Merge pull request #59 from clulab/nonlinearity

Reorganized TokenClassificationHead to use the Sequential module. Added optional nonlinearity.

Author: MihaiSurdeanu

apps/src/main/scala/org/clulab/scala_transformers/apps/TokenClassifierExampleApp.scala

@@ -4,8 +4,8 @@ import org.clulab.scala_transformers.encoder.TokenClassifier
 
 object TokenClassifierExampleApp extends App {
   // Choose one of these.
-  // val tokenClassifier = TokenClassifier.fromFiles("../microsoft-deberta-v3-base-mtl/avg_export")
-  val tokenClassifier = TokenClassifier.fromResources("/org/clulab/scala_transformers/models/microsoft_deberta_v3_base_mtl/avg_export")
+  val tokenClassifier = TokenClassifier.fromFiles("../microsoft-deberta-v3-base-mtl/avg_export")
+  //val tokenClassifier = TokenClassifier.fromResources("/org/clulab/scala_transformers/models/microsoft_deberta_v3_base_mtl/avg_export")
   // val tokenClassifier = TokenClassifier.fromFiles("../models/google_electra_small_discriminator_mtl/avg_export")
   // val tokenClassifier = TokenClassifier.fromResources("/org/clulab/scala_transformers/models/google_electra_small_discriminator_mtl/avg_export")
   // val tokenClassifier = TokenClassifier.fromFiles("../models/roberta_base_mtl/avg_export")
@@ -14,7 +14,8 @@ object TokenClassifierExampleApp extends App {
   //val words = Seq("EU", "rejects", "German", "call", "to", "boycott", "British", "lamb", ".")
   //val words = Seq("John", "Doe", "went", "to", "China", ".")
   //val words = Seq("John", "Doe", "went", "to", "China", ".")
-  val words = Seq("Ras1", "has", "phosphorylated", "Mek2", ".")
+  //val words = Seq("Ras1", "has", "phosphorylated", "Mek2", ".")
+  val words = Seq("Jack", "Doe", "and", "John", "Doe", "were", "friends", "in", "yyy", "zzz", ".")
 
   println(s"Words: ${words.mkString(", ")}")
 

encoder/src/main/python/averaging_trainer.py

@@ -83,28 +83,29 @@ def average_checkpoints(self,
         main_model = self.load_model(checkpoints[0], config, tasks)
         print("Done loading.")
 
-        self.print_some_params(main_model, "before averaging:")
+        # TODO: update print
+        # self.print_some_params(main_model, "before averaging:")
 
         for i in range(1, len(checkpoints)):  # Skip 0, which is the main_model.
             print(f"Loading satellite checkpoint[{i}] {checkpoints[i]}...")
             satellite_model = self.load_model(checkpoints[i], config, tasks)
-            self.print_some_params(satellite_model, "satellite model:")
+            #self.print_some_params(satellite_model, "satellite model:") # TODO
 
             print("Adding its parameter weights to the main model...")
             for key, value in main_model.state_dict().items():
                 if value.data.type() != "torch.LongTensor":
                     value.data += satellite_model.state_dict()[key].data.clone()
             print("Done adding")
 
-        self.print_some_params(main_model, "after summing:")
+        # self.print_some_params(main_model, "after summing:") # TODO
         if len(checkpoints) > 1:
             print("Computing average weights...")
             for value in main_model.state_dict().values():
                 if value.data.type() != "torch.LongTensor":
                     value.data /= len(checkpoints)
             print("Done computing.")
 
-        self.print_some_params(main_model, "after averaging:")
+        # self.print_some_params(main_model, "after averaging:") # TODO
         print("Saving averaged model...")
         main_model.save_pretrained(path_to_save)
         main_model.export_model(tasks, tokenizer, path_to_export)

encoder/src/main/python/token_classifier.py

@@ -14,6 +14,9 @@
 from transformers.modeling_outputs import TokenClassifierOutput
 from typing import Any, Callable, List, Optional, Union
 
