Merge pull request #258 from MohamedAliBouhaouala/enhance/intent-classifier

Enhance/intent classifier

.gitignore

@@ -4,3 +4,4 @@ certbot
 docker/**/data
 node_modules/
 /docker/**/db_data
+.idea/

nlu/data_loaders/jisfdl.py

@@ -210,10 +210,6 @@ class JISFDL(tfbp.DataLoader):
 
         return encoded_texts, encoded_intents, encoded_slots, intent_names, slot_names
 
-    def get_prediction_data(self) -> str:
-        helper = JsonHelper()
-        dataset = helper.read_dataset_json_file('predict.json')
-        return dataset["text"]
 
     def encode_text(self, text: str, tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast]):
         return self.encode_texts([text], tokenizer)

nlu/data_loaders/tflcdl.py

@@ -125,14 +125,6 @@ class TFLCDL(tfbp.DataLoader):
         else:
             raise ValueError("Unknown method!")
 
-    def get_prediction_data(self):
-        # The predict file contains a single JSON object whose only key is text.
-        data = self.json_helper.read_dataset_json_file("predict.json")
-        text = self.strip_numbers(data["text"])
-        encoded_texts = np.array(self.tfidf.transform(
-            [text]).toarray())  # type: ignore
-        return np.array([text]), encoded_texts
-
     def encode_text(self, text: str):
         sanitized_text = self.strip_numbers(text)
         return self.tfidf.transform([sanitized_text]).toarray() # type: ignore

nlu/models/intent_classifier.py

@@ -12,8 +12,8 @@ if platform == "darwin":
 else:
     from keras.optimizers import Adam
 
-from keras.losses import SparseCategoricalCrossentropy
 from keras.metrics import SparseCategoricalAccuracy
+from focal_loss import SparseCategoricalFocalLoss
 import numpy as np
 
 from data_loaders.jisfdl import JISFDL
@@ -42,6 +42,8 @@ class IntentClassifier(tfbp.Model):
         "num_epochs": 2,
         "dropout_prob": 0.1,
         "intent_num_labels": 7,
+        "gamma": 2,
+        "k": 3
     }
     data_loader: JISFDL
 
@@ -128,7 +130,7 @@ class IntentClassifier(tfbp.Model):
         # Hyperparams, Optimizer and Loss function
         opt = Adam(learning_rate=3e-5, epsilon=1e-08)
 
-        losses = SparseCategoricalCrossentropy()
+        losses = SparseCategoricalFocalLoss(gamma=self.hparams.gamma)
 
         metrics = [SparseCategoricalAccuracy("accuracy")]
 
@@ -172,32 +174,48 @@ class IntentClassifier(tfbp.Model):
 
         return scores
 
-    @tfbp.runnable
-    def predict(self):
-        text = self.data_loader.get_prediction_data()
-
-        info = self.get_prediction(text)
-
-        print(self.summary())
-        print("Text : " + text)
-        print(json.dumps(info, indent=2))
-
-        return json.dumps(info, indent=2)
 
     def get_prediction(self, text: str):
         inputs = self.data_loader.encode_text(text, self.tokenizer)
         intent_probas = self(inputs)  # type: ignore
 
         intent_probas_np = intent_probas.numpy()
-        
+
         # Get the indices of the maximum values
         intent_id = intent_probas_np.argmax(axis=-1)[0]
-        
+
         # get the confidences for each intent
         intent_confidences = intent_probas_np[0]
 
-        return {
+        margin = self.compute_normalized_confidence_margin(intent_probas_np)
+        output = {
             "text": text,
             "intent": {"name": self.extra_params["intent_names"][intent_id],
                        "confidence": float(intent_confidences[intent_id])},
+            "margin": margin,
         }
+
+        return output
+
+    def compute_top_k_confidence(self, probs, k=3):
+        sorted_probas = np.sort(probs[0])[::-1]  # Sort in descending order
+        top_k_sum = np.sum(sorted_probas[:k])
+        return top_k_sum
+
+    def compute_normalized_confidence_margin(self, probs):
+        highest_proba = np.max(probs[0])
+        sum_of_probas = self.compute_top_k_confidence(probs, self.hparams.k)
+        # Normalized margin
+        normalized_margin = highest_proba / sum_of_probas
+        return normalized_margin
+
+    @tfbp.runnable
+    def predict(self):
+        while True:
+
+            text = input("Provide text: ")
+            output = self.get_prediction(text)
+            print(output)
+            # Optionally, provide a way to exit the loop
+            if input("Try again? (y/n): ").lower() != 'y':
+                break

nlu/models/slot_filler.py

@@ -151,16 +151,19 @@ class SlotFiller(tfbp.Model):
 
     @tfbp.runnable
     def predict(self):
-        text = self.data_loader.get_prediction_data()
+        while True:
+            text = input("Provide text: ")
+            info = self.get_prediction(text)
 
-        info = self.get_prediction(text)
+            print(self.summary())
+            print("Text : " + text)
+            print(info)
+            
+            # Optionally, provide a way to exit the loop
+            if input("Try again? (y/n): ").lower() != 'y':
+                break
 
-        print(self.summary())
-        print("Text : " + text)
-        print(json.dumps(info, indent=2))
 
-        return json.dumps(info, indent=2)
-    
     def get_slots_prediction(self, text: str, inputs, slot_probas):
         slot_probas_np = slot_probas.numpy()
         # Get the indices of the maximum values

nlu/models/tflc.py

@@ -95,19 +95,27 @@ class TFLC(tfbp.Model):
 
         self.calculate_metrics(y_test, y_pred, languages)
 
+    def preprocess_text(self, text):
+        # The predict file contains a single JSON object whose only key is text.
+        stripped_text = self.strip_numbers(text)
+        encoded_text = np.array(self.tfidf.transform(
+            [stripped_text]).toarray())  # type: ignore
+        return np.array([stripped_text]), encoded_text
+
     @tfbp.runnable
     def predict(self):
         languages = list(self.extra_params['languages'])
-        texts, encoded_texts = self.data_loader.get_prediction_data()
+        input_provided = input("Provide text: ")
+        text, encoded_text = self.preprocess_text(input_provided)
         # converting a one hot output to language index
-        probas = super().predict(encoded_texts)
+        probas = super().predict(encoded_text)
         predictions = np.argmax(probas, axis=1)
 
         results = []
         for idx, prediction in enumerate(predictions):
             print('The sentence "{}" is in {}.'.format(
-                texts[idx], languages[prediction].upper()))
+                text[idx], languages[prediction].upper()))
-            results.append({'text': texts[idx], 'language': prediction})
+            results.append({'text': text[idx], 'language': prediction})
         return results
 
     def get_prediction(self, text: str):

nlu/requirements.txt

@@ -6,4 +6,5 @@ scikit_learn==1.2.2
 fastapi==0.100.0
 uvicorn[standard]==0.23.1
 autopep8==2.0.2
+focal-loss==0.0.7
 h5py --only-binary=h5py

nlu/utils/json_helper.py

@@ -6,7 +6,7 @@ class JsonHelper:
 
     def __init__(self, model:str = "intent_classifier"):
         self.data_folder=os.path.join("data",model)
-        
+
     def read_dataset_json_file(self, filename):
         file_path = os.path.join(self.data_folder, filename)
         if os.path.exists(file_path):