feat: enable 0-indexed labels for csoaa (#3533)

* feat: enable 0-indexed labels for csoaa * clang * address comments * info output * 177
VowpalWabbit · Dec 16, 2021 · 471c4f0 · 471c4f0
1 parent 47cf09a
commit 471c4f0
Show file tree

Hide file tree

Showing 13 changed files with 206 additions and 15 deletions.
diff --git a/test/core.vwtest.json b/test/core.vwtest.json
@@ -4440,6 +4440,41 @@
     "depends_on": [
       373
     ]
+  },
+  {
+    "id": 375,
+    "desc": "Use csoaa with 1-indexing flag",
+    "vw_command": "-d train-sets/csoaa_1_ind.dat --csoaa 4 -p csoaa_1_ind.predict --indexing 1",
+    "diff_files": {
+      "stderr": "train-sets/ref/csoaa_1_ind.stderr",
+      "csoaa_1_ind.predict": "pred-sets/ref/csoaa_1_ind.predict"
+    },
+    "input_files": [
+      "train-sets/csoaa_1_ind.dat"
+    ]
+  },
+  {
+    "id": 376,
+    "desc": "Use csoaa with auto-detect 1-indexing",
+    "vw_command": "-d train-sets/csoaa_1_ind.dat --csoaa 4 -p csoaa_1_ind_auto.predict",
+    "diff_files": {
+      "stderr": "train-sets/ref/csoaa_1_ind_auto.stderr",
+      "csoaa_1_ind_auto.predict": "pred-sets/ref/csoaa_1_ind.predict"
+    },
+    "input_files": [
+      "train-sets/csoaa_1_ind.dat"
+    ]
+  },
+  {
+    "id": 377,
+    "desc": "Use csoaa with on 1-indexed data set with 0-indexing flag",
+    "vw_command": "-d train-sets/csoaa_1_ind.dat --csoaa 4 --indexing 0",
+    "diff_files": {
+      "stderr": "train-sets/ref/csoaa_1_ind_0_flag.stderr",
+      "stdout": "train-sets/ref/csoaa_1_ind_0_flag.stdout"
+    },
+    "input_files": [
+      "train-sets/csoaa_1_ind.dat"
+    ]
   }
-
 ]
diff --git a/test/pred-sets/ref/csoaa_1_ind.predict b/test/pred-sets/ref/csoaa_1_ind.predict
@@ -0,0 +1,5 @@
+1
+1
+1
+3
+4
diff --git a/test/test-sets/ref/backwards.stdout b/test/test-sets/ref/backwards.stdout
@@ -1,3 +1,4 @@
+[info] label 3 found -- labels are now considered 1-indexed.
 Id 
 Min label:-1
 Max label:1

diff --git a/test/train-sets/csoaa_1_ind.dat b/test/train-sets/csoaa_1_ind.dat
@@ -0,0 +1,5 @@
+1:0 2:1 3:1 4:1 | a b c
+1:1 2:1 3:0 4:1 | b c d
+1:0 2:1 3:1 4:1 | a c e
+1:1 2:1 3:1 4:0 | b d f
+1:1 2:0 3:1 4:1 | d e f
diff --git a/test/train-sets/ref/audit2.stdout b/test/train-sets/ref/audit2.stdout
@@ -4,6 +4,7 @@
 	f^a:57421:1:0@0	f^b:62865:1:0@0	f^c:228993:1:0@0	e^x:125153:1:0@0	e^y:27665:1:0@0	e^z:259285:1:0@0	Constant:202097:1:0@0	e^x*f^a:96493:1:0@0	e^x*f^b:93489:1:0@0	e^x*f^c:189985:1:0@0	e^y*f^a:195965:1:0@0	e^y*f^b:190625:1:0@0	e^y*f^c:91057:1:0@0	e^z*f^a:166385:1:0@0	e^z*f^b:171053:1:0@0	e^z*f^c:71485:1:0@0
 0
 	f^a:57422:1:0@0	f^b:62866:1:0@0	f^c:228994:1:0@0	e^x:125154:1:0@0	e^y:27666:1:0@0	e^z:259286:1:0@0	Constant:202098:1:0@0	e^x*f^a:96494:1:0@0	e^x*f^b:93490:1:0@0	e^x*f^c:189986:1:0@0	e^y*f^a:195966:1:0@0	e^y*f^b:190626:1:0@0	e^y*f^c:91058:1:0@0	e^z*f^a:166386:1:0@0	e^z*f^b:171054:1:0@0	e^z*f^c:71486:1:0@0
+[info] label 3 found -- labels are now considered 1-indexed.
 -0.324225
 	f^a:57420:1:-0.0540374@4	f^c:228992:1:-0.0540374@4	e^x:125152:1:-0.0540374@4	Constant:202096:1:-0.0540374@4	e^x*f^a:96492:1:-0.0540374@4	e^x*f^c:189984:1:-0.0540374@4	f^y:154732:1:0@0	e^x*f^y:246988:1:0@0
 -0.324225

