pytorch/xla

Pull Requests Commits

Add pass to remove in-place operations on output parameters

asuhan committed 7 years ago

4b1edad5

Keep track of evaluated aten::size

asuhan committed 7 years ago

7ded34a6

Initial changes for latest PyTorch

asuhan committed 7 years ago

abd38b35

Merge pull request #62 from pytorch/test_util_restructure1

jysohn23 committed 7 years ago

Verified d4fbf359

Do not import xla_model.py from test_util.py, as this ends up picking up the xla_model.py from the installed module, instead of the local one.

dlibenzi committed 7 years ago

b709d55e

Add tensorboardX support for all our models

jysohn23 committed 7 years ago

Verified 161686d9

Addressed review comments and ran google-pyformat

jysohn23 committed 7 years ago

cd35dc52

Handle cases for metrics_debug

jysohn23 committed 7 years ago

ea068be1

Remove explicit writer from cifar and imagenet models

jysohn23 committed 7 years ago

b3e9cfcf

Fix processed examples calculated

jysohn23 committed 7 years ago

7ec1597c

Create new classes for writing and logging metrics

jysohn23 committed 7 years ago

1d410082

Merge branch 'jysohn/tbx' of https://github.com/pytorch/xla into jysohn/tbx

jysohn23 committed 7 years ago

016b6aa2

Add tensorboardX support for convergence tests

jysohn23 committed 7 years ago

f6c85c9b

Add tensorboardX support for test_train_mnist.py

jysohn23 committed 7 years ago

4f33a5c2

Dedicate IO thread pool to non CPU bound closures.

dlibenzi committed 7 years ago

fa355a71

Make execute parallel run the multiple sessions in parallel.

dlibenzi committed 7 years ago

e40f3117

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

dlibenzi committed 7 years ago

b5b308c5

Allow debug metrics to be dumped within the train loop.

dlibenzi committed 7 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 7 years ago

d77401d9

Use XLA_CHECK macros to sump stack trace.

dlibenzi committed 7 years ago

f549953b

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

dlibenzi committed 7 years ago

e21a4566

Fix build script.

dlibenzi committed 7 years ago

13341ab0

Reformat Python code using Google pyformat.

dlibenzi committed 7 years ago

57926630

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

dlibenzi committed 7 years ago

32ab44c7

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

dlibenzi committed 7 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 7 years ago

cbf11819

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

dlibenzi committed 7 years ago

2cc427ee

Fix typo: set_inputs_gardients -> set_input_gradients

asuhan committed 7 years ago

b9c2bed1

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

dlibenzi committed 7 years ago

54a33873

Fix translator.cpp formatting

asuhan committed 7 years ago

452ad483

Newer Older