Updated code for conda, python 3.7, PyTorch 1.8, cuda 10/11 on ubuntu 20.04 #592
Conversation
Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
Signed-off-by: Sangeetha Siddegowda <[email protected]> Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
…tom markers with pytorch layer names Also reverted back ONNX export mode to default(eval) from training Signed-off-by: Abhi Khobare <[email protected]> Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
Signed-off-by: Abhi Khobare <[email protected]> Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
… nodes. Signed-off-by: Abhi Khobare <[email protected]> Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
Signed-off-by: Abhi Khobare <[email protected]> Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
Signed-off-by: Sendil Krishna <[email protected]> Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
Signed-off-by: Bharath Ramaswamy <[email protected]> Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
TF tests fail because using tf.contrib modules removed in TF 2.x Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
added conda environment Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
updated to cmake 3.18+ to support cuda architectures Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
CPU tests passing in conda Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
…f quantizers Signed-off-by: Sangeetha Siddegowda <[email protected]> Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
…rner cases Signed-off-by: Abhi Khobare <[email protected]> Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
Signed-off-by: Abhi Khobare <[email protected]> Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
… 10 and 11 Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
Signed-off-by: Mikael Bourges-Sevenier <[email protected]>
|
PR is failing because your CI is using an older version of cmake. Without newer version, Cuda and lapack are not correctly found and linked on conda environments. @quic-akhobare please upgrade. |
|
@mikeseven Thank you for pushing this update. Sorry about the delayed response. And, a quick note - we are working on updating AIMET to support TF2.0. @quic-akhobare @quic-bharathr could you please review this update? |
|
Can't wait to test the new version with TF 2. Any tentative timeline? |
Difficult to project an ETA at the moment. So one thing we realized that several contrib modules got deprecated and removed in TF 2.4. E.g. AIMET currently depends on contrib.quantize.python.graph_matcher module in TF 1.15. This is not present in TF 2.4. We are looking at alternatives. @mikeseven If you are familiar with this module by chance, any suggestions would be welcome. |
# Conflicts: # Jenkins/Dockerfile # TrainingExtensions/common/CMakeLists.txt # TrainingExtensions/tensorflow/src/QcQuantizeOpDeprecated.hpp # TrainingExtensions/tensorflow/test/python/test_qc_quantize_op_deprecated.py # TrainingExtensions/torch/src/python/aimet_torch/onnx_utils.py # TrainingExtensions/torch/src/python/aimet_torch/torchscript_utils.py # TrainingExtensions/torch/test/python/test_onnx_utils.py # TrainingExtensions/torch/test/python/test_quantizer.py # dobuildntest.sh # packaging/dependencies/reqs_pip_common.txt # packaging/version.txt
This PR updates AIMET to almost latest versions of various dependencies in a conda environment.
Tensorflow remains at 1.15 since AIMET uses deprecated contrib code not available in TF 2. Hopefully, AIMET team would upgrade one day!
Conda TF 1.15 conflicts with Pytorch on cuda, forcing Pytorch to be CPU only. There is no such issue with TF 2.
Most tests passes (
dobuildntest.sh -u) though many have errors regardless of CPU or GPU. AIMET team, please check.No update on docker containers. Still on old versions with many tests not passing.
AIMET team, please help make this code robust.