diff --git a/test/train-sets/ref/csoaa_1_ind.stderr b/test/train-sets/ref/csoaa_1_ind.stderr
@@ -0,0 +1,23 @@
+predictions = csoaa_1_ind.predict
+Num weight bits = 18
+learning rate = 0.5
+initial_t = 0
+power_t = 0.5
+using no cache
+Reading datafile = train-sets/csoaa_1_ind.dat
+num sources = 1
+Enabled reductions: gd, scorer-identity, csoaa
+Input label = cs
+Output pred = multiclass
+average  since         example        example  current  current  current
+loss     last          counter         weight    label  predict features
+0.000000 0.000000            1            1.0    known        1        4
+0.500000 1.000000            2            2.0    known        1        4
+0.500000 0.500000            4            4.0    known        3        4
+
+finished run
+number of examples = 5
+weighted example sum = 5.000000
+weighted label sum = 0.000000
+average loss = 0.600000
+total feature number = 20
diff --git a/test/train-sets/ref/csoaa_1_ind_0_flag.stderr b/test/train-sets/ref/csoaa_1_ind_0_flag.stderr
@@ -0,0 +1,22 @@
+Num weight bits = 18
+learning rate = 0.5
+initial_t = 0
+power_t = 0.5
+using no cache
+Reading datafile = train-sets/csoaa_1_ind.dat
+num sources = 1
+Enabled reductions: gd, scorer-identity, csoaa
+Input label = cs
+Output pred = multiclass
+average  since         example        example  current  current  current
+loss     last          counter         weight    label  predict features
+1.000000 1.000000            1            1.0    known        0        4
+1.000000 1.000000            2            2.0    known        1        4
+0.750000 0.500000            4            4.0    known        3        4
+
+finished run
+number of examples = 5
+weighted example sum = 5.000000
+weighted label sum = 0.000000
+average loss = 0.800000
+total feature number = 20
diff --git a/test/train-sets/ref/csoaa_1_ind_0_flag.stdout b/test/train-sets/ref/csoaa_1_ind_0_flag.stdout
@@ -0,0 +1,5 @@
+[warning] label 4 is not in {0,3}. This won't work for 0-indexed actions.
+[warning] label 4 is not in {0,3}. This won't work for 0-indexed actions.
+[warning] label 4 is not in {0,3}. This won't work for 0-indexed actions.
+[warning] label 4 is not in {0,3}. This won't work for 0-indexed actions.
+[warning] label 4 is not in {0,3}. This won't work for 0-indexed actions.
diff --git a/test/train-sets/ref/csoaa_1_ind_auto.stderr b/test/train-sets/ref/csoaa_1_ind_auto.stderr
@@ -0,0 +1,23 @@
+predictions = csoaa_1_ind_auto.predict
+Num weight bits = 18
+learning rate = 0.5
+initial_t = 0
+power_t = 0.5
+using no cache
+Reading datafile = train-sets/csoaa_1_ind.dat
+num sources = 1
+Enabled reductions: gd, scorer-identity, csoaa
+Input label = cs
+Output pred = multiclass
+average  since         example        example  current  current  current
+loss     last          counter         weight    label  predict features
+0.000000 0.000000            1            1.0    known        1        4
+0.500000 1.000000            2            2.0    known        1        4
+0.500000 0.500000            4            4.0    known        3        4
+
+finished run
+number of examples = 5
+weighted example sum = 5.000000
+weighted label sum = 0.000000
+average loss = 0.600000
+total feature number = 20
diff --git a/test/train-sets/ref/help.stdout b/test/train-sets/ref/help.stdout
@@ -467,6 +467,7 @@ Cost Sensitive Active Learning Options:
   --csa_debug                           Print debug stuff for cs_active
 Cost Sensitive One Against All Options:
   --csoaa arg           One-against-all multiclass with <k> costs
+  --indexing arg        Choose between 0 or 1-indexing. Choices: {0, 1}
 Cost Sensitive One Against All with Label Dependent Features Options:
   --csoaa_ldf arg       Use one-against-all multiclass learning with label 
                         dependent features

