generated new fontManager Cache file /home/aistudio/.cache/paddle/dataset/mnist/t10k-images-idx3-ubyte.gz not found, downloading https://dataset.bj.bcebos.com/mnist/t10k-images-idx3-ubyte.gz Begin to download item 403/403 [============================>.] - ETA: 0s - 389us/it Download finished Cache file /home/aistudio/.cache/paddle/dataset/mnist/t10k-labels-idx1-ubyte.gz not found, downloading https://dataset.bj.bcebos.com/mnist/t10k-labels-idx1-ubyte.gz Begin to download item 2/2 [===========================>..] - ETA: 0s - 642us/it Download finished W1124124 09:37:17.144690 311 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1 W1124124 09:37:17.149226 311 device_context.cc:465] device: 0, cuDNN Version: 7.6. 2021-11-24 09:37:22,326 | RAM: [*] Loading model from ./ckpt INFO:RAM:[*] Loading model from ./ckpt 2021-11-24 09:37:22,336 | RAM: [*] Loaded ram_6_8x8_1_model_best.pdparams checkpoint @ epoch 290 with best valid acc of 98.917 INFO:RAM:[*] Loaded ram_6_8x8_1_model_best.pdparams checkpoint @ epoch 290 with best valid acc of 98.917 /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/tensor/creation.py:130: DeprecationWarning: `np.object` is a deprecated alias for the builtin `object`. To silence this warning, use `object` by itself. Doing this will not modify any behavior and is safe. Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations if data.dtype == np.object: 2021-11-24 09:37:28,007 | RAM: [*] Test Acc: 9883.0/10000 (98.83% - 1.17%) INFO:RAM:[*] Test Acc: 9883.0/10000 (98.83% - 1.17%)
六、总结
这里就说下复现遇到的小坑：
1.rsample（location网络）： paddle.distribution.Normal，没有torch.distribution.Normal().rsample方法， 参考torch源码实现后，在对齐精度时发现不能完全对齐，该操作有随机性，差0.3%左右，不过影响不大；
def rsample(loc, scale): shape = loc.shape normal_ = paddle.nn.initializer.Normal() eps = paddle.empty(shape, dtype=loc.dtype) normal_(eps) return loc + eps * scale
2.索引：
在glimpse的retina的extract_patch方法内，根据输入的位置信息lt[bsz,2]，对图片进行采样（8*8patch）。
def extract_patch(self, x, l, size):
... patch = []
for i in range(B):
subset=x[i, :, start[i, 1] : end[i, 1], start[i, 0] : end[i, 0]]
patch.append(subset)
return paddle.to_tensor(np.stack(patch))- 我在改完代码后发现paddle的训练速度250steps/s,而torch为900steps/s (本地3060)
- 然后逐个模块定位，最后发现glimpse里面这个采样的操作特别慢，128的bsz，for循环对每张图片切片得到8*8的patch，paddle需要0.07，torch需要0.003s左右，差了二三十倍，整体训练差了四五倍。
- 我先试试基础api，select_index能对所有的图片截取相同的某几行或几列，不能达到取不同块的目的。
- 考虑到索引操作都不需要梯度了，试了下numpy，发现速度较快，128bsz，迭代10次，共1280次索引，对比如下：
| slice x1280 | Paddle | Torch | Numpy |
|---|---|---|---|
| Time | 0.727s | 0.0219s | 0.00099s |
- 最后改完后paddle速度达到1200，比参考代码快了1.33倍（aistudio上1800steps/s）：
总之，在遇到精度、速度差距大时，从上到下一层层慢慢debug就行啦~
