xla
Opti para
#85
Merged
Overview
Commits
15
Changes
View On GitHub
Commits
  • Avoid tensors sync on fused train.
    dlibenzi committed 7 years ago
  • Rename a few tensor API for consistency.
    dlibenzi committed 7 years ago
  • Generate the XLA computations in parallel.
    dlibenzi committed 7 years ago
  • Smoother rate.
    dlibenzi committed 7 years ago
  • No need to re-check for nullptr as GetApplyOrder() already filtered out the tensors which are noth graph backed.
    dlibenzi committed 7 years ago
  • Allow tensors to reference data from other tensors, so that new generations of inputs and gradient can be tracked back (via the unique ID) to the same tensor kind.
    dlibenzi committed 7 years ago
  • Use a context to capture the ApplyPendingGraph() metadata, to allow sped up operations in case the new apply matches the one within the saved context.
    dlibenzi committed 7 years ago
  • Minor code refactoring for the loader wrapper batch counting.
    dlibenzi committed 7 years ago
  • In case of multiple tensors providing data, select the older one.
    dlibenzi committed 7 years ago
  • Added more counters to track parallel apply status.
    dlibenzi committed 7 years ago
  • Now that we route CHECKs to exceptions, care should be taken when running lambdas.
    dlibenzi committed 7 years ago
  • Re-format python file according to Google pyformat.
    dlibenzi committed 7 years ago
  • Refactor some code into a function.
    dlibenzi committed 7 years ago
  • Minor code refactoring.
    dlibenzi committed 7 years ago
  • Refactored the execute cached path.
    dlibenzi committed 7 years ago
FAQ Terms Privacy Refunds Impressum
Loading