Skip to content

Commit

Permalink
Shared example merge reduction (#1873)
Browse files Browse the repository at this point in the history 
* pre-merge and remove shared features before sending examples down the reduction stack

* copy-paste error in vcxproj file

* fixing bad iterator dereference

* Enable Version and Tag override for NuGet pack (#1870)

* First attempt at softmax learner for cbadf (#1839)

* First attempt at softmax learner for cbadf

* some bugfixes and proper handling of shared and non-shared examples

* add test for cb_adf softmax

* Removing redundant safe_probability definition

* Update gd_mf_weights.cc right namespace (#1874)

* Update gd_mf_weights.cc

Use right indices to write out features for the right namespace, correcting bug found in #1871

* Update gd_mf_weights.cc

Add back rank to the right quadratic calculation

* Fixed warning message ips -> mtr (#1875)

* Fix dsjson parser regression and add smoke-test (#1878)

* Fix dsjson parser regression and add smoke-test

Note: The test is ignored under the C# Unit Tests, because the test
driver is not set up to run the native parser. This is an area we should
improve outside of the bugfix.

* Suppress test for softmax learner

* fixing bugs, finding/fixing wrong tests

* fixing finish_example, unit tests

* cleaning up CB reductions

* copying predictions up and down the stack

* removing shared example in place. Disabling finish_example

* modify unit tests to match outputs. Add patch file to undo this change later

* moving patch file to test directory

* Removing unused variables, function
  • Loading branch information
@peterychang @JohnLangford
peterychang authored and JohnLangford committed Jun 3, 2019
1 parent 9206466 commit 5d4676c
Show file tree
Hide file tree
Showing 32 changed files with 648 additions and 343 deletions.
319 changes: 319 additions & 0 deletions test/fix_cbtest_feature_count.patch
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,319 @@
diff --git b/test/train-sets/ref/cb_adf_crash2.stderr a/test/train-sets/ref/cb_adf_crash2.stderr
index 2a6a7c6c..8784d074 100644
--- b/test/train-sets/ref/cb_adf_crash2.stderr
+++ a/test/train-sets/ref/cb_adf_crash2.stderr
@@ -8,11 +8,11 @@ Reading datafile = train-sets/cb_adf_crash_2.data
num sources = 1
average since example example current current current
loss last counter weight label predict features
- n.a. n.a. 1 1.0 unknown 0:1... 2
+ n.a. n.a. 1 1.0 unknown 0:1... 3

finished run
number of examples = 1
weighted example sum = 1.000000
weighted label sum = 0.000000
average loss = n.a.
-total feature number = 4
+total feature number = 3
diff --git b/test/train-sets/ref/cb_adf_dm.stderr a/test/train-sets/ref/cb_adf_dm.stderr
index 5ee390d3..d790ee21 100644
--- b/test/train-sets/ref/cb_adf_dm.stderr
+++ a/test/train-sets/ref/cb_adf_dm.stderr
@@ -8,7 +8,7 @@ Reading datafile = train-sets/cb_test.ldf
num sources = 1
average since example example current current current
loss last counter weight label predict features
-2.000000 2.000000 1 1.0 known 0:0... 12
+2.000000 2.000000 1 1.0 known 0:0... 18
1.000000 0.000000 2 2.0 known 1:-0.0849907... 8

finished run
@@ -16,4 +16,4 @@ number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 1.000000
-total feature number = 31
+total feature number = 34
diff --git b/test/train-sets/ref/cb_adf_dr.stderr a/test/train-sets/ref/cb_adf_dr.stderr
index 13366ba5..265c65e4 100644
--- b/test/train-sets/ref/cb_adf_dr.stderr
+++ a/test/train-sets/ref/cb_adf_dr.stderr
@@ -8,7 +8,7 @@ Reading datafile = train-sets/cb_test.ldf
num sources = 1
average since example example current current current
loss last counter weight label predict features
-2.000000 2.000000 1 1.0 known 0:0... 12
+2.000000 2.000000 1 1.0 known 0:0... 18
0.931417 -0.137166 2 2.0 known 1:0.121728... 8

