fix chatglm2 evaluation error: hypothesis emtpy (#348)

* fix evaluation error: hypothesis emtpy

* fix pipeline

* fix bug

examples/pytorch/chatglm6b/chatglm_trainer.py

@@ -16,6 +16,8 @@ class Seq2SeqTrainer(EpochBasedTrainer):
         if ignore_pad_token_for_loss:
             tokens = np.where(tokens != -100, tokens,
                               self.tokenizer.pad_token_id)
+        tokens = np.where(tokens < self.tokenizer.vocab_size, tokens,
+                          self.tokenizer.pad_token_id)
         return [
             t for t in self.tokenizer.batch_decode(
                 tokens, skip_special_tokens=True) if t != '</s>'
@@ -59,7 +61,9 @@ class Seq2SeqTrainer(EpochBasedTrainer):
 
         gen_kwargs['input_ids'] = generation_inputs
         gen_kwargs['pad_token_id'] = self.tokenizer.pad_token_id
-        generated_tokens = self.model.generate(**gen_kwargs)
+        self.model.eval()
+        with torch.no_grad():
+            generated_tokens = self.model.generate(**gen_kwargs)
         generated_tokens = generated_tokens[:, generation_inputs.size()[-1]:]
 
         # in case the batch is shorter than max length, the output should be padded

examples/pytorch/chatglm6b/finetune.py

@@ -192,8 +192,15 @@ if config['model']['type'] == 'chatglm6b':
     model_config['model']['prefix_projection'] = args.prefix_projection
 
 tokenizer = ChatGLMTokenizer.from_pretrained(model_dir, trust_remote_code=True)
+
+device_map_kwargs = {}
+device_kwargs = {}
+if args.use_lora != 0:
+    device_kwargs['device_map'] = 'auto'
+    # No placement for model, leave the model to `device_map`
+    device_kwargs['device'] = 'cpu'
 model = Model.from_pretrained(
-    model_dir, cfg_dict=model_config, device_map='auto')
+    model_dir, cfg_dict=model_config, **device_map_kwargs)
 
 if args.ptuning_checkpoint is not None:
     # Evaluation
@@ -378,8 +385,7 @@ trainer = Seq2SeqTrainer(
     seed=args.seed,
     data_collator=data_collator,
     remove_unused_data=True,
-    # No placement for model, leave the model to `device_map`
-    device='cpu',
-    cfg_modify_fn=cfg_modify_fn)
+    cfg_modify_fn=cfg_modify_fn,
+    **device_kwargs)
 trainer.tokenizer = tokenizer
 trainer.train()

examples/pytorch/chatglm6b/text_generation_metric.py

@@ -53,7 +53,7 @@ class TextGenerationMetric(Metric):
         }
         for pred, label in zip(preds, labels):
             hypothesis = list(jieba.cut(pred))
-            if len(hypothesis) == 0:
+            if len(hypothesis) == 0 or ''.join(hypothesis) == '.':
                 hypothesis = ['</s>']
             reference = list(jieba.cut(label))
             rouge = Rouge()

modelscope/pipelines/nlp/text_generation_pipeline.py

@@ -6,6 +6,7 @@ from typing import Any, Dict, Optional, Union
 
 import torch
 
+from modelscope import snapshot_download
 from modelscope.metainfo import Pipelines
 from modelscope.models.base import Model
 from modelscope.outputs import (ModelOutputBase, OutputKeys,
@@ -192,9 +193,14 @@ class ChatGLM6bTextGenerationPipeline(Pipeline):
                  quantization_bit=None,
                  use_bf16=False,
                  **kwargs):
-        from modelscope.models.nlp.chatglm.text_generation import ChatGLMForConditionalGeneration
-        model = ChatGLMForConditionalGeneration(model) if isinstance(
-            model, str) else model
+        from modelscope.models.nlp.chatglm.text_generation import ChatGLMForConditionalGeneration, ChatGLMConfig
+        if isinstance(model, str):
+            model_dir = snapshot_download(
+                model) if not os.path.exists(model) else model
+            config = ChatGLMConfig.from_pretrained(model_dir)
+            model = ChatGLMForConditionalGeneration(config).half()
+            if torch.cuda.is_available():
+                model = model.cuda()
         if quantization_bit is not None:
             model = model.quantize(quantization_bit)
         if use_bf16:
@@ -225,9 +231,14 @@ class ChatGLM6bV2TextGenerationPipeline(Pipeline):
                  quantization_bit=None,
                  use_bf16=False,
                  **kwargs):
-        from modelscope.models.nlp import ChatGLM2ForConditionalGeneration, ChatGLM2Tokenizer
-        model = ChatGLM2ForConditionalGeneration(model) if isinstance(
-            model, str) else model
+        from modelscope.models.nlp import ChatGLM2ForConditionalGeneration, ChatGLM2Tokenizer, ChatGLM2Config
+        if isinstance(model, str):
+            model_dir = snapshot_download(
+                model) if not os.path.exists(model) else model
+            config = ChatGLM2Config.from_pretrained(model_dir)
+            model = ChatGLM2ForConditionalGeneration(config)
+            if torch.cuda.is_available():
+                model = model.cuda()
         if quantization_bit is not None:
             model = model.quantize(quantization_bit)
         if use_bf16: