【Hackathon 8th No.23】Improved Training of Wasserstein GANs 论文复现 #1146

XvLingWYY · 2025-04-27T05:20:55Z

PR types

New Features

PR changes

Others

Describe

Add wgangp

paddle-bot · 2025-04-27T05:21:00Z

Thanks for your contribution!

XvLingWYY · 2025-04-27T05:23:39Z

@robinbg

Update wgangp_toy.py

Update functions.py

XvLingWYY · 2025-05-02T07:29:54Z

所有模型和训练日志
通过网盘分享的文件：wgangp
链接: https://pan.baidu.com/s/1geiL5wWudYI44sCMR1WhKQ?pwd=wgan 提取码: wgan

lijialin03

麻烦修改如下问题～
toy: 结果基本符合预期
mnist: 网盘里似乎只有预测结果的图片而没有target，实际运行发现，按照当前参数训练后再评估，mse loss约0.47，生成的图片与target相比都不相同
cifar10：网盘中无预测结果，实际运行时发现在eval的最后一个iter会报错

lijialin03 · 2025-05-08T11:27:43Z

examples/wgangp/conf/wgangp_cifar10.yaml

+    output_dim: 3072
+    label_num: 10
+    use_label: true
+    batch_size: 64


如果训练或者评估用的batch_size与这个不同会报错

好的谢谢老师，已完成此项修改。

lijialin03 · 2025-05-08T11:31:27Z

examples/wgangp/wgamgp_cifar10_model.py

+        if noise is None:
+            noise = paddle.randn([self.batch_size, 128], dtype=paddle.float32)


看了一下似乎只有生成noise这里使用到了batch_size，但实际上这个值可以通过labels.shape[0]得到，如果允许在config文件中定义，反而有可能出现二者不一致的情况，因此这里是否可以直接改为noise = paddle.randn([labels.shape[0], 128], dtype=paddle.float32)

好的谢谢老师，已完成此项修改。

lijialin03 · 2025-05-08T11:34:30Z

examples/wgangp/conf/wgangp_cifar10.yaml

+  inceptionscore:
+    eps: 1e-16
+    splits: 5
+  batch_size: 64


当这个值改为1时会报错
另外,即使不更改这个值，运行到eval最后一个iter时也会报错：

File "/PaddleScience/examples/wgangp/wgamgp_cifar10_model.py", line 166, in forward outputs = outputs * label_scale + label_offset ValueError: (InvalidArgument) Broadcast dimension mismatch. Operands could not be broadcast together with the shape of X = [64, 128, 4, 4] and the shape of Y = [16, 128, 1, 1]. Received [64] in X is not equal to [16] in Y at i:0. [Hint: Expected x_dims_array[i] == y_dims_array[i] || (x_dims_array[i] <= 1 && x_dims_array[i] != 0) || (y_dims_array[i] <= 1 && y_dims_array[i] != 0) == true, but received x_dims_array[i] == y_dims_array[i] || (x_dims_array[i] <= 1 && x_dims_array[i] != 0) || (y_dims_array[i] <= 1 && y_dims_array[i] != 0):0 != true:1.] (at ../paddle/phi/kernels/funcs/common_shape.h:71)

老师修改后未复现出这个错误，应该是没有了。

lijialin03 · 2025-05-08T11:36:18Z

examples/wgangp/conf/wgangp_cifar10.yaml

+DATA:
+  input_keys: [ "labels" ]
+  label_keys: [ "real_data" ]
+  data_path: D:\Data\CIFAR-10\raw\cifar-10-python.tar.gz


改为相对地址如：data_path: "./data/cifar-10-python.tar.gz"

好的谢谢老师，已完成此项修改。

lijialin03 · 2025-05-08T11:37:02Z

examples/wgangp/conf/wgangp_cifar10.yaml

+  pretrained_gen_model_path:
+  pretrained_dis_model_path:


