readme

README.md

@@ -0,0 +1,14 @@
+# Transformer Encoder - RoBERTa
+## Modyfikacje
+1. ????
+
+# Transformer Decoder - GPT-2
+## Modyfikacje
+1. ????
+
+
+# Transformer Encoder-Decoder - T5
+## Modyfikacje
+1. Zamrożenie pierwszych 20 warstw
+
+# Transformer w trybie few-shot/zero-shot learning - ?????

run_translation_freezing.py

@@ -260,8 +260,12 @@ class DataTrainingArguments:
 def freeze_model_weights(model: torch.nn.Module) -> None:
     count = 0
     for param in model.parameters():
-        logger.info(count)
-        param.requires_grad = False
+        count += 1
+        if count < 20:
+            logger.info(f'Freezing layer {count}')
+            param.requires_grad = False
+        else:
+            logger.info(f'Ignoring layer {count}')
 
 
 def main():