No WebGPU support for layers required by 3D segmentation CNNs due to no 4D Upsampling or Transposed 3D Convolution support

[jphdotam]

Please make sure that this is a feature request. As per our GitHub Policy, we only address code/doc bugs, performance issues, feature requests and build/installation issues on GitHub. tag:feature_template

System information

• TensorFlow.js version (you are using): 4.0.0
• Are you willing to contribute it (Yes/No): Beyond my skills

Describe the feature and the current behavior/state. To perform 3D segmentation e.g. a BS * H * W * Depth/time with e.g. a UNet you require a way to 'upsample' on the 'right' side of the U. This is performed using Interpolation/Upsampling or using a transposed convolution. Unfortunately 4D Upsampling doesn't seem to be supported by TFJS. Transposed convolutions are, but don't have GPU support:

Error during inference: TypeError: Unknown op 'Conv3DBackpropInputV2'.

Could one of these be implemented?

Will this change the current api? How? No

Who will benefit with this feature? Anyone using 3D segmentation networks.

Any Other info.