改为：

pretrained_gen_model_path: null pretrained_dis_model_path: null

好的谢谢老师，已完成此项修改。

lijialin03 · 2025-05-08T12:01:17Z

examples/wgangp/wgangp_cifar10.py

+                for i in range(
+                    cfg["EVAL"]["batch_size"]
+                    if cfg["EVAL"]["batch_size"] < cfg.VIS.num
+                    else cfg.VIS.num
+                ):


同理可以把batch_size设置为1并删除这层循环吗

这里不能把batch_size变为1，变为1后生成的图片质量大幅变差，我没有找到原因，所以我认为应该无法把batch_size变为1

lijialin03 · 2025-05-08T12:03:23Z

examples/wgangp/wgangp_mnist_model.py

+    def forward(self, x):
+        y = paddle.randn([self.batch_size, 128])
+        y = self.linear1(y)
+        y = self.relu1(y)
+        y = paddle.reshape(y, [-1, 4 * self.dim, 4, 4])
+        y = self.conv2d_transpose1(y)
+        y = self.relu2(y)
+        y = y[:, :, :7, :7]
+        y = self.conv2d_transpose2(y)
+        y = self.relu3(y)
+        y = self.conv2d_transpose3(y)
+        y = self.sigmoid(y)
+        return paddle.reshape(y, [-1, self.output_dim])


这个forward中没有用到x吗？这里的self.batch_size是否也是x.shape[0]

好的，谢谢老师，已完成此项修改。

lijialin03 · 2025-05-08T12:10:11Z

examples/wgangp/wgangp_toy_model.py

+        self.batch_size = batch_size
+
+    def forward(self, x):
+        y = self.generator(self.batch_size)


这里的self.batch_size同理
tensorflow由于通常是静态图模式，需要placeholder，因此可能需要在模型init阶段就确定batch_size，但是paddle和torch都常用动态图模式，就不需要这么做。而且事实上，一般在模型开发过程中会进行反复调优，batch_size作为一个超参数，也会对模型效果、显存占用等方面造成影响，因此也会进行调整，如果在init阶段就确定它，不仅调试麻烦，也可能会导致无法使用之前训练的模型做pretain

好的，谢谢老师，已完成此项修改。

lijialin03 · 2025-05-08T12:14:32Z

mkdocs.yml

@@ -54,6 +54,7 @@ nav:
        - NeuralOperator: zh/examples/neuraloperator.md
        - Brusselator3D: zh/examples/brusselator3d.md
        - Transformer4SR: zh/examples/transformer4sr.md
+        - Wgan_gp: zh/examples/wgan_gp.md


放在地球科学分类吧
另外代码改动会影响文档内容，可能需要先合入代码再合入文档，可以麻烦把文档相关的文件另提一个pr吗

lijialin03 · 2025-05-09T03:01:54Z

examples/wgangp/wgamgp_cifar10_model.py

@@ -0,0 +1,619 @@
+import math


文件名错了打成m了

好的，谢谢老师，已完成此项修改。

luotao1 · 2025-05-12T08:15:35Z

@XvLingWYY
请参考下 Code Review 注意事项

和提交 PR 注意事项

lijialin03 · 2025-05-12T08:25:56Z

对于相同目的的反复修改，在git push之前在本地修改的时候，可以使用git commit --amend，而不是git commit -m "新的信息"，这样可以把修改改在上一个commit上，然后再push就会更新commit而不是创建新的

XvLingWYY added 4 commits April 27, 2025 13:11

Create wgangp.yaml

0b02a82

Add files via upload

83ecd01

Add files via upload

0f56f37

Add files via upload

51befb3

paddle-bot bot added the contributor label Apr 27, 2025

Delete examples/wgangp/conf/wgangp.yaml

532f852

XvLingWYY changed the title ~~【Hackathon 8th No.23】RFC：Improved Training of Wasserstein GANs 论文复现~~ 【Hackathon 8th No.23】Improved Training of Wasserstein GANs 论文复现 Apr 27, 2025

