| Commit message (Collapse) | Author | Age |
|
|
|
|
|
|
|
|
| |
Now the layout of tensorflow/contrib/cudnn_rnn is similar to
tensorflow/contrib/layers
Delete repeated tests in cudnn_rnn_ops_test (new tests in cudnn_rnn_test).
PiperOrigin-RevId: 175487631
|
|
|
|
|
|
| |
wrappers
PiperOrigin-RevId: 174523358
|
|
|
|
| |
PiperOrigin-RevId: 173560463
|
|
|
|
|
|
| |
wrappers
PiperOrigin-RevId: 173494053
|
|
|
|
|
|
| |
Also add reuse arg to LSTMBlockCell.
PiperOrigin-RevId: 167030950
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
1. Use a lossy way to save/restore cudnn biases during checkpointing.
Cudnn uses 2 biases each gate for all RNNs while tf uses one. To allow cudnn checkpoints
to be compatible with both Cudnn and platform-independent impls, previously both
individual bias and summed biases each gate were stored.
The new way only stores the bias sum for each gate, and split it half-half when
restoring from a cudnn graph. Doing this does not cause problems since RNNs do not use
weight-decay to regularize.
2. Use inheritance instead of branching
* Split RNNParamsSaveable to 1 base class and 4 subclasses.
* Extract common routines and only overwrite rnn-type-specific pieces in subclasses.
PiperOrigin-RevId: 166413989
|
|
|
|
| |
Change: 151185292
|
|
|
|
|
|
| |
which is now hidden, it should be added to the _allowed_symbols list in the appropriate __init__.py file.
Change: 145943844
|
|
|
|
|
|
|
|
|
|
|
| |
TESTED:
- opensource_build passed:
https://ci.tensorflow.org/job/tensorflow-cl-presubmit-multijob/8309/
https://ci.tensorflow.org/job/tensorflow-cl-presubmit-multijob/8641/
https://ci.tensorflow.org/job/tensorflow-cl-presubmit-multijob/9331/
https://ci.tensorflow.org/job/tensorflow-cl-presubmit-multijob/9406/
- passed unit tests
Change: 141227871
|
|
It is about 2-3x faster compared to rnn_cell.LSTMCell and lstm_ops.LSTMBlockCell.
Cudnn LSTM speedup Cudnn LSTM speedup
over rnn.LSTMCell over rnn.LSTMBlockCell
large 200.00% 192.27%
medium 247.75% 228.38%
small 500.00% 438.10%
The step-time per second for each model size.
Cudnn LSTM rnn_cell.LSTMCell lstm_ops.LSTMBlockCell
large 0.0854 0.2562 0.2496
medium 0.0222 0.0772 0.0729
small 0.0042 0.0252 0.0226
TESTED:
- opensource_build
https://ci.tensorflow.org/job/tensorflow-cl-presubmit-multijob/4568/
- passed unit tests
Change: 131472315
|