Skip to content
This repository was archived by the owner on Jul 7, 2023. It is now read-only.

v1.5.5 #646

Merged
merged 25 commits into from
Mar 10, 2018
Merged

v1.5.5 #646

Changes from 1 commit
Commits
Show all changes
25 commits
Select commit Hold shift + click to select a range
ca967c1
Internal change
Mar 2, 2018
908981e
Set evaluation_master to master in RunConfig
Mar 2, 2018
e2e61ac
Add option to do soft EM instead of hard EM
a-googler Mar 5, 2018
f92d901
Add an inv_temp hparam for controlling softness of EM
a-googler Mar 6, 2018
f36f82c
More Librispeech subsets to help with mixed clean and noisy data trai…
a-googler Mar 6, 2018
40d1f15
proper em - P(c_i) is computed using ema_count instead of actual counts
a-googler Mar 6, 2018
8320faf
increase pseudo-count to 1.0 and now there's no NaN in training
a-googler Mar 7, 2018
d83d992
Use logits instead of probs to compute supervised attention loss.
a-googler Mar 8, 2018
7056827
Why do we need stop gradient here?
a-googler Mar 8, 2018
e1e8fbb
Add expected_attention_loss_type hparam to supervised_attention to al…
a-googler Mar 8, 2018
5ee776d
ema_count trainable should be False; this was causing the weird dp be…
a-googler Mar 8, 2018
7293efc
Fix multi-logit loss computation error.
aidangomez Mar 8, 2018
75d2aef
Basic autoencoder and improvements in image modality.
Mar 9, 2018
c4e6fab
Change batch size for hparam config
Mar 9, 2018
9ae5bc2
Make Vanilla GAN work, based on Compare GAN code.
Mar 9, 2018
1568e9b
internal
Mar 9, 2018
95053b4
Bump release number.
Mar 9, 2018
9a638df
Documentation for cloud TPU for Image Transformer. Additional default…
Mar 9, 2018
3de51ab
Add smoothed L0 prior and trainable logits for cluster probabilities.
Mar 9, 2018
6a6d9fe
Added the ema count smoothing update inside the else.
Mar 9, 2018
6e846f2
Make text_encoder unicode conversion a pass-through
Mar 9, 2018
d8080a1
Pass in decode_hp to _interactive_input_fn and remove summaries when
a-googler Mar 10, 2018
c7495b5
six.iteritems for Py3
Mar 10, 2018
329123f
Update Travis tests for Py3 to run TF 1.6
Mar 10, 2018
688f4d5
Update ISSUE_TEMPLATE
Mar 10, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
Why do we need stop gradient here? 
PiperOrigin-RevId: 188341955
  • Loading branch information
T2T Team authored and Ryan Sepassi committed Mar 9, 2018
commit 7056827fda3dd94ab70b4f13e0c5d268545ab2a2
2 changes: 1 addition & 1 deletion tensor2tensor/layers/discretization.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -109,7 +109,7 @@ def nearest_neighbor(x,
else:
nearest_idx = tf.argmax(-dist, axis=-1)
nearest_hot = tf.one_hot(nearest_idx, block_v_size)
return tf.stop_gradient(nearest_hot)
return nearest_hot


def embedding_lookup(x,
Expand Down