+# This global variable indicates the position of the linear layer in TokenClassificationHead.classifier
+linear_pos = 1
+
 # This class is adapted from: https://towardsdatascience.com/how-to-create-and-train-a-multi-task-transformer-model-18c54a146240
 class TokenClassificationModel(PreTrainedModel):    
     def __init__(self, config: AutoConfig, transformer_name: str) -> None:
@@ -112,9 +115,10 @@ def save_pretrained(self,
         # https://pytorch.org/tutorials/beginner/saving_loading_models.html
         torch.save(self.state_dict(), f"{save_directory}/pytorch_model.bin")
         print("pickle saving done.")
-        #super().save_pretrained(save_directory, is_main_process, state_dict, save_function, push_to_hub, max_shard_size, safe_serialization, **kwargs)
+
+        # some optional debugging prints:
         #for i in range(5):
-        #  key = f"output_heads.{i}.classifier.weight"
+        #  key = f"output_heads.{i}.classifier[linear_pos].weight"
         #  print(f"{key} = {self.state_dict()[key]}")
 
     def from_pretrained(self, pretrained_model_name_or_path: str, *model_args, **kwargs) -> None:
@@ -129,8 +133,8 @@ def from_pretrained(self, pretrained_model_name_or_path: str, *model_args, **kwa
             print("Done loading.")
 
     def export_task(self, task_head, task: Task, task_checkpoint) -> None:
-        numpy_weights = task_head.classifier.weight.cpu().detach().numpy()
-        numpy_bias = task_head.classifier.bias.cpu().detach().numpy()
+        numpy_weights = task_head.classifier[linear_pos].weight.cpu().detach().numpy()
+        numpy_bias = task_head.classifier[linear_pos].bias.cpu().detach().numpy()
         labels = task.labels
         #print(f"Shape of weights: {numpy_weights.shape}")
         #print(f"Weights are:\n{numpy_weights}")
@@ -220,25 +224,31 @@ def export_model(self, tasks: List[Task], tokenizer: AutoTokenizer, checkpoint_d
 class TokenClassificationHead(nn.Module):
     def __init__(self, hidden_size: int, num_labels: int, task_id, dual_mode: bool=False, dropout_p: float=0.1):
         super().__init__()
-        self.dropout = nn.Dropout(dropout_p)
+        #self.dropout = nn.Dropout(dropout_p)
         self.dual_mode = dual_mode
-        self.classifier = nn.Linear(
-            hidden_size * 2 if (self.dual_mode and Parameters.use_concat) else hidden_size,
-            num_labels
-        )
+        # make sure to adjust linear_pos if the position of Linear here changes!
+        self.classifier = nn.Sequential(
+            # nn.ReLU(), # it works better wo/ a nonlinearity here
+            nn.Dropout(dropout_p),
+            nn.Linear(
+              hidden_size * 2 if (self.dual_mode and Parameters.use_concat) else hidden_size,
+              num_labels
+            )
+          )
         self.num_labels = num_labels
         self.task_id = task_id
         self._init_weights()
 
     def _init_weights(self):
-        self.classifier.weight.data.normal_(mean=0.0, std=0.02)
-        # torch.nn.init.xavier_normal_(self.classifier.weight.data)
-        if self.classifier.bias is not None:
-            self.classifier.bias.data.zero_()    
+      print("Nothing to initialize here.")
+      self.classifier[linear_pos].weight.data.normal_(mean=0.0, std=0.02)
+      #torch.nn.init.xavier_normal_(self.classifier[linear_pos].weight.data)
+      if self.classifier[linear_pos].bias is not None:
+        self.classifier[linear_pos].bias.data.zero_()    
 
     def summarize(self, task_id):
         print(f"Task {self.task_id} with {self.num_labels} labels.")
-        print(f"Dropout is {self.dropout}")
+        #print(f"Dropout is {self.dropout}")
         print(f"Classifier layer is {self.classifier}")
 
     def concatenate(self, sequence_output, head_positions):
@@ -250,16 +260,12 @@ def concatenate(self, sequence_output, head_positions):
         # Concatenate the hidden states from modifier + head.
         modifier_head_states = torch.cat([sequence_output, head_states], dim=2) if Parameters.use_concat else torch.add(sequence_output, head_states) 
         #print(f"modifier_head_states.size = {modifier_head_states.size()}")
-        #print("EXIT")
-        #exit(1)
         return modifier_head_states
 
