אימון מודל נסיוני

machine-learn/scrape_data/level0/Gemini-synthetic-v2/5-creating_model.py

@@ -0,0 +1,158 @@
+import spacy
+from spacy.lang.he import Hebrew
+from spacy.tokenizer import Tokenizer
+from spacy.training.example import Example
+from spacy.lang.char_classes import LIST_PUNCT, LIST_ELLIPSES, LIST_QUOTES, LIST_CURRENCY, LIST_ICONS
+from spacy.util import minibatch, compounding
+import json
+import random
+import logging
+from tqdm import tqdm
+
+# הגדרת הלוגר
+logging.basicConfig(
+    level=logging.INFO,
+    format='%(asctime)s - %(levelname)s - %(message)s',
+    handlers=[
+        logging.StreamHandler()
+    ]
+)
+logger = logging.getLogger(__name__)
+
+def custom_tokenizer(nlp):
+    default_tokenizer = Tokenizer(nlp.vocab)
+    nlp2 = Hebrew()
+
+    LIST_BREAKING_WORDS = [r'—', r'--', r'-', r'\+']
+    LIST_AMPERSAND = [r"[\x2D&]"]
+    LIST_MORE = [r"״", "\."]
+
+    custom_patterns = (
+        LIST_QUOTES +
+        LIST_ELLIPSES +
+        LIST_BREAKING_WORDS +
+        LIST_AMPERSAND +
+        LIST_CURRENCY +
+        LIST_PUNCT +
+        LIST_ICONS +
+        LIST_MORE
+    )
+
+    prefixes = nlp2.Defaults.prefixes + custom_patterns + [r'^(?!וו)ו']
+    infixes = nlp2.Defaults.infixes + custom_patterns
+    suffixes = nlp2.Defaults.suffixes + custom_patterns
+
+    prefix_regex = spacy.util.compile_prefix_regex(prefixes)
+    infix_regex = spacy.util.compile_infix_regex(infixes)
+    suffix_regex = spacy.util.compile_suffix_regex(suffixes)
+
+    nlp2.tokenizer.prefix_search = prefix_regex.search
+    nlp2.tokenizer.infix_finditer = infix_regex.finditer
+    nlp2.tokenizer.suffix_search = suffix_regex.search
+
+    return nlp2.tokenizer
+
+def main():
+    try:
+        nlp = spacy.blank("he")
+        nlp.tokenizer = custom_tokenizer(nlp)
+
+        test_text = "תומר כהן- הישראלי הבכיר בלינקדין"
+        doc = nlp(test_text)
+        logger.info(f"Tokenized Text: {[token.text for token in doc]}")
+
+
+        if "ner" not in nlp.pipe_names:
+            ner = nlp.add_pipe("ner", last=True)
+        else:
+            ner = nlp.get_pipe("ner")
+
+        if "SINGER" not in ner.labels:
+            ner.add_label("SINGER")
+
+        ner.add_label("SONG")
+        ner.add_label("MISC")
+        ner.add_label("ALBUM")
+        ner.add_label("GENRE")
+
+
+        json_files = [
+            'cleaned_new-data.json'
+        ]
+
+        training_data = []
+        for json_file in json_files:
+            with open(json_file, 'r', encoding='utf-8') as f:
+                data = json.load(f)
+                for example_text, example_entities in data:
+                    entities = example_entities.get('entities', [])
+                    example = Example.from_dict(nlp.make_doc(example_text), {'entities': entities})
+                    training_data.append(example)
+
+        logger.info(f"Loaded {len(training_data)} training examples.")
+
+        nlp.begin_training()
+
+        patience = 5
+        min_delta = 1
+        best_loss = float('inf')
+        patience_counter = 0
+        best_model_path = "best_model"
+        n_iter = 100
+        batch_size = 64
+        drop_size = 0.45
+        iteration_data = {}
+
+        logger.info("Starting training...")
+
+        for itn in tqdm(range(n_iter), desc="Training Iterations"):
+            random.shuffle(training_data)
+            losses = {}
+            for i in range(0, len(training_data), batch_size):
+                batch = training_data[i:i + batch_size]
+                nlp.update(batch, drop=drop_size, losses=losses)
+            logger.info(f"Iteration {itn}: {losses}")
+            iteration_data[itn] = losses.copy()
+
+            current_loss = losses.get('ner', float('inf'))
+            if current_loss < best_loss - min_delta:
+                best_loss = current_loss
+                patience_counter = 0
+                # שמירת המודל הטוב ביותר
+                nlp.to_disk(best_model_path)
+                logger.info(f"New best model found at iteration {itn} with loss {current_loss}. Saved to {best_model_path}.")
+            else:
+                patience_counter += 1
+                logger.info(f"No improvement in iteration {itn}. Patience counter: {patience_counter}/{patience}")
+
+            if patience_counter >= patience:
+                logger.info(f"Early stopping at iteration {itn} due to no improvement.")
+                break
+
+        with open(r"C:\Users\me\Documents\GitHub\Singles-Sorter\machine-learn\model_name.txt", 'r', encoding='utf-8') as f:
+            model_name = f.read().strip()
+            logger.info(f'Final Model Name: {model_name}')
+
+        try:
+            with open('iteration_data.json', 'w', encoding='utf-8') as f:
+                json.dump(iteration_data, f, ensure_ascii=False, indent=2)
+            logger.info("Saved iteration data to JSON file.")
+        except Exception as e:
+            logger.error(f'Error saving iteration data to JSON file: {e}')
+
+        # טעינת המודל הטוב ביותר לפני השמירה עם השם הסופי
+        nlp = spacy.load(best_model_path)
+        nlp.meta['name'] = 'singer_ner_he'
+        nlp.meta['description'] = 'Model for recognizing singer names in Hebrew song titles'
+        nlp.meta['author'] = 'nhlocal'
+        nlp.meta['email'] = '[email protected]'
+        nlp.meta['license'] = 'MIT'
+        nlp.meta['tags'] = ['NER', 'Hebrew', 'Singer', 'Named Entity Recognition', 'Text Classification']
+        nlp.to_disk(model_name)
+        logger.info(f"Saved final model to {model_name}.")
+
+    except Exception as e:
+        logger.error(f"An error occurred during training: {e}")
+
+if __name__ == "__main__":
+    main()