pytorch/xla

Pull Requests Commits

Add tensorboardX support for convergence tests

jysohn23 committed 6 years ago

f6c85c9b

Add tensorboardX support for test_train_mnist.py

jysohn23 committed 6 years ago

4f33a5c2

Dedicate IO thread pool to non CPU bound closures.

dlibenzi committed 6 years ago

fa355a71

Make execute parallel run the multiple sessions in parallel.

dlibenzi committed 6 years ago

e40f3117

Use the number of devices as default value for the number of releaser threads.

dlibenzi committed 6 years ago

b5b308c5

Allow debug metrics to be dumped within the train loop.

dlibenzi committed 6 years ago

ca97e619

With many cores, and until we make use of the new XRTWriteLiteral op, we create and destroy many handles, and one single releaser thread cannot keep up.

dlibenzi committed 6 years ago

d77401d9

Use XLA_CHECK macros to sump stack trace.

dlibenzi committed 6 years ago

f549953b

Use a unique ID to sort tensors during the sync operation.

dlibenzi committed 6 years ago

e21a4566

Fix build script.

dlibenzi committed 6 years ago

13341ab0

Reformat Python code using Google pyformat.

dlibenzi committed 6 years ago

57926630

Import latest TF features and avoid tuple decomposition calls/RPCs.

dlibenzi committed 6 years ago

32ab44c7

Bring new TF commits with the auto-tuple-decompose XRT feature.

dlibenzi committed 6 years ago

8c9cf384

The Python GC does not immediately release the collected objects, so we need to wait for it to happen before heading out.

dlibenzi committed 6 years ago

cbf11819

Make an attempt to release the pending XLA tensors on exit.

dlibenzi committed 6 years ago

2cc427ee

Fix typo: set_inputs_gardients -> set_input_gradients

asuhan committed 6 years ago

b9c2bed1

Added MNIST training test using the PyTorch legacy (or almost) interface.

dlibenzi committed 6 years ago

54a33873

Fix translator.cpp formatting

asuhan committed 6 years ago

452ad483

Use implicit once-like initialization for function static variables and simplify code further.

dlibenzi committed 6 years ago

50cc3f5c

Made the computation client live forever (lie all the static data does in g3).

dlibenzi committed 6 years ago

36c1a1ff

Merge pull request #41 from pytorch/asuhan/remove_mkl_dnn_dep

jysohn23 committed 6 years ago

Verified d2f255bb

Implement condition variable based triggering for handle releases instead of sleeps.

dlibenzi committed 6 years ago

bc4dd1e1

Remove mkldnn dependency

asuhan committed 6 years ago

4afca440

dlibenzi committed 6 years ago

ede9c967

Move handle release in background. (#37)

dlibenzi committed 6 years ago

551c2e89

The compare model was broken ... this is why resnet18 was working.

dlibenzi committed 6 years ago

19cbea9a

Enable Continuous Integration for pytorch/xla

jysohn23 committed 6 years ago

Verified c26880ad

Enable Continuous Integration for pytorch/xla

jysohn23 committed 6 years ago

37fa9059

Move the metric snapshotting after the release handles call, since that already has its metric.

dlibenzi committed 6 years ago

091a334e

Refactored the tests and added resnet18 test.

dlibenzi committed 6 years ago

81317626

Newer Older