关于数据集的问题 #491

New Issue

claunia · 2026-01-29T21:48:12Z

claunia commented

2026-01-29 21:48:12 +00:00

Originally created by @izhaolinger on GitHub (May 21, 2024).

非常感谢作者的贡献，GFPGAN的效果让人惊艳。我尝试使用私有数据集定制化训练，但train_gfpgan_v1配置文件中所需的数据集格式，我不太理解，请问：

train - dataroot_gt参数下是512x512尺寸的高清图片吗？
val - dataroot_lq和dataroot_gt参数下应该填写什么路径呢？
如果我想在v1.2模型上做微调，我应该怎么加载预训练模型呢？

感谢，期待回复！

Originally created by @izhaolinger on GitHub (May 21, 2024). 非常感谢作者的贡献，GFPGAN的效果让人惊艳。我尝试使用私有数据集定制化训练，但train_gfpgan_v1配置文件中所需的数据集格式，我不太理解，请问： 1. train - dataroot_gt参数下是512x512尺寸的高清图片吗？ 2. val - dataroot_lq和dataroot_gt参数下应该填写什么路径呢？ 3. 如果我想在v1.2模型上做微调，我应该怎么加载预训练模型呢？感谢，期待回复！

claunia closed this issue

2026-01-29 21:48:13 +00:00

claunia commented

2026-01-29 21:48:13 +00:00

@doniaa24 commented on GitHub (May 22, 2024):

Training Data Format: Yes, for the train - dataroot_gt parameter, you should use high-definition images with a size of 512x512 pixels.
Validation Data Paths:
val - dataroot_lq: This should be the path to your low-quality or damaged images. These are the images that your model will attempt to restore.
val - dataroot_gt: This path should lead to the clean or undamaged versions of the images specified in dataroot_lq. It is crucial for assessing the model's performance during validation.
Note: Ensure that both directories (dataroot_lq and dataroot_gt) contain the same number of images as each damaged image in dataroot_lq should have its matching clean version in dataroot_gt for effective training and validation.

@doniaa24 commented on GitHub (May 22, 2024): Training Data Format: Yes, for the train - dataroot_gt parameter, you should use high-definition images with a size of 512x512 pixels. Validation Data Paths: val - dataroot_lq: This should be the path to your low-quality or damaged images. These are the images that your model will attempt to restore. val - dataroot_gt: This path should lead to the clean or undamaged versions of the images specified in dataroot_lq. It is crucial for assessing the model's performance during validation. Note: Ensure that both directories (dataroot_lq and dataroot_gt) contain the same number of images as each damaged image in dataroot_lq should have its matching clean version in dataroot_gt for effective training and validation.

claunia commented

2026-01-29 21:48:14 +00:00

@izhaolinger commented on GitHub (May 29, 2024):

Training Data Format: Yes, for the train - dataroot_gt parameter, you should use high-definition images with a size of 512x512 pixels. Validation Data Paths: val - dataroot_lq: This should be the path to your low-quality or damaged images. These are the images that your model will attempt to restore. val - dataroot_gt: This path should lead to the clean or undamaged versions of the images specified in dataroot_lq. It is crucial for assessing the model's performance during validation. Note: Ensure that both directories (dataroot_lq and dataroot_gt) contain the same number of images as each damaged image in dataroot_lq should have its matching clean version in dataroot_gt for effective training and validation.

非常感谢您的回复，很有帮助！

@izhaolinger commented on GitHub (May 29, 2024): > Training Data Format: Yes, for the train - dataroot_gt parameter, you should use high-definition images with a size of 512x512 pixels. Validation Data Paths: val - dataroot_lq: This should be the path to your low-quality or damaged images. These are the images that your model will attempt to restore. val - dataroot_gt: This path should lead to the clean or undamaged versions of the images specified in dataroot_lq. It is crucial for assessing the model's performance during validation. Note: Ensure that both directories (dataroot_lq and dataroot_gt) contain the same number of images as each damaged image in dataroot_lq should have its matching clean version in dataroot_gt for effective training and validation. 非常感谢您的回复，很有帮助！

claunia commented

2026-01-29 21:48:14 +00:00

@Bobchenyx commented on GitHub (Dec 8, 2024):

非常感谢作者的贡献，GFPGAN的效果让人惊艳。我尝试使用私有数据集定制化训练，但train_gfpgan_v1配置文件中所需的数据集格式，我不太理解，请问：

train - dataroot_gt参数下是512x512尺寸的高清图片吗？

val - dataroot_lq和dataroot_gt参数下应该填写什么路径呢？

如果我想在v1.2模型上做微调，我应该怎么加载预训练模型呢？

感谢，期待回复！

您好, 请问可以加下您联系方式嘛? 想咨询有关模型微调相关的问题

@Bobchenyx commented on GitHub (Dec 8, 2024): > 非常感谢作者的贡献，GFPGAN的效果让人惊艳。我尝试使用私有数据集定制化训练，但train_gfpgan_v1配置文件中所需的数据集格式，我不太理解，请问： > > 1. train - dataroot_gt参数下是512x512尺寸的高清图片吗？ > 2. val - dataroot_lq和dataroot_gt参数下应该填写什么路径呢？ > 3. 如果我想在v1.2模型上做微调，我应该怎么加载预训练模型呢？ > > 感谢，期待回复！您好, 请问可以加下您联系方式嘛? 想咨询有关模型微调相关的问题

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: TencentARC/GFPGAN#491