| Commit message (Collapse) | Author | Age |
|
|
| |
Mismatch after the fix in #17815
|
|
|
|
|
|
|
| |
packages:
data
training
tensor_forest
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* Prevent ctc_loss op from segfaulting when given empty batch.
PiperOrigin-RevId: 163663460
* New "SavedModel: Practical Uses" and "SavedModel: Architecture" documents.
PiperOrigin-RevId: 163669809
* Minor cleanup
PiperOrigin-RevId: 163685423
* Add regression variance over individual trees to TensorForest inference.
PiperOrigin-RevId: 163695881
|
|
|
|
|
|
| |
ultimate interpretability.
PiperOrigin-RevId: 163466324
|
|
|
|
| |
PiperOrigin-RevId: 162944683
|
|
|
|
| |
PiperOrigin-RevId: 162222071
|
|
|
|
| |
PiperOrigin-RevId: 161949540
|
|
|
|
|
|
| |
training time, instead of only after FinalizeTreeOp.
PiperOrigin-RevId: 161663317
|
|
|
|
| |
PiperOrigin-RevId: 161396592
|
|
|
|
| |
PiperOrigin-RevId: 159852889
|
|
|
|
| |
PiperOrigin-RevId: 159728380
|
|
|
|
| |
PiperOrigin-RevId: 159718610
|
|
|
|
|
|
|
|
| |
mix of dense and sparse features. Move feature processing to tensor_forest to have the most sane interface (just dict of tensors or single tensor).
Had to move data_ops because of python3 import weirdness.
Change value == bias to go right instead of left because this inherently handles sparse one-hot categorical data (if a node's bias is set at 1 (the only value it could pick), both values of 1 and 0 are not strictly > than 1, so that node would always go left).
Change: 143970989
|
|
|
|
|
| |
Empirically, n ~ 75 gives the most speedup for the bootstrap method.
Change: 142684023
|
|
of useless template code. Load ops at import time, which is more idiomatic with the rest of tf/contrib.
Also rename TensorForests string_to_float op to be more descriptive and unique (reinterpret_string_to_float).
Change: 141756823
|