使用 paddle 复现论文 FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence. 官方开源 pytorch 代码 在这.
详见 FixMatch-Paddle/pipeline/
- Step1 模型结构对齐
- Step2 评估指标对齐
- Step3 损失函数和反向传播对齐
- Step4 训练对齐
- python 3.7
- paddle 2.1
- visualdl
- numpy
- tqdm
模型和训练日志在 Release 中，也可到百度网盘提取，链接：https://pan.baidu.com/s/14j8AQnvLgBFgCfZzHbfpxQ 提取码：kimn
| #Labels | 40 | 250 | 4000 |
|---|---|---|---|
| Paper (RA) | 86.19 ± 3.37 | 94.93 ± 0.65 | 95.74 ± 0.05 |
| pytorch code | 93.60 | 95.22 | 95.77 |
| paddle code | 93.59 (epoch 190) | 95.20 (epoch 255) | 95.83 (epoch 156) |
| model_best | [email protected] | [email protected] | [email protected] |
cifar10 数据集: cifar-10-python.tar.gz
-
单卡：
- 命令行1：使用 4000（40、250） 个有标签样本，从零开始训练:
python train.py --dataset cifar10 \ --num-labeled 4000 \ --arch wideresnet --batch-size 64 --lr 0.03 \ --expand-labels --seed 5 \ --log-out results/[email protected]/logs \ --out results/[email protected] \ --data-file path/to/cifar10- 命令行2：从预训练模型继续训练，添加命令行参数
--resume path/to/latest-ckpt
-
单机 4 卡：
python -m paddle.distributed.launch --gpus '0,1,2,3' train.py \ --dataset cifar10 \ --num-labeled 4000 \ --arch wideresnet --batch-size 64 --lr 0.095 \ --expand-labels --seed 5 \ --local_rank 0 \ --log-out results/[email protected]/logs \ --out results/[email protected] \ --data-file path/to/cifar10 \ --resume path/to/latest-ckpt
- 方式1: 加载模型中保存的 dict:
import paddle
state=paddle.load("path/to/moddel_best.pdparams")
print(state.keys()) # 查看model中保存的dict
print('epoch:', state['epoch']) # 查看是哪个epoch保存的最佳模型
print(f"best acc: {state['best_acc']}") # 查看模型的测试集精度- 方式2: 使用命令行
python test.py --dataset cifar10 \
--arch wideresnet --batch-size 64 \
--expand-labels --seed 5 \
--data-file path/to/cifar10 \
--model-best path/to/best-ckpt
输出如下：
Test Iter: 1/ 22. Data: 0.895s. Batch: 0.942s. Loss: 0.2222. top1: 93.97. top5: 99.78.
Test Iter: 2/ 22. Data: 0.448s. Batch: 0.485s. Loss: 0.1925. top1: 94.75. top5: 99.67.
Test Iter: 3/ 22. Data: 0.299s. Batch: 0.332s. Loss: 0.2224. top1: 94.64. top5: 99.70.
Test Iter: 4/ 22. Data: 0.224s. Batch: 0.260s. Loss: 0.2255. top1: 94.70. top5: 99.78.
Test Iter: 5/ 22. Data: 0.183s. Batch: 0.216s. Loss: 0.2200. top1: 94.69. top5: 99.82.
Test Iter: 6/ 22. Data: 0.173s. Batch: 0.207s. Loss: 0.2203. top1: 94.79. top5: 99.81.
Test Iter: 7/ 22. Data: 0.148s. Batch: 0.181s. Loss: 0.2156. top1: 94.80. top5: 99.81.
Test Iter: 8/ 22. Data: 0.130s. Batch: 0.161s. Loss: 0.2135. top1: 94.81. top5: 99.75.
Test Iter: 9/ 22. Data: 0.123s. Batch: 0.154s. Loss: 0.2136. top1: 94.79. top5: 99.78.
Test Iter: 10/ 22. Data: 0.123s. Batch: 0.153s. Loss: 0.2218. top1: 94.73. top5: 99.75.
Test Iter: 11/ 22. Data: 0.111s. Batch: 0.142s. Loss: 0.2158. top1: 94.89. top5: 99.78.
Test Iter: 12/ 22. Data: 0.102s. Batch: 0.132s. Loss: 0.2067. top1: 95.07. top5: 99.80.
Test Iter: 13/ 22. Data: 0.100s. Batch: 0.129s. Loss: 0.2059. top1: 95.09. top5: 99.81.
Test Iter: 14/ 22. Data: 0.101s. Batch: 0.130s. Loss: 0.2047. top1: 95.03. top5: 99.81.
Test Iter: 15/ 22. Data: 0.094s. Batch: 0.124s. Loss: 0.2059. top1: 94.93. top5: 99.81.
Test Iter: 16/ 22. Data: 0.088s. Batch: 0.118s. Loss: 0.2019. top1: 95.03. top5: 99.82.
Test Iter: 17/ 22. Data: 0.087s. Batch: 0.116s. Loss: 0.2062. top1: 94.96. top5: 99.83.
Test Iter: 18/ 22. Data: 0.089s. Batch: 0.117s. Loss: 0.2100. top1: 94.90. top5: 99.84.
Test Iter: 19/ 22. Data: 0.084s. Batch: 0.112s. Loss: 0.2080. top1: 94.94. top5: 99.84.
Test Iter: 20/ 22. Data: 0.080s. Batch: 0.108s. Loss: 0.2073. top1: 94.96. top5: 99.84.
Test Iter: 21/ 22. Data: 0.078s. Batch: 0.106s. Loss: 0.2058. top1: 95.00. top5: 99.85.
Test Iter: 22/ 22. Data: 0.080s. Batch: 0.108s. Loss: 0.2073. top1: 94.99. top5: 99.85.
top-1 acc: 94.99
top-5 acc: 99.85
| 信息 | 说明 |
|---|---|
| 发布者 | 刘辰、吴海涛、石华榜、杨瑞智、许观 |
| 时间 | 2021.10.25 |
| 框架版本 | paddlepaddle==2.1.2 |
| 应用场景 | 半监督图像分类 |
| 支持硬件 | GPU × 4 |
| 预训练模型下载 | 链接：https://pan.baidu.com/s/14j8AQnvLgBFgCfZzHbfpxQ 提取码：kimn |
| AI Studio 地址 | FixMatch-Paddle |
@article{sohn2020fixmatch,
title={FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence},
author={Kihyuk Sohn and David Berthelot and Chun-Liang Li and Zizhao Zhang and Nicholas Carlini and Ekin D. Cubuk and Alex Kurakin and Han Zhang and Colin Raffel},
journal={arXiv preprint arXiv:2001.07685},
year={2020},
}