finished run
@@ -16,4 +16,4 @@ number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 0.931417
-total feature number = 31
+total feature number = 34
diff --git b/test/train-sets/ref/cb_adf_mtr.stderr a/test/train-sets/ref/cb_adf_mtr.stderr
index f0759d88..b7401476 100644
--- b/test/train-sets/ref/cb_adf_mtr.stderr
+++ a/test/train-sets/ref/cb_adf_mtr.stderr
@@ -7,7 +7,7 @@ Reading datafile = train-sets/cb_test.ldf
num sources = 1
average since example example current current current
loss last counter weight label predict features
-2.000000 2.000000 1 1.0 known 0:0... 9
+2.000000 2.000000 1 1.0 known 0:0... 15
1.000000 0.000000 2 2.0 known 1:0... 6

finished run
@@ -15,4 +15,4 @@ number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 1.000000
-total feature number = 23
+total feature number = 27
diff --git b/test/train-sets/ref/cb_adf_rank.stderr a/test/train-sets/ref/cb_adf_rank.stderr
index 1e9968e0..a7d50aac 100644
--- b/test/train-sets/ref/cb_adf_rank.stderr
+++ a/test/train-sets/ref/cb_adf_rank.stderr
@@ -8,7 +8,7 @@ Reading datafile = train-sets/cb_test.ldf
num sources = 1
average since example example current current current
loss last counter weight label predict features
-2.000000 2.000000 1 1.0 known 0:0... 9
+2.000000 2.000000 1 1.0 known 0:0... 15
1.000000 0.000000 2 2.0 known 1:0... 6

finished run
@@ -16,4 +16,4 @@ number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 1.000000
-total feature number = 23
+total feature number = 27
diff --git b/test/train-sets/ref/cb_adf_sm.stderr a/test/train-sets/ref/cb_adf_sm.stderr
index cfd35a18..55dc63a1 100644
--- b/test/train-sets/ref/cb_adf_sm.stderr
+++ a/test/train-sets/ref/cb_adf_sm.stderr
@@ -8,13 +8,13 @@ Reading datafile = train-sets/cb_adf_sm.data
num sources = 1
average since example example current current current
loss last counter weight label predict features
-6.000000 6.000000 1 1.0 known 0:0... 9
-3.000000 0.000000 2 2.0 known 2:-0.532113... 9
-1.500000 0.000000 4 4.0 known 2:-0.165079... 12
+6.000000 6.000000 1 1.0 known 0:0... 12
+3.000000 0.000000 2 2.0 known 2:-0.532113... 12
+1.500000 0.000000 4 4.0 known 2:-0.165079... 16

finished run
number of examples = 4
weighted example sum = 4.000000
weighted label sum = 0.000000
average loss = 1.500000
-total feature number = 47
+total feature number = 52
diff --git b/test/train-sets/ref/cbe_adf_bag.stderr a/test/train-sets/ref/cbe_adf_bag.stderr
index bf60365f..967dd62a 100644
--- b/test/train-sets/ref/cbe_adf_bag.stderr
+++ a/test/train-sets/ref/cbe_adf_bag.stderr
@@ -8,7 +8,7 @@ Reading datafile = train-sets/cb_test.ldf
num sources = 1
average since example example current current current
loss last counter weight label predict features
-0.666667 0.666667 1 1.0 known 0:0.333333... 9
+0.666667 0.666667 1 1.0 known 0:0.333333... 15
0.333333 0.000000 2 2.0 known 0:0.5... 6

finished run
@@ -16,4 +16,4 @@ number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 0.333333
-total feature number = 23
+total feature number = 27
diff --git b/test/train-sets/ref/cbe_adf_cover.stderr a/test/train-sets/ref/cbe_adf_cover.stderr
index 6880a093..3035d093 100644
--- b/test/train-sets/ref/cbe_adf_cover.stderr
+++ a/test/train-sets/ref/cbe_adf_cover.stderr
@@ -9,7 +9,7 @@ Reading datafile = train-sets/cb_test.ldf
num sources = 1
average since example example current current current
loss last counter weight label predict features
-0.666667 0.666667 1 1.0 known 0:0.333333... 9
+0.666667 0.666667 1 1.0 known 0:0.333333... 15
0.333333 0.000000 2 2.0 known 0:0.5... 6

