bug: set metamodel nclasses *before* building metamodel

uclh-criu · Jan 28, 2025 · c971a8b · c971a8b
1 parent 56270cf
commit c971a8b
Showing 1 changed file with 10 additions and 8 deletions.
diff --git a/src/scripts/miade.py b/src/scripts/miade.py
@@ -18,7 +18,6 @@
 from shutil import rmtree
 from typing import Optional, List
 from pydantic import BaseModel, validator
-from datetime import datetime
 
 from tokenizers import ByteLevelBPETokenizer
 from gensim.models import Word2Vec
@@ -746,13 +745,13 @@ def make(config_filepath: Path, temp_dir: Path = Path("./.temp"), output: Path =
 
             meta_model_categories.append(meta_spec.config.general.category_name)
 
-            meta_model = MiADE_MetaCAT(tokenizer=tokenizer, embeddings=embeddings, config=meta_spec.config)
-
             annotations_path = meta_spec.annotations.get_or_download(temp_dir)
 
             if meta_spec.synthetic_data:
                 synthetic_csv_path = meta_spec.synthetic_data.get_or_download(temp_dir)
-                synthetic_data = pl.read_csv(synthetic_csv_path)[["text", meta_model_name]]
+                synthetic_data = pl.read_csv(synthetic_csv_path, null_values=["nan"])[
+                    ["text", "start", "end", meta_model_name]
+                ].filter(pl.col(meta_model_name).is_not_null())
                 prepped_synthetic_data_path = temp_dir / Path(f"prepped_synthetic_data_{meta_model_name}")
                 synthetic_data.write_csv(prepped_synthetic_data_path)
                 synthetic_label_counts = _get_label_counts_for_synthetic_data(synthetic_data, meta_model_name)
@@ -765,16 +764,19 @@ def make(config_filepath: Path, temp_dir: Path = Path("./.temp"), output: Path =
                 synthetic_label_counts,
             )
 
+            log.info(f"Number of examples: {label_counts}")
+
             labels = list(label_counts.keys())
             total_label_count = sum(label_counts.values())
 
-            meta_model.config.general["category_value2id"] = dict(zip(labels, range(0, len(labels))))
+            meta_spec.config.general["category_value2id"] = dict(zip(labels, range(0, len(labels))))
             if meta_spec.balance_weights:
-                meta_model.config.train["class_weights"] = [
+                meta_spec.config.train["class_weights"] = [
                     1 - (label_counts[label] / total_label_count) for label in labels
                 ]
-            meta_model.config.model.nclasses = len(labels)
-            meta_model.config.model["last_trained_on"] = datetime.now().strftime("%y%m%d%H%M%S")
+            meta_spec.config.model.nclasses = len(labels)
+
+            meta_model = MiADE_MetaCAT(tokenizer=tokenizer, embeddings=embeddings, config=meta_spec.config)
 
             log.info(
                 f"Starting MetaCAT training for {meta_spec.config.general['category_name']} for {meta_spec.config.train.nepochs} epoch(s) "