diff --git a/test/train-sets/ref/oaa_mixed_probabilities.stdout b/test/train-sets/ref/oaa_mixed_probabilities.stdout
@@ -1 +1,2 @@
+[info] label 0 found -- labels are now considered 0-indexed.
 [warning] label 4 is not in {0,3}. This won't work for 0-indexed actions.
diff --git a/vowpalwabbit/csoaa.cc b/vowpalwabbit/csoaa.cc
@@ -21,22 +21,34 @@ namespace CSOAA
 struct csoaa
 {
   uint32_t num_classes = 0;
+  int indexing = -1;
+  bool search = false;
   polyprediction* pred = nullptr;
   ~csoaa() { free(pred); }
 };
 
 template <bool is_learn>
 inline void inner_loop(single_learner& base, example& ec, uint32_t i, float cost, uint32_t& prediction, float& score,
-    float& partial_prediction)
+    float& partial_prediction, int indexing)
 {
   if (is_learn)
   {
     ec.weight = (cost == FLT_MAX) ? 0.f : 1.f;
     ec.l.simple.label = cost;
-    base.learn(ec, i - 1);
+    if (indexing == 0) { base.learn(ec, i); }
+    else
+    {
+      base.learn(ec, i - 1);
+    }
   }
   else
-    base.predict(ec, i - 1);
+  {
+    if (indexing == 0) { base.predict(ec, i); }
+    else
+    {
+      base.predict(ec, i - 1);
+    }
+  }
 
   partial_prediction = ec.partial_prediction;
   if (ec.partial_prediction < score || (ec.partial_prediction == score && i < prediction))
@@ -52,12 +64,44 @@ inline void inner_loop(single_learner& base, example& ec, uint32_t i, float cost
 template <bool is_learn>
 void predict_or_learn(csoaa& c, single_learner& base, example& ec)
 {
+  if (!c.search)
+  {
+    for (auto& cost : ec.l.cs.costs)
+    {
+      auto& lbl = cost.class_index;
+      // Update indexing
+      if (c.indexing == -1 && lbl == 0)
+      {
+        logger::log_info("label 0 found -- labels are now considered 0-indexed.");
+        c.indexing = 0;
+      }
+      else if (c.indexing == -1 && lbl == c.num_classes)
+      {
+        logger::log_info("label {0} found -- labels are now considered 1-indexed.", c.num_classes);
+        c.indexing = 1;
+      }
+
+      // Label validation
+      if (c.indexing == 0 && lbl >= c.num_classes)
+      {
+        logger::log_warn(
+            "label {0} is not in {{0,{1}}}. This won't work for 0-indexed actions.", lbl, c.num_classes - 1);
+        lbl = 0;
+      }
+      else if (c.indexing == 1 && (lbl < 1 || lbl > c.num_classes))
+      {
+        logger::log_warn("label {0} is not in {{1,{1}}}. This won't work for 1-indexed actions.", lbl, c.num_classes);
+        lbl = static_cast<uint32_t>(c.num_classes);
+      }
+    }
+  }
+
   COST_SENSITIVE::label ld = std::move(ec.l.cs);
 
   // Guard example state restore against throws
   auto restore_guard = VW::scope_exit([&ld, &ec] { ec.l.cs = std::move(ld); });
 
-  uint32_t prediction = 1;
+  uint32_t prediction = (c.indexing == 0) ? 0 : 1;
   float score = FLT_MAX;
   size_t pt_start = ec.passthrough ? ec.passthrough->size() : 0;
   ec.l.simple = {0.};
@@ -68,7 +112,7 @@ void predict_or_learn(csoaa& c, single_learner& base, example& ec)
   if (!ld.costs.empty())
   {
     for (auto& cl : ld.costs)
-      inner_loop<is_learn>(base, ec, cl.class_index, cl.x, prediction, score, cl.partial_prediction);
+      inner_loop<is_learn>(base, ec, cl.class_index, cl.x, prediction, score, cl.partial_prediction, c.indexing);
     ec.partial_prediction = score;
   }
   else if (dont_learn)
@@ -77,17 +121,30 @@ void predict_or_learn(csoaa& c, single_learner& base, example& ec)
     ec._reduction_features.template get<simple_label_reduction_features>().reset_to_default();
 
