A modern programming language that makes developers happier.

kotlinlang

<@UQGG9A0LX> do you think it is possible to build <https://arxiv.org/abs/1511.06434|DCGANs> with KotlinDL? They have stuff like <https://pytorch.org/docs/stable/generated/torch.nn.LeakyReLU.html,|LeaklyReLU>, <https://pytorch.org/docs/stable/generated/torch.nn.BatchNorm2d.html|BatchNorm2d> for example?

Is there a way to work directly with tensors in KoltinDL or you can only build models sequentially with the provided layers?

If not, what is the lib to work with tensor that supports GPU? 

<@U01TFT8DQSY>, you should be able to use <https://www.tensorflow.org/jvm/install|tensorflow> or <https://github.com/deeplearning4j|deeplearning4j> ecosystem

At the current release 0.1.1, KotlinDL supports only Sequential models with VGG style architecture, since release 0.2 will support Functional API to build models like ResNet or MobileNet. BatchNorm2d could be executed in inference mode only in 0.2, Conv2dTranspose or Upsampling2D will be supported later, closer to the end of 2021. I suppose GANs architectures are not the main goal now. About tensor manipulations - it's not a goal of KotlinDL, for tensor manipulations could be used Java API Tensorflow (old or new).

<@UQGG9A0LX> you run inference on `Array&lt;Array&lt;FloatArray&gt;&gt;` <https://github.com/JetBrains/KotlinDL/blob/master/docs/loading_trained_model_for_inference.md|here> . I am sure you would prefer to have a genuine tensor library behind KotlinDL

why not integrating with <https://github.com/mipt-npm/kmath|kmath>