finished run
@@ -17,4 +17,4 @@ number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 0.333333
-total feature number = 23
+total feature number = 27
diff --git b/test/train-sets/ref/cbe_adf_cover_dr.json.stderr a/test/train-sets/ref/cbe_adf_cover_dr.json.stderr
index b351a486..818a5cc4 100644
--- b/test/train-sets/ref/cbe_adf_cover_dr.json.stderr
+++ a/test/train-sets/ref/cbe_adf_cover_dr.json.stderr
@@ -8,7 +8,7 @@ Reading datafile = train-sets/cb_test.json
num sources = 1
average since example example current current current
loss last counter weight label predict features
-0.666667 0.666667 1 1.0 known 0:0.333333... 9
+0.666667 0.666667 1 1.0 known 0:0.333333... 15
0.333333 0.000000 2 2.0 known 0:0.5... 6

finished run
@@ -16,4 +16,4 @@ number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 0.333333
-total feature number = 23
+total feature number = 27
diff --git b/test/train-sets/ref/cbe_adf_cover_dr.stderr a/test/train-sets/ref/cbe_adf_cover_dr.stderr
index 08b1f32a..979eb436 100644
--- b/test/train-sets/ref/cbe_adf_cover_dr.stderr
+++ a/test/train-sets/ref/cbe_adf_cover_dr.stderr
@@ -8,7 +8,7 @@ Reading datafile = train-sets/cb_test.ldf
num sources = 1
average since example example current current current
loss last counter weight label predict features
-0.666667 0.666667 1 1.0 known 0:0.333333... 9
+0.666667 0.666667 1 1.0 known 0:0.333333... 15
0.333333 0.000000 2 2.0 known 0:0.5... 6

finished run
@@ -16,4 +16,4 @@ number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 0.333333
-total feature number = 23
+total feature number = 27
diff --git b/test/train-sets/ref/cbe_adf_cover_dr256.json.stderr a/test/train-sets/ref/cbe_adf_cover_dr256.json.stderr
index a4909f33..78196105 100644
--- b/test/train-sets/ref/cbe_adf_cover_dr256.json.stderr
+++ a/test/train-sets/ref/cbe_adf_cover_dr256.json.stderr
@@ -8,7 +8,7 @@ Reading datafile = train-sets/cb_test256.json
num sources = 1
average since example example current current current
loss last counter weight label predict features
-0.666667 0.666667 1 1.0 known 0:0.333333... 9
+0.666667 0.666667 1 1.0 known 0:0.333333... 15
0.333333 0.000000 2 2.0 known 0:0.5... 6
0.333333 0.333333 4 4.0 known 1:0.591752... 6
0.297761 0.262189 8 8.0 known 1:0.666667... 6
@@ -23,4 +23,4 @@ number of examples = 260
weighted example sum = 260.000000
weighted label sum = 0.000000
average loss = 0.068875
-total feature number = 2210
+total feature number = 2730
diff --git b/test/train-sets/ref/cbe_adf_dsjson.stderr a/test/train-sets/ref/cbe_adf_dsjson.stderr
index abef6034..ee6fca6e 100644
--- b/test/train-sets/ref/cbe_adf_dsjson.stderr
+++ a/test/train-sets/ref/cbe_adf_dsjson.stderr
@@ -9,13 +9,13 @@ Reading datafile = train-sets/decisionservice.json
num sources = 1
average since example example current current current
loss last counter weight label predict features
--0.102041 -0.102041 1 1.0 known 0:0.0833333... 361
--0.051020 0.000000 2 2.0 known 6:0.816667... 361
--0.040816 -0.020408 3 3.0 known 6:0.816667... 361
+-0.102041 -0.102041 1 1.0 known 0:0.0833333... 433
+-0.051020 0.000000 2 2.0 known 6:0.816667... 433
+-0.040816 -0.020408 3 3.0 known 6:0.816667... 433

