[Refine] Fix several bugs on custom device

[HydrogenSulfate]

PR types

Bug fixes

PR changes

Others

Describe

This pull request updates documentation and code to improve clarity, device handling, and compatibility across several examples and model architectures. The most significant changes include correcting device usage in code (removing unnecessary .cuda() and device assignments), refining training and evaluation command documentation for models, and ensuring proper data type handling in plotting routines.

Documentation improvements:

• Swapped and clarified the "模型训练命令" and "模型评估命令" sections in the Chinese documentation for nsfnet.md and nsfnet4.md, ensuring that commands for training and evaluation are correctly labeled and use the appropriate arguments. [1] [2] [3] [4]
• Added missing "模型评估命令" sections in example documentation files for cgcnn.md, phygeonet.md, and velocity_gan.md to improve usability and completeness. [1] [2] [3]
• Updated dataset structure formatting in unetformer.md for clarity, changing the code block type to sh.

Device handling and compatibility:

• Removed unnecessary .cuda() calls in phycrnet/functions.py and phycrnet.py, making code device-agnostic and compatible with both CPU and GPU execution. [1] [2] [3] [4] [5]
• Eliminated explicit device assignment in training and evaluation routines for tgcn/run.py, and removed device transfer in unetformer/vaihingen_test.py, simplifying code and reducing potential device mismatches. [1] [2] [3]
• Removed device string manipulation and .to(self.device) calls in the stafnet.py model, further improving device compatibility. [1] [2] [3]

Data type and plotting fixes:

• Ensured proper conversion to Python scalars and explicit data type specification for plotting routines in VP_NSFNet4.py, preventing type errors and improving compatibility. [1] [2] [3] [4]

Miscellaneous:

• Added a missing pass statement to the ModuleNotFoundError exception in chemprop_molecule_utils.py to avoid runtime errors.

[paddle-bot]

Thanks for your contribution!