search

Bihaqo / t3f

Tensor Train decomposition on TensorFlow
https://t3f.readthedocs.io/en/latest/index.html
MIT License
219 stars 55 forks source link

Slow execution of some ops #104

Closed Bihaqo closed 6 years ago

Bihaqo commented 6 years ago

T3f on CPU scales worse than lineary with e.g. multipliying a batch of TT-matrices by a TT-matrix. This is weird.

Also, matrix-by-matrix multiplication is slower than that of TTPY.

See profile_cpu_vs_gpu branch

Bihaqo commented 6 years ago

The reason turned out to be slow tf.transpose for tensors of more than 5 dims. To mitigate, apply the following patch and compile TensorFlow from sources.