FS#64574 - [tensorflow] OneDeviceStrategy/MirrowedStrategy broken after upgrade to 2.0.0.5
Attached to Project:
Community Packages
Opened by Oliver Kowalke (olk) - Tuesday, 19 November 2019, 20:26 GMT
Last edited by Sven-Hendrik Haase (Svenstaro) - Thursday, 12 December 2019, 05:31 GMT
Opened by Oliver Kowalke (olk) - Tuesday, 19 November 2019, 20:26 GMT
Last edited by Sven-Hendrik Haase (Svenstaro) - Thursday, 12 December 2019, 05:31 GMT
|
Details
Description: Using OneDeviceStrategy(device='/gpu:0')
results in error "RuntimeError:
/job:localhost/replica:0/task:0/device:GPU:0 unknown
device." after upgrading to version 2.0.0.5.
MirroredStrategy() uses CPU instread of GPU. Maybe an upstream bug. Additional info: * 2.0.0.5 Steps to reproduce: run example that uses a strategy: strategy = tf.distribute.OneDeviceStrategy(device="/gpu:0") #strategy = tf.distribute.MirroredStrategy() with strategy.scope(): network_layers = small.layers(input_shape) |
This task depends upon
Closed by Sven-Hendrik Haase (Svenstaro)
Thursday, 12 December 2019, 05:31 GMT
Reason for closing: Fixed
Thursday, 12 December 2019, 05:31 GMT
Reason for closing: Fixed
2.0.0.4 was broken because of Python-3.8 update (see
FS#64528)So 2.0.0.3 needs to be tested - could you tell me at which date 2.0.0.4 was updated?
Because I do a complete system downgrade by pinning pacman to a specific date - I'd like to choose a date for downgrading one day before 2.0.0.4.
Could you concider to add a Arch package for keras-tuner and tensorflow_dataset?