added use_bias=False to linear projections

Kyubyong · Mar 3, 2019 · 6715edc · 6715edc
1 parent e61abd9
commit 6715edc
Show file tree

Hide file tree

Showing 44 changed files with 21,192 additions and 21,191 deletions.
diff --git a/README.md b/README.md
@@ -52,9 +52,10 @@ python train.py --logdir myLog --batch_size 256 --dropout_rate 0.5
 
 * STEP 3. Or download the pretrained models.
 ```
-wget -qO- --show-progress https://dl.dropbox.com/s/4o7zwef7kzma4q4/log.tar.gz | tar xz
+wget -qO- --show-progress https://dl.dropbox.com/s/efv2gmq5hu3np43/log.tar.gz | tar xz
 ```
 
+
 ## Training Loss Curve
 <img src="fig/loss.png">
 
@@ -77,7 +78,7 @@ python test.py --ckpt log/1/iwslt2016_E19L2.62-29146 (OR yourCkptFile OR yourCkp
 
 |tst2013 (dev) | tst2014 (test) |
 |--|--|
-|26.93|23.16|
+|26.69|22.46|
 
 ## Notes
 * Beam decoding will be added soon.

diff --git a/eval/1/iwslt2016_E01L5.85B0.00 b/eval/1/iwslt2016_E01L5.85B0.00
diff --git a/eval/1/iwslt2016_E01L5.95B0.26 b/eval/1/iwslt2016_E01L5.95B0.26
diff --git a/eval/1/iwslt2016_E02L5.16B2.20 b/eval/1/iwslt2016_E02L5.16B2.20
diff --git a/eval/1/iwslt2016_E02L5.24B1.36 b/eval/1/iwslt2016_E02L5.24B1.36
diff --git a/eval/1/iwslt2016_E03L4.41B8.06 b/eval/1/iwslt2016_E03L4.41B8.06
diff --git a/eval/1/iwslt2016_E03L4.57B4.16 b/eval/1/iwslt2016_E03L4.57B4.16
diff --git a/eval/1/iwslt2016_E04L4.05B12.44 b/eval/1/iwslt2016_E04L4.05B12.44
diff --git a/eval/1/iwslt2016_E04L4.06B15.63 b/eval/1/iwslt2016_E04L4.06B15.63
diff --git a/eval/1/iwslt2016_E05L3.51B19.15 b/eval/1/iwslt2016_E05L3.51B19.15
diff --git a/eval/1/iwslt2016_E05L3.68B16.64 b/eval/1/iwslt2016_E05L3.68B16.64
diff --git a/eval/1/iwslt2016_E06L3.43B21.47 b/eval/1/iwslt2016_E06L3.43B21.47
diff --git a/eval/1/iwslt2016_E06L3.46B19.83 b/eval/1/iwslt2016_E06L3.46B19.83
diff --git a/eval/1/iwslt2016_E07L3.32B22.57 b/eval/1/iwslt2016_E07L3.32B22.57
diff --git a/eval/1/iwslt2016_E07L3.36B22.13 b/eval/1/iwslt2016_E07L3.36B22.13
diff --git a/eval/1/iwslt2016_E08L3.09B23.45 b/eval/1/iwslt2016_E08L3.09B23.45
diff --git a/eval/1/iwslt2016_E08L3.28B23.33 b/eval/1/iwslt2016_E08L3.28B23.33
diff --git a/eval/1/iwslt2016_E09L3.15B24.30 b/eval/1/iwslt2016_E09L3.15B24.30
diff --git a/eval/1/iwslt2016_E09L3.25B23.94 b/eval/1/iwslt2016_E09L3.25B23.94
diff --git a/eval/1/iwslt2016_E10L2.97B25.17 b/eval/1/iwslt2016_E10L2.97B25.17
diff --git a/eval/1/iwslt2016_E10L3.00B24.28 b/eval/1/iwslt2016_E10L3.00B24.28
diff --git a/eval/1/iwslt2016_E11L2.85B25.39 b/eval/1/iwslt2016_E11L2.85B25.39
