Add support for Paddle and VisualDL

[lamhoangtung]

Recently, there was Paddle, a new deep learning framework that gained a lot of traction in production usage, along with their own equivalent for Tensorboard (or TensorboardX) visualization tools called VisualDL. We and many company had started to use Paddle and VisualDL in ours DL workflows.

So, it would be great if ClearML can extend the support to Paddle and VisualDL in a future version. For now with our test, everything from Iteration count to metric visualization using VisualDL are not working with ClearML yet.

[bmartinn]

Hi @lamhoangtung Sure that sounds like a great match for automagic :) A few things that will help accelerate this process:

1. Simple sample code of Paddle training (maybe something similar to the mnist examples). (It should cover model load/save, and logging metrics)

2. GPU / CUDA package matrix link (like the one torch has), in order to have full clearml-agent support

BTW: From the looks of it VisualDL is very similar to Tensorboard , so hopefully we could reuse the code :)

[lamhoangtung]

Thanks for the interest @bmartinn

I will try to find some spare time to update a simple sample code of Paddle training soon.

For now, here is the package matrix link, luckily they had something like Pytorch ;)

https://www.paddlepaddle.org.cn/install/quick

[adri]

I'd also love to see better Paddle support. Does this example help? https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.7/tools/train.py

[ainoam]

Thanks for the example @adri. We'll look for the right opportunity to add such support.

[adri]

@ainoam thanks for the answer! Is there something i can contribute? I saw these bindings here https://github.com/allegroai/clearml/tree/master/clearml/binding/frameworks Is that the way to go?

[ainoam]

That's exactly the place @adri. We'll definitely appreciate your contribution if you're feeling up for it. If you need any help, head over to the clearml-dev channel on ClearML's slack space.