feat: Added the parameter-free online learning algorithm FreeGrad

test/core.vwtest.json

@@ -4359,5 +4359,32 @@
     "input_files": [
       "train-sets/rcv1_multiclass.dat"
     ]
+  },
+  {
+    "id": 369,
+    "desc": "Freegrad",
+    "vw_command": "--freegrad -d train-sets/0001.dat -f models/freegrad.model -k -c --passes 4",
+    "diff_files": {
+      "stderr": "train-sets/ref/freegrad.stderr"
+    },
+    "input_files": [
+      "train-sets/0001.dat"
+    ]
+  },
+  {
+    "id": 370,
+    "desc": "Freegrad load model",
+    "vw_command": "-d train-sets/0001.dat -i models/freegrad.model -k -t -p freegrad.predict",
+    "diff_files": {
+      "stderr": "train-sets/ref/freegrad_load.stderr",
+      "freegrad.predict": "pred-sets/ref/freegrad.predict"
+    },
+    "input_files": [
+      "train-sets/0001.dat",
+      "models/freegrad.model"
+    ],
+    "depends_on": [
+      369
+    ]
   }
-]
+]

test/pred-sets/ref/freegrad.predict

@@ -0,0 +1,200 @@
+1
+0
+0
+0
+0
+0.965797
+0
+0
+0
+0.782102
+0
+0
+0
+0
+1
+1
+1
+0
+0
+0.038974
+1
+1
+0
+1
+0
+0
+0
+0
+1
+0.377762
+1
+0
+0
+0
+1
+0
+1
+0
+1
+0.797009
+0
+1
+0
+0
+0
+0
+0.028436
+0.010243
+1
+0.357274
+1
+1
+0
+0
+0.901110
+0.017772
+0
+0
+1
+0.170360
+1
+0
+1
+0
+1
+0
+0.002721
+0
+0
+0.627691
+0
+1
+1
+0
+0.994092
+1
+0
+0
+0
+0.332845
+0
+0
+1
+0
+0
+0
+1
+1
+1
+0
+0
+0.988268
+1
+0
+1
+0
+1
+0
+0.856255
+0.356337
+0
+1
+0
+1
+0.005745
+1
+0
+0
+0
+0.469874
+1
+0
+0
+1
+0
+0
+0.820517
+1
+0.992771
+0.126444
+0
+1
+0
+1
+1
+0.972440
+0
+1
+0
+0.634233
+0
+1
+0
+1
+0
+0
+1
+1
+1
+0.214775
+0
+0
+1
+1
+1
+1
+0.966476
+1
+0
+0.156040
+1
+1
+1
+0.216000
+0
+1
+1
+0
+1
+0.258705
+1
+0
+0
+0.947910
+0
+1
+1
+0.033638
+1
+0.800311
+0.868122
+0
+0
+0.984126
+0.005594
+0
+0
+1
+1
+0.710266
+0.985484
+0.024802
+1
+0
+0
+0
+1
+0
+0
+0.633951
+1
+0.030655
+0
+0
+0.006107
+1
+1
+0
+0
+0.167157

test/train-sets/ref/freegrad.stderr

@@ -0,0 +1,36 @@
+final_regressor = models/freegrad.model
+Enabling FreeGrad based optimization
+Algorithm used: FreeGrad
+Num weight bits = 18
+learning rate = 0.5
+initial_t = 0
+power_t = 0.5
+decay_learning_rate = 1
+creating cache_file = train-sets/0001.dat.cache
+Reading datafile = train-sets/0001.dat
+num sources = 1
+Enabled reductions: freegrad, scorer-identity, count_label
+Input label = simple
+Output pred = scalar
+average  since         example        example  current  current  current
+loss     last          counter         weight    label  predict features
+1.000000 1.000000            1            1.0   1.0000   0.0000       51
+0.570071 0.140141            2            2.0   0.0000   0.3744      104
+0.290144 0.010218            4            4.0   0.0000   0.0364      135
+0.270084 0.250024            8            8.0   0.0000   0.0000      146
+0.295007 0.319929           16           16.0   1.0000   0.0000      143
+0.292001 0.288995           32           32.0   1.0000   0.3208       70
+0.275237 0.258473           64           64.0   0.0000   0.5163       34
+0.258924 0.242611          128          128.0   0.0000   0.2918       30
+0.209539 0.209539          256          256.0   0.0000   0.0000       72 h
+0.197807 0.186074          512          512.0   0.0000   0.1012       37 h
+
+finished run
+number of examples per pass = 180
+passes used = 4
+weighted example sum = 720.000000
+weighted label sum = 320.000000
+average loss = 0.159649 h
+best constant = 0.444444
+best constant's loss = 0.246914
+total feature number = 55132

test/train-sets/ref/freegrad_load.stderr

@@ -0,0 +1,33 @@
+only testing
+predictions = freegrad.predict
+Enabling FreeGrad based optimization
+Algorithm used: FreeGrad
+Num weight bits = 18
+learning rate = 0.5
+initial_t = 0
+power_t = 0.5
+using no cache
+Reading datafile = train-sets/0001.dat
+num sources = 1
+Enabled reductions: freegrad, scorer-identity, count_label
+Input label = simple
+Output pred = scalar
+average  since         example        example  current  current  current
+loss     last          counter         weight    label  predict features
+0.000000 0.000000            1            1.0   1.0000   1.0000       51
+0.000000 0.000000            2            2.0   0.0000   0.0000      104
+0.000000 0.000000            4            4.0   0.0000   0.0000      135
+0.000146 0.000292            8            8.0   0.0000   0.0000      146
+0.003041 0.005935           16           16.0   1.0000   1.0000       24
+0.006027 0.009014           32           32.0   0.0000   0.0000       32
+0.006277 0.006528           64           64.0   0.0000   0.0000       61
+0.011065 0.015854          128          128.0   1.0000   1.0000      106
+
+finished run
+number of examples = 200
+weighted example sum = 200.000000
+weighted label sum = 91.000000
+average loss = 0.016990
+best constant = 0.455000
+best constant's loss = 0.247975
+total feature number = 15482

test/train-sets/ref/help.stdout

@@ -486,6 +486,14 @@ Explore Evaluation Options:
 Debug: Metrics Options:
   --extra_metrics arg   Specify filename to write metrics to. Note: There is no
                         fixed schema
+FreeGrad Options:
+  --freegrad            Diagonal FreeGrad Algorithm
+  --restart             Use the FreeRange restarts
+  --project             Project the outputs to adapt to both the lipschitz and 
+                        comparator norm
+  --radius arg          Radius of the l2-ball for the projection. If not 
+                        supplied, an adaptive radius will be used.
+  --fepsilon arg (=1, ) Initial wealth
 Follow the Regularized Leader Options:
   --ftrl                FTRL: Follow the Proximal Regularized Leader
   --coin                Coin betting optimizer

vowpalwabbit/CMakeLists.txt

@@ -113,6 +113,7 @@ set(vw_all_headers
   fast_pow10.h
   feature_group.h
   ftrl.h
+  freegrad.h
   gd_mf.h
   gd_predict.h
   gd.h
@@ -296,6 +297,7 @@ set(vw_all_sources
   explore_eval.cc
   feature_group.cc
   ftrl.cc
+  freegrad.cc
   gd_mf.cc
   gd.cc
   gen_cs_example.cc