new weights with masks over padding

src/_jslml.mjs

@@ -192,7 +192,7 @@ class JslBERTBlock {
                     }, this);
                     if (Mvdef) perc = Activation.masked_softmax(perc, Mv);
                     else perc = Activation.softmax(perc);
-                }else{//if no attention for this query, we add uniform attention
+                }else{//if no attention for this query, we add a uniform attention
                     perc = uniform(K_enc_head.length);
                 }
 
@@ -234,12 +234,13 @@ class JslBERT {
      * @param {Perceptron[]} encoders encoders for inputs embedding (token, position, segment)
      * @param {JslBERTBlock[]} blocks stacked blocks for incoding 
      */
-    constructor(encoders, blocks){
+    constructor(encoders, blocks, masked=false){
         this.encoders = encoders;
         this.blocks = blocks;
+        this.masked = masked;
     }
 
-    predict(X, masked=false){// X = [[[TOKEN], [POSITION], [SEGMENT]], [[], [], []], ...]
+    predict(X){// X = [[[TOKEN], [POSITION], [SEGMENT]], [[], [], []], ...]
 
         //incoding inputs
         let X_enc = [];
@@ -254,7 +255,9 @@ class JslBERT {
         // console.log("Word", X_enc);
 
         //the mask, when the token starts with 1 then 0; otherwise 1
-        const M = masked? X.map(x => 1 - x[0]): undefined;
+        const M = this.masked? X.map(x => 1 - x[0][0]): undefined;
+
+        // console.log("Mask", M);
 
         this.blocks.forEach(block => {
             X_enc = block.predict(X_enc, X_enc, X_enc, M, M);
@@ -274,6 +277,7 @@ class JslBERT {
 // }
 
 class TagEncoder{
+
     constructor(tag_list, embedding={predict: x => x}){//new EmptyNeuron()
         this.tag_list = tag_list;
         this.embedding = embedding;