     def forward(self, sequence_output, pooled_output, head_positions, labels=None, attention_mask=None, **kwargs):
         #print(f"sequence_output size = {sequence_output.size()}")
         sequence_output_for_classification = sequence_output if not self.dual_mode else self.concatenate(sequence_output, head_positions)
-
-        sequence_output_dropout = self.dropout(sequence_output_for_classification)
-        logits = self.classifier(sequence_output_dropout)
+        logits = self.classifier(sequence_output_for_classification)
 
         loss = None
         if labels is not None:

encoder/src/main/scala/org/clulab/scala_transformers/encoder/Encoder.scala

@@ -1,12 +1,13 @@
 package org.clulab.scala_transformers.encoder
 
 import ai.onnxruntime.{OnnxTensor, OrtEnvironment, OrtSession}
+import org.clulab.scala_transformers.encoder.math.Mathematics
 import org.clulab.scala_transformers.encoder.math.Mathematics.{Math, MathMatrix}
 
 import java.io.DataInputStream
 import java.util.{HashMap => JHashMap}
 
-class Encoder(val encoderEnvironment: OrtEnvironment, val encoderSession: OrtSession) {
+class Encoder(val encoderEnvironment: OrtEnvironment, val encoderSession: OrtSession, nonLin: Option[NonLinearity] = None) {
   /**
     * Runs the inference using a transformer encoder over a batch of sentences
     *
@@ -19,6 +20,19 @@ class Encoder(val encoderEnvironment: OrtEnvironment, val encoderSession: OrtSes
 
     val result: OrtSession.Result = encoderSession.run(inputs)
     val outputs = Math.fromResult(result)
+
+    if(nonLin.isDefined) {
+      for (matrix <- outputs) {
+        for (i <- 0 until Math.rows(matrix)) {
+          val row = Math.row(matrix, i)
+          for (j <- 0 until Math.cols(matrix)) {
+            val orig = Math.get(row, j)
+            Math.set(row, j, nonLin.get.compute(orig))
+          }
+        }
+      }
+    }
+
     outputs
   }
 
@@ -36,8 +50,10 @@ class Encoder(val encoderEnvironment: OrtEnvironment, val encoderSession: OrtSes
 object Encoder {
   val ortEnvironment = OrtEnvironment.getEnvironment
 
+  // val nonLinearity = Some(new ReLU())
+
   protected def fromSession(ortSession: OrtSession): Encoder =
-      new Encoder(ortEnvironment, ortSession)
+      new Encoder(ortEnvironment, ortSession, None) // nonLinearity // ms: skipping the nonlinearity is better
 
   protected def ortSessionFromFile(fileName: String): OrtSession =
       ortEnvironment.createSession(fileName, new OrtSession.SessionOptions)

encoder/src/main/scala/org/clulab/scala_transformers/encoder/NonLinearity.scala

@@ -0,0 +1,15 @@
+package org.clulab.scala_transformers.encoder
+
+import org.clulab.scala_transformers.encoder.math.EjmlMath.MathValue
+
+import java.lang
+
+trait NonLinearity {
+  def compute(input: MathValue): MathValue
+}
+
+class ReLU extends NonLinearity {
+  override def compute(input: MathValue): MathValue = {
+    lang.Float.max(0, input)
+  }
+}

encoder/src/main/scala/org/clulab/scala_transformers/encoder/math/EjmlMath.scala

@@ -132,6 +132,11 @@ object EjmlMath extends Math {
     rowVector.get(index)
   }
 
+  def set(rowVector: MathRowVector, index: Int, value: MathValue): Unit = {
+    assert(isRowVector(rowVector))
+    rowVector.set(index, value)
+  }
+
   def mkMatrixFromRows(values: Array[Array[MathValue]]): MathRowMatrix = {
     new FMatrixRMaj(values)
   }

encoder/src/main/scala/org/clulab/scala_transformers/encoder/math/Math.scala

@@ -23,6 +23,7 @@ trait Math {
   def horcat(leftRowVector: MathRowVector, rightRowVector: MathRowVector): MathRowVector
   def toArray(rowVector: MathRowVector): Array[MathValue]
   def get(rowVector: MathRowVector, index: Int): MathValue
+  def set(rowVector: MathRowVector, index: Int, value: MathValue): Unit
   def mkMatrixFromRows(values: Array[Array[MathValue]]): MathRowMatrix
   // For this, the array is specified in column-major order,
   // but it should be converted to the normal representation.