XvLingWYY added 21 commits April 27, 2025 18:16

Update wgangp_cifar10.yaml

f0da2e9

Update functions.py

5f18b0b

Update wgangp_cifar10.py

c3d753d

Update model.py

84230d1

Update wgangp_cifar10.py

8331599

Update wgangp_mnist.py

326da7a

Update wgangp_toy.py

b06cee6

Update wgangp_toy.py

Update functions.py

93118bf

Update functions.py

Update model.py

ba7481b

Update wgangp_cifar10.py

afbd7d8

Update wgangp_mnist.py

e3fddbd

Update wgangp_toy.py

3068ff6

Update wgangp_cifar10.yaml

0c2b21c

Update model.py

bdf54d2

Update wgan_gp.md

c0e34b3

Update wgangp_cifar10.yaml

127650b

Update model.py

617c9f3

Update wgan_gp.md

acf302f

Update wgan_gp.md

9549f1a

Update wgan_gp.md

6871216

Add files via upload

2c38a16

XvLingWYY added 7 commits May 1, 2025 16:25

Add files via upload

4194dcc

Update wgangp_cifar10.yaml

5933fbf

Update wgangp_cifar10.yaml

6cdbc38

Add files via upload

0442eee

Add files via upload

15f855c

Add files via upload

865b731

Add files via upload

cbe55e3

lijialin03 reviewed May 9, 2025

View reviewed changes

XvLingWYY added 17 commits May 9, 2025 23:06

Merge branch 'PaddlePaddle:develop' into develop

d9dd40b

Delete examples/wgangp directory

6a4dd20

Create wgangp_cifar10.yaml

59181b3

Add files via upload

ea46640

Add files via upload

e5ce98f

Add files via upload

30bd672

Add files via upload

16eb6df

Add files via upload

8bae84a

Add files via upload

5b7b437

Add files via upload

af72624

Add files via upload

3cbafe5

Add files via upload

d0d8249

Add files via upload

5b43407

Add files via upload

cf80107

Delete examples/wgangp_toy_model.py

56e5d68

Add files via upload

6c872c8

Add files via upload

0722c3c

XvLingWYY added 2 commits May 12, 2025 19:00

Add files via upload

71a103f

Add files via upload

071fbd7

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

【Hackathon 8th No.23】Improved Training of Wasserstein GANs 论文复现 #1146

【Hackathon 8th No.23】Improved Training of Wasserstein GANs 论文复现 #1146

XvLingWYY commented Apr 27, 2025

paddle-bot bot commented Apr 27, 2025

XvLingWYY commented Apr 27, 2025

XvLingWYY commented May 2, 2025

lijialin03 left a comment •

edited

Loading

lijialin03 May 8, 2025

XvLingWYY May 12, 2025

lijialin03 May 8, 2025

XvLingWYY May 12, 2025

lijialin03 May 8, 2025

XvLingWYY May 12, 2025

lijialin03 May 8, 2025

XvLingWYY May 12, 2025

lijialin03 May 8, 2025

XvLingWYY May 12, 2025

lijialin03 May 8, 2025

XvLingWYY May 12, 2025

lijialin03 May 8, 2025

XvLingWYY May 12, 2025

lijialin03 May 8, 2025

XvLingWYY May 12, 2025

lijialin03 May 8, 2025

lijialin03 May 9, 2025

XvLingWYY May 12, 2025

luotao1 commented May 12, 2025

lijialin03 commented May 12, 2025

		if noise is None:
		noise = paddle.randn([self.batch_size, 128], dtype=paddle.float32)

【Hackathon 8th No.23】Improved Training of Wasserstein GANs 论文复现 #1146

Are you sure you want to change the base?

【Hackathon 8th No.23】Improved Training of Wasserstein GANs 论文复现 #1146

Conversation

XvLingWYY commented Apr 27, 2025

PR types

PR changes

Describe

paddle-bot bot commented Apr 27, 2025

XvLingWYY commented Apr 27, 2025

XvLingWYY commented May 2, 2025

lijialin03 left a comment • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

luotao1 commented May 12, 2025

lijialin03 commented May 12, 2025

lijialin03 left a comment •

edited

Loading