Closed fchollet closed 1 year ago
I've made more extensive modifications on my fork, that are becoming too heavy for a single PR. Closing this. Happy to reopen a PR for a more focused modification of your choice (e.g. switch to predict_on_batch()
).
...and make it faster via graph functions (i.e. calling models via
.predict_on_batch()
rather than__call__()
), and addjit_compile
option (XLA compilation).On my system this is 1.3-1.4x faster. I expect it may be up to 2x faster on GPU.