v1.2.3#318
Conversation
PiperOrigin-RevId: 168274573
…ded to scatter the representations after attention back into the right positions. Added test2dGatherAndScatter, which tests for invertibility of 2d gather and scatter functions. PiperOrigin-RevId: 168283655
…. Now only applied for the attention expert. PiperOrigin-RevId: 168292028
…oregression/masking. PiperOrigin-RevId: 168302680
…ttention in local_1d and 2d attention functions. Adding a flag for image summaries in dot_product_attention because we need to figure out the best way to get image summaries in 2d functions. PiperOrigin-RevId: 168472353
PiperOrigin-RevId: 168570496
PiperOrigin-RevId: 168579149
PiperOrigin-RevId: 168579271
PiperOrigin-RevId: 168599850
…_timing_signal_1d_given_position in common_attention.py that takes timing positions (as a tensor of shape [batch, length]). PiperOrigin-RevId: 168601518
PiperOrigin-RevId: 168757167
PiperOrigin-RevId: 168765336
PiperOrigin-RevId: 168782469
…cify individual checkpoints. PiperOrigin-RevId: 168867603
…inference and beam search work in prepend mode. After this change, inference in prepend mode requires batch size 1, since padding is not properly ignored. PiperOrigin-RevId: 168889211
…during PiperOrigin-RevId: 169116643
PiperOrigin-RevId: 169135518
…variable scopes so that they don't share parameters. PiperOrigin-RevId: 169147528
PiperOrigin-RevId: 169162566
PiperOrigin-RevId: 169166125
PiperOrigin-RevId: 169187769
…Dataset.apply()`. PiperOrigin-RevId: 169264919
PiperOrigin-RevId: 169279745
PiperOrigin-RevId: 169297690
PiperOrigin-RevId: 169299088
PiperOrigin-RevId: 169299895
PiperOrigin-RevId: 169312588
PiperOrigin-RevId: 169345762
PiperOrigin-RevId: 169347220
PiperOrigin-RevId: 169354689
PiperOrigin-RevId: 169412526
PiperOrigin-RevId: 169476287
PiperOrigin-RevId: 169482949
PiperOrigin-RevId: 169554635
PiperOrigin-RevId: 169566059
PiperOrigin-RevId: 169625024
|
We found a Contributor License Agreement for you (the sender of this pull request), but were unable to find agreements for the commit author(s). If you authored these, maybe you used a different email address in the git commits than was used to sign the CLA (login here to double check)? If these were authored by someone else, then they will need to sign a CLA as well, and confirm that they're okay with these being contributed to Google. |
|
Travis failed. Hold on merge. Fixing... |
PiperOrigin-RevId: 169629386
… tuple and list PiperOrigin-RevId: 169632380
|
Looking good now. Ready to merge. |
|
thank you guys for your good work, t2t is really amazing, just want to ask if the decoding speed has been improved in the new version, and if it is safe to upgrade to 1.2.3 from 1.2.2 and continuing training on models that was trained on the previous version. (From 1.2.1 to 1.2.2, the change in batching scheme force me to dump all old models :(( ) |
No description provided.