As the tensorflow code uses numpy to PCA the output embeddings, it is not possible to take advantage of this when adding torch-vggish to other networks (the usecase for this is relatively small). It would be useful to reimplement the PCA algorithm so that it can operate on Torch tensors.
As the tensorflow code uses numpy to PCA the output embeddings, it is not possible to take advantage of this when adding torch-vggish to other networks (the usecase for this is relatively small). It would be useful to reimplement the PCA algorithm so that it can operate on Torch tensors.