notes

docs/acceptable_mistakes.txt

@@ -0,0 +1,5 @@
+
+Greeting function misinterpretation
+Misinterpreting a greeting sentence as an information-querying sentence is an acceptable mistake.
+Lots of people make it when greeted with "How are ya?"
+

docs/concepts_to_graph.txt

@@ -0,0 +1,10 @@
+
+
+A list of things to consider fleshing out in the web
+
+
+Qualia
+
+
+
+

docs/feature_justification.txt

@@ -0,0 +1,22 @@
+
+
+Arguments justifying different features
+
+
+
+Information Priority
+Pieces of information have different priorities and should be handled accordingly.
+Keep a 'hot' web of the most important parts of info that should immediately be searchable every frame.
+This can link into 'colder' webs that store lower priority information.
+Will have this with both relations and nodes in the web.
+To justify for nodes, some people in your life are of higher importance than others to remember.
+You don't keep all people you know immediately accessible in your brain.
+To justify for relations, relations between specific numbers are thrown out almost immediately when done solving a math problem.
+
+
+Layered webs
+If you have a conversation about the color of a hat, you will not store the color of the hat in your immediately accessible brain long term.
+This lets you not constantly scan a huge repertoire of information when you want to perform day-to-day tasks.
+Thus, I want to create an information heirarchy to mimic this and make engineering potentially easier.
+
+

docs/mantas.txt

@@ -0,0 +1,6 @@
+
+1. keep it simple, stupid
+
+2. make everything easily debug-able
+
+3. work on a single sentence at a time

improvement_ideas.txt

@@ -1 +1,11 @@
-summarize previous statements to provide context instead of using full statement
+
+summarize previous statements to provide context instead of using full statement
+
+Come up with strategies for culling the number of connections in the knowledge web
+ - Particularly, connections used during immediate processing
+   - don't need to keep track of the color of the paper that a sentence you're going to forget about is written on
+ - Assign value weight to different nodes in web and use that to periodically clean web (?)
+
+Layered webs
+ - Have more fine grained processing data get pushed into a lower priority web that is not immediately searched all the time
+ - ie low value detail relations, like colors or other qualities of objects, relations between people, etc

src/main/java/org/studiorailgun/conversation/evaluators/GreetingEval.java

@@ -0,0 +1,20 @@
+package org.studiorailgun.conversation.evaluators;
+
+/**
+ * Evaluates a greeting
+ */
+public class GreetingEval {
+    
+    /**
+     * Evaluates a greeting
+     * @param input The sentence
+     */
+    public static void evaluate(String input){
+        switch(input){
+            case "Hello": {
+                
+            } break;
+        }
+    }
+
+}

src/main/java/org/studiorailgun/conversation/semantic/SentenceSubjectParser.java

@@ -36,6 +36,10 @@ public class SentenceSubjectParser {
         // System.out.println(model.functions().get(0).signature().getInputs().values().iterator().next().shape);
     }
 
+    /**
+     * Evaluates a sentence
+     * @param sentence The sentence to evaluate
+     */
     public void evaluate(String sentence){
         //run predict
         TInt64 tensor = TInt64.scalarOf(10);

src/main/python/conversation/sentence/function.py

@@ -0,0 +1,105 @@
+import tensorflow as tf
+keras = tf.keras
+from tensorflow import Tensor
+from keras.api.layers import TextVectorization, Embedding, LSTM, Dense, Input
+from keras.api.models import Sequential
+import numpy as np
+import numpy.typing as npt
+
+
+
+#
+#
+#  Description: The purpose of this model is to qualify sentences into different 'functions'
+#
+#    - "Utility" - A sentence whose purpose is to perform some non-informational duty in a conversation.
+#                  An example of this could be a greeting or farewell.
+#
+#    - "Transfer" - Transfers a piece of information to another participant. For instance, describing
+#                   the details of an object.
+#
+#    - "Query" - Queries a piece of information from another participant.
+#
+#    - "Imperative" - Commands a participant to do something.
+#
+
+
+
+
+# Model constants.
+max_features: int = 20000
+embedding_dim: int = 128
+sequence_length: int = 500
+epochs: int = 50
+max_tokens: int = 5000
+output_sequence_length: int = 4
+
+
+# read sentences
+data_path: str = './data/semantic/subject.txt'
+data_raw: str = open(data_path).read()
+vocab: list[str] = data_raw.split('\n')
+
+# read labels
+label_data_path: str = './data/semantic/subject_label.txt'
+label_data_raw: str = open(label_data_path).read()
+labels: list[int] = list(map(int,label_data_raw.split()))
+
+# init vectorizer
+textVec: TextVectorization = TextVectorization(
+    max_tokens=max_tokens,
+    output_mode='int',
+    output_sequence_length=output_sequence_length,
+    pad_to_max_tokens=True)
+
+# Add the vocab to the tokenizer
+textVec.adapt(vocab) 
+input_data: list[str] = vocab
+data: Tensor = textVec.call(input_data)
+
+# construct model
+model: Sequential = Sequential([
+    keras.Input(shape=(None,), dtype="int64"),
+    Embedding(max_features + 1, embedding_dim),
+    LSTM(64),
+    Dense(1, activation='sigmoid')
+])
+
+#compile the model
+# model.build(keras.Input(shape=(None,), dtype="int64"))
+model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])
+
+
+# fit the training data
+npData = np.array(data)
+npLabel = np.array(labels)
+model.fit(npData,npLabel,epochs=epochs)
+
+
+# evaluate here
+
+
+# predict
+predictTargetRaw: list[str] = ['saf']
+predictTargetToken: list[int] = textVec.call(predictTargetRaw)
+npPredict: npt.NDArray[np.complex64] = np.array(predictTargetToken)
+# print(npPredict)
+result: list[int] = model.predict(npPredict)
+print("predict result:")
+print(predictTargetToken)
+print(result)
+print(data)
+print(labels)
+
+
+
+# save the model so keras can reload
+# savePath: str = './data/semantic/model.keras'
+# model.save(savePath)
+
+# export the model so java can leverage it
+exportPath: str = './data/semantic/model'
+model.export(exportPath)
+
+# tf.keras.utils.get_file('asdf')
+# asdf: str = 'a'