feat(torch): add multilabel classification

jolibrain · Aug 12, 2022 · 90d536e · 90d536e
1 parent 24aa37c
commit 90d536e
Show file tree

Hide file tree

Showing 5 changed files with 61 additions and 1 deletion.
diff --git a/docs/api.md b/docs/api.md
@@ -1124,6 +1124,7 @@ gpu           | bool   | yes      | false   | Whether to use GPU
 gpuid         | int or array | yes | 0      | GPU id, use single int for single GPU, `-1` for using all GPUs, and array e.g. `[1,3]` for selecting among multiple GPUs
 extract_layer | string | yes      | ""      | Returns tensor values from intermediate layers. In bert models "hidden_state" allows to extract raw hidden_states values to return as output. If set to 'last', simply returns the tensor values from last layer.
 forward_method | string | yes | "" | Executes a custom function from within a traced/JIT model, instead of the standard forward()
+multi_label | bool | yes | false | Model outputs an independent score for each class
 
 
 - XGBoost

diff --git a/src/backends/torch/torchlib.cc b/src/backends/torch/torchlib.cc
@@ -82,6 +82,7 @@ namespace dd
     _bbox = tl._bbox;
     _segmentation = tl._segmentation;
     _ctc = tl._ctc;
+    _multi_label = tl._multi_label;
     _loss = tl._loss;
     _template_params = tl._template_params;
     _dtype = tl._dtype;
@@ -295,6 +296,11 @@ namespace dd
         _regression = true;
       }
 
+    if (mllib_dto->multi_label)
+      {
+        _multi_label = true;
+      }
+
     if (_template == "bert")
       {
         if (!self_supervised.empty())
@@ -1500,7 +1506,12 @@ namespace dd
                 // XXX: why is (!_timeserie) needed here? Aren't _timeserie and
                 // _classification mutually exclusive?
                 if (extract_layer.empty() && !_timeserie && _classification)
-                  output = torch::softmax(output, 1).to(cpu);
+                  {
+                    if (_multi_label)
+                      output = torch::sigmoid(output).to(cpu);
+                    else
+                      output = torch::softmax(output, 1).to(cpu);
+                  }
                 else if (extract_layer.empty() && !_timeserie && ctc)
                   output = torch::softmax(output, 2).to(cpu);
                 else

diff --git a/src/backends/torch/torchlib.h b/src/backends/torch/torchlib.h
@@ -110,6 +110,7 @@ namespace dd
     bool _bbox = false;           /**< select detection type problem */
     bool _segmentation = false;   /**< select segmentation type problem */
     bool _ctc = false;            /**< select OCR type problem */
+    bool _multi_label = false;    /**< whether model outputs multiple labels */
     std::string _loss = "";       /**< selected loss*/
     double _reg_weight
         = 1; /**< for detection models, weight for bbox regression loss. */

diff --git a/src/dto/mllib.hpp b/src/dto/mllib.hpp
@@ -176,6 +176,13 @@ namespace dd
       };
       DTO_FIELD(UnorderedFields<Any>, template_params);
 
+      DTO_FIELD_INFO(multi_label)
+      {
+        info->description
+            = "Model outputs an independent score for each class";
+      }
+      DTO_FIELD(Boolean, multi_label) = false;
+
       // Libtorch predict options
       DTO_FIELD_INFO(forward_method)
       {

diff --git a/tests/ut-torchapi.cc b/tests/ut-torchapi.cc
@@ -238,6 +238,46 @@ TEST(torchapi, service_predict_native_bw)
   fileops::remove_dir(native_resnet_repo);
 }
 
+TEST(torchapi, service_predict_multi_label)
+{
+  // create service
+  JsonAPI japi;
+  std::string sname = "imgserv";
+  std::string jstr
+      = "{\"mllib\":\"torch\",\"description\":\"resnet-50\",\"type\":"
+        "\"supervised\",\"model\":{\"repository\":\""
+        + incept_repo
+        + "\"},\"parameters\":{\"input\":{\"connector\":\"image\",\"height\":"
+          "224,\"width\":224,\"rgb\":true,\"scale\":0.0039},\"mllib\":{"
+          "\"nclasses\":1000,\"multi_label\":true}}}";
+  std::string joutstr = japi.jrender(japi.service_create(sname, jstr));
+  ASSERT_EQ(created_str, joutstr);
+
+  // predict
+  std::string jpredictstr
+      = "{\"service\":\"imgserv\",\"parameters\":{\"input\":{\"height\":224,"
+        "\"width\":224},\"output\":{\"best\":10}},\"data\":"
+        "[\""
+        + incept_repo + "cat.jpg\"]}";
+
+  joutstr = japi.jrender(japi.service_predict(jpredictstr));
+  JDoc jd = JDoc();
+  std::cout << "joutstr=" << joutstr << std::endl;
+  jd.Parse<rapidjson::kParseNanAndInfFlag>(joutstr.c_str());
+  ASSERT_TRUE(!jd.HasParseError());
+  ASSERT_EQ(200, jd["status"]["code"]);
+  ASSERT_TRUE(jd["body"]["predictions"].IsArray());
+  ASSERT_EQ(jd["body"]["predictions"][0]["classes"].Size(), 10);
+
+  // multi_label = no softmax = sums of probs is not 1
+  double sum = 0.0;
+  for (size_t i = 0; i < jd["body"]["predictions"][0]["classes"].Size(); i++)
+    {
+      sum += jd["body"]["predictions"][0]["classes"][i]["prob"].GetDouble();
+    }
+  ASSERT_GT(sum, 1.0);
+}
+
 #if !defined(CPU_ONLY)
 TEST(torchapi, service_predict_fp16)
 {