Support reusing existing model and config for finetuning, fix: #1942

Author: hankcs

hanlp/common/torch_component.py

@@ -246,12 +246,13 @@ def fit(self,
             first_device = -1
         if _device_placeholder and first_device >= 0:
             _dummy_placeholder = self._create_dummy_placeholder_on(first_device)
-        if finetune:
-            if isinstance(finetune, str):
-                self.load(finetune, devices=devices, **self.config)
-            else:
-                self.load(save_dir, devices=devices, **self.config)
-            self.config.finetune = finetune
+        if finetune or self.model:
+            if not self.model:
+                if isinstance(finetune, str):
+                    self.load(finetune, devices=devices, **self.config)
+                else:
+                    self.load(save_dir, devices=devices, **self.config)
+            self.config.finetune = finetune or True
             self.vocabs.unlock()  # For extending vocabs
             logger.info(
                 f'Finetune model loaded with {sum(p.numel() for p in self.model.parameters() if p.requires_grad)}'

plugins/hanlp_demo/hanlp_demo/zh/train/finetune_ner.py

@@ -9,6 +9,7 @@
 
 cdroot()
 
+# 0. Prepare your dataset for finetuning
 your_training_corpus = 'data/ner/finetune/word_to_iobes.tsv'
 your_development_corpus = your_training_corpus  # Use a different one in reality
 save_dir = 'data/ner/finetune/model'
@@ -25,18 +26,22 @@
 '''
         )
 
+# 1. Load a pretrained model for finetuning
 ner = TransformerNamedEntityRecognizer()
+ner.load(hanlp.pretrained.ner.MSRA_NER_ELECTRA_SMALL_ZH)
+
+# 2. Override hyper-parameters
+ner.config['epochs'] = 50  # Since the corpus is small, overfit it
+
+# 3. Fit on your dataset
 ner.fit(
     trn_data=your_training_corpus,
     dev_data=your_development_corpus,
     save_dir=save_dir,
-    epochs=50,  # Since the corpus is small, overfit it
-    finetune=hanlp.pretrained.ner.MSRA_NER_ELECTRA_SMALL_ZH,
-    # You MUST set the same parameters with the fine-tuning model:
-    average_subwords=True,
-    transformer='hfl/chinese-electra-180g-small-discriminator',
+    **ner.config
 )
 
+# 4. Test it out on your data points
 HanLP = hanlp.pipeline()\
     .append(hanlp.load(hanlp.pretrained.tok.FINE_ELECTRA_SMALL_ZH), output_key='tok')\
     .append(ner, output_key='ner')