diff --git a/eval/1/iwslt2016_E11L3.11B25.26 b/eval/1/iwslt2016_E11L3.11B25.26
diff --git a/eval/1/iwslt2016_E12L2.95B25.17 b/eval/1/iwslt2016_E12L2.95B25.17
diff --git a/eval/1/iwslt2016_E12L2.96B25.59 b/eval/1/iwslt2016_E12L2.96B25.59
diff --git a/eval/1/iwslt2016_E13L2.86B25.51 b/eval/1/iwslt2016_E13L2.86B25.51
diff --git a/eval/1/iwslt2016_E13L2.93B25.98 b/eval/1/iwslt2016_E13L2.93B25.98
diff --git a/eval/1/iwslt2016_E14L2.74B25.89 b/eval/1/iwslt2016_E14L2.74B25.89
diff --git a/eval/1/iwslt2016_E14L2.75B25.51 b/eval/1/iwslt2016_E14L2.75B25.51
diff --git a/eval/1/iwslt2016_E15L2.76B25.56 b/eval/1/iwslt2016_E15L2.76B25.56
diff --git a/eval/1/iwslt2016_E15L2.78B25.65 b/eval/1/iwslt2016_E15L2.78B25.65
diff --git a/eval/1/iwslt2016_E16L2.78B26.34 b/eval/1/iwslt2016_E16L2.78B26.34
diff --git a/eval/1/iwslt2016_E16L2.84B26.83 b/eval/1/iwslt2016_E16L2.84B26.83
diff --git a/eval/1/iwslt2016_E17L2.69B26.44 b/eval/1/iwslt2016_E17L2.69B26.44
diff --git a/eval/1/iwslt2016_E17L2.78B26.69 b/eval/1/iwslt2016_E17L2.78B26.69
diff --git a/eval/1/iwslt2016_E18L2.67B26.07 b/eval/1/iwslt2016_E18L2.67B26.07
diff --git a/eval/1/iwslt2016_E18L2.76B26.15 b/eval/1/iwslt2016_E18L2.76B26.15
diff --git a/eval/1/iwslt2016_E19L2.62B26.93 b/eval/1/iwslt2016_E19L2.62B26.93
diff --git a/eval/1/iwslt2016_E19L2.70B26.43 b/eval/1/iwslt2016_E19L2.70B26.43
diff --git a/eval/1/iwslt2016_E20L2.61B26.73 b/eval/1/iwslt2016_E20L2.61B26.73
diff --git a/eval/1/iwslt2016_E20L2.62B26.29 b/eval/1/iwslt2016_E20L2.62B26.29
diff --git a/modules.py b/modules.py
@@ -178,9 +178,9 @@ def multihead_attention(queries, keys, values,
     d_model = queries.get_shape().as_list()[-1]
     with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
         # Linear projections
-        Q = tf.layers.dense(queries, d_model) # (N, T_q, d_model)
-        K = tf.layers.dense(keys, d_model) # (N, T_k, d_model)
-        V = tf.layers.dense(values, d_model) # (N, T_k, d_model)
+        Q = tf.layers.dense(queries, d_model, use_bias=False) # (N, T_q, d_model)
+        K = tf.layers.dense(keys, d_model, use_bias=False) # (N, T_k, d_model)
+        V = tf.layers.dense(values, d_model, use_bias=False) # (N, T_k, d_model)
 
         # Split and concat
         Q_ = tf.concat(tf.split(Q, num_heads, axis=2), axis=0) # (h*N, T_q, d_model/h)

diff --git a/test/1/iwslt2016_E17L2.78-26078B22.46 b/test/1/iwslt2016_E17L2.78-26078B22.46
diff --git a/test/1/iwslt2016_E19L2.62-29146B23.16 b/test/1/iwslt2016_E19L2.62-29146B23.16