finished run
number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = -0.040816
-total feature number = 1104
+total feature number = 1299
diff --git b/test/train-sets/ref/cbe_adf_epsilon.stderr a/test/train-sets/ref/cbe_adf_epsilon.stderr
index 98e59395..dacf6bfc 100644
--- b/test/train-sets/ref/cbe_adf_epsilon.stderr
+++ a/test/train-sets/ref/cbe_adf_epsilon.stderr
@@ -8,7 +8,7 @@ Reading datafile = train-sets/cb_test.ldf
num sources = 1
average since example example current current current
loss last counter weight label predict features
-0.666667 0.666667 1 1.0 known 0:0.333333... 9
+0.666667 0.666667 1 1.0 known 0:0.333333... 15
0.333333 0.000000 2 2.0 known 1:0.95... 6

finished run
@@ -16,4 +16,4 @@ number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 0.333333
-total feature number = 23
+total feature number = 27
diff --git b/test/train-sets/ref/cbe_adf_first.stderr a/test/train-sets/ref/cbe_adf_first.stderr
index af2100bb..35158afb 100644
--- b/test/train-sets/ref/cbe_adf_first.stderr
+++ a/test/train-sets/ref/cbe_adf_first.stderr
@@ -8,7 +8,7 @@ Reading datafile = train-sets/cb_test.ldf
num sources = 1
average since example example current current current
loss last counter weight label predict features
-0.666667 0.666667 1 1.0 known 0:0.333333... 9
+0.666667 0.666667 1 1.0 known 0:0.333333... 15
0.333333 0.000000 2 2.0 known 1:0.5... 6

finished run
@@ -16,4 +16,4 @@ number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 0.333333
-total feature number = 23
+total feature number = 27
diff --git b/test/train-sets/ref/cbe_adf_softmax.stderr a/test/train-sets/ref/cbe_adf_softmax.stderr
index 543bc037..3ea3464b 100644
--- b/test/train-sets/ref/cbe_adf_softmax.stderr
+++ a/test/train-sets/ref/cbe_adf_softmax.stderr
@@ -8,7 +8,7 @@ Reading datafile = train-sets/cb_test.ldf
num sources = 1
average since example example current current current
loss last counter weight label predict features
-0.666667 0.666667 1 1.0 known 0:0.333333... 9
+0.666667 0.666667 1 1.0 known 0:0.333333... 15
0.333333 0.000000 2 2.0 known 1:0.559575... 6

finished run
@@ -16,4 +16,4 @@ number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 0.333333
-total feature number = 23
+total feature number = 27
diff --git b/test/train-sets/ref/no_shared_features.stderr a/test/train-sets/ref/no_shared_features.stderr
index 23c135ee..83a007ce 100644
--- b/test/train-sets/ref/no_shared_features.stderr
+++ a/test/train-sets/ref/no_shared_features.stderr
@@ -14,4 +14,4 @@ number of examples = 1
weighted example sum = 1.000000
weighted label sum = 0.000000
average loss = 0.000000
-total feature number = 9
+total feature number = 8
diff --git b/test/train-sets/ref/sparse.stderr a/test/train-sets/ref/sparse.stderr
index 5ee390d3..d790ee21 100644
--- b/test/train-sets/ref/sparse.stderr
+++ a/test/train-sets/ref/sparse.stderr
@@ -8,7 +8,7 @@ Reading datafile = train-sets/cb_test.ldf
num sources = 1
average since example example current current current
loss last counter weight label predict features
-2.000000 2.000000 1 1.0 known 0:0... 12
+2.000000 2.000000 1 1.0 known 0:0... 18
1.000000 0.000000 2 2.0 known 1:-0.0849907... 8

finished run
@@ -16,4 +16,4 @@ number of examples = 3
weighted example sum = 3.000000
weighted label sum = 0.000000
average loss = 1.000000
-total feature number = 31
+total feature number = 34
4 changes: 2 additions & 2 deletions test/pred-sets/ref/cb_adf_dr.predict
View file
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
0:0,1:0,2:0

1:-0.0207921,0:0.477788
1:0.121728,0:0.457171

1:-0.0844552,0:0.648064
1:0.09676,0:0.596982

Loading

0 comments on commit 5d4676c

Please sign in to comment.