     base.multipredict(ec, 0, c.num_classes, c.pred, false);
-    for (uint32_t i = 1; i <= c.num_classes; i++)
+    if (c.indexing == 0)
     {
-      add_passthrough_feature(ec, i, c.pred[i - 1].scalar);
-      if (c.pred[i - 1].scalar < c.pred[prediction - 1].scalar) prediction = i;
+      for (uint32_t i = 0; i <= c.num_classes; i++)
+      {
+        add_passthrough_feature(ec, i, c.pred[i].scalar);
+        if (c.pred[i].scalar < c.pred[prediction].scalar) prediction = i;
+      }
+      ec.partial_prediction = c.pred[prediction].scalar;
+    }
+    else
+    {
+      for (uint32_t i = 1; i <= c.num_classes; i++)
+      {
+        add_passthrough_feature(ec, i, c.pred[i - 1].scalar);
+        if (c.pred[i - 1].scalar < c.pred[prediction - 1].scalar) prediction = i;
+      }
+      ec.partial_prediction = c.pred[prediction - 1].scalar;
     }
-    ec.partial_prediction = c.pred[prediction - 1].scalar;
   }
   else
   {
     float temp;
-    for (uint32_t i = 1; i <= c.num_classes; i++) inner_loop<false>(base, ec, i, FLT_MAX, prediction, score, temp);
+    for (uint32_t i = 1; i <= c.num_classes; i++)
+      inner_loop<false>(base, ec, i, FLT_MAX, prediction, score, temp, c.indexing);
   }
 
   if (ec.passthrough)
@@ -124,13 +181,15 @@ base_learner* csoaa_setup(VW::setup_base_i& stack_builder)
   VW::workspace& all = *stack_builder.get_all_pointer();
   auto c = VW::make_unique<csoaa>();
   option_group_definition new_options("Cost Sensitive One Against All");
-  new_options.add(
-      make_option("csoaa", c->num_classes).keep().necessary().help("One-against-all multiclass with <k> costs"));
+  new_options
+      .add(make_option("csoaa", c->num_classes).keep().necessary().help("One-against-all multiclass with <k> costs"))
+      .add(make_option("indexing", c->indexing).one_of({0, 1}).keep().help("Choose between 0 or 1-indexing"));
 
   if (!options.add_parse_and_check_necessary(new_options)) return nullptr;
 
   if (options.was_supplied("probabilities"))
   { THROW("Error: csoaa does not support probabilities flag, please use oaa or multilabel_oaa"); }
+  c->search = options.was_supplied("search");
 
   c->pred = calloc_or_throw<polyprediction>(c->num_classes);
   size_t ws = c->num_classes;

diff --git a/vowpalwabbit/oaa.cc b/vowpalwabbit/oaa.cc
@@ -40,9 +40,14 @@ struct oaa
 void learn_randomized(oaa& o, VW::LEARNER::single_learner& base, example& ec)
 {
   // Update indexing
-  if (o.indexing == -1 && ec.l.multi.label == 0) { o.indexing = 0; }
+  if (o.indexing == -1 && ec.l.multi.label == 0)
+  {
+    logger::log_info("label 0 found -- labels are now considered 0-indexed.");
+    o.indexing = 0;
+  }
   else if (o.indexing == -1 && ec.l.multi.label == o.k)
   {
+    logger::log_info("label {0} found -- labels are now considered 1-indexed.", o.k);
     o.indexing = 1;
   }
 
@@ -98,9 +103,14 @@ template <bool print_all, bool scores, bool probabilities>
 void learn(oaa& o, VW::LEARNER::single_learner& base, example& ec)
 {
   // Update indexing
-  if (o.indexing == -1 && ec.l.multi.label == 0) { o.indexing = 0; }
+  if (o.indexing == -1 && ec.l.multi.label == 0)
+  {
+    logger::log_info("label 0 found -- labels are now considered 0-indexed.");
+    o.indexing = 0;
+  }
   else if (o.indexing == -1 && ec.l.multi.label == o.k)
   {
+    logger::log_info("label {0} found -- labels are now considered 1-indexed.", o.k);
     o.indexing = 1;
   }