[prompt] add maskedlmverbalizer & fix collate_fn (#3889)

Author: LemonNoel

docs/advanced_guide/prompt.md

@@ -355,7 +355,7 @@ template = AutoTemplate.create_from(prompt="{'prefix': None, 'encoder': 'mlp', '
 
 ### 离散型标签词映射
 
-``ManualVerbalizer`` 支持构造 ``{'mask'}`` 对应的标签词映射，支持多``{'mask'}``，直接作用于 ``AutoMaskedLM`` 模型结构。当标签对应的预测词长度大于 ``1`` 时，默认取均值。
+``ManualVerbalizer`` 支持构造 ``{'mask'}`` 对应的标签词映射，同一标签可对应多个不同长度的词，直接作用于 ``AutoMaskedLM`` 模型结构。当标签对应的预测词长度大于 ``1`` 时，默认取均值；当标签对应多个 `{'mask'}` 时，默认与单个 `{mask}` 效果等价。
 
 **调用 API**
 
@@ -373,6 +373,22 @@ verbalizer = ManualVerbalizer(tokenizer=tokenizer,
 - ``label_words`` : 原标签到预测词之间的映射字典。
 - ``tokenizer`` : 预训练模型的 tokenizer，用于预测词的编码。
 
+``MaskedLMVerbalizer`` 同样支持构造 ``{'mask'}`` 对应的标签词映射，映射词与模板中的 `{'mask'}` 逐字对应，因此，映射词长度应与 `{'mask'}` 数量保持一致。当定义的标签词映射中同一标签对应多个词时，仅有第一个映射词生效。在自定义的 `compute_metric` 函数中需先调用 `verbalizer.aggregate_multiple_mask` 将多 `{'mask'}` 合并后再计算评估函数，默认使用乘积的方式。
+
+**调用 API**
+from paddlenlp.prompt import MaskedLMVerbalizer
+from paddlenlp.transformers import AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("ernie-3.0-base-zh")
+verbalizer = MaskedLMVerbalizer(tokenizer=tokenizer,
+                                label_words={'负向': '不', '正向': '很'})
+```
+
+其中初始化参数定义如下
+
+- ``label_words`` : 原标签到预测词之间的映射字典。
+- ``tokenizer`` : 预训练模型的 tokenizer，用于预测词的编码。
+
 ### 连续型标签词映射
 
 标签词映射分类器 ``SoftVerbalizer`` 修改了原 ``AutoMaskedLM`` 的模型结构，将预训练模型最后一层“隐藏层-词表”替换为“隐藏层-标签”的映射。该层网络的初始化参数由标签词映射中的预测词词向量来决定，如果预测词长度大于 ``1`` ，则使用词向量均值进行初始化。当前支持的预训练模型包括 ``ErnieForMaskedLM`` 、 ``BertForMaskedLM`` 、 ``AlbertForMaskedLM`` 和 ``RobertaForMaskedLM`` 。可用于实现 WARP 算法。

paddlenlp/prompt/prompt_utils.py

@@ -96,8 +96,10 @@ def __call__(self, features: List[Dict[str, Any]]) -> Dict[str, Any]:
                         length = len(value)
                         new_values[index][0, :length, :length] = value
                     values = new_values
-                elif key != "labels":
+                elif key in ("soft_token_ids", "encoder_ids"):
                     for index, value in enumerate(values):
                         values[index] = value + [0] * (max_length - len(value))
+                elif key != "labels":
+                    continue
                 batch[key] = self._convert_to_tensors(values)
         return batch

paddlenlp/prompt/verbalizer.py

@@ -27,7 +27,9 @@
 from paddlenlp.transformers import PretrainedTokenizer, PretrainedModel
 from paddlenlp.utils.log import logger
 
-__all__ = ["Verbalizer", "ManualVerbalizer", "SoftVerbalizer"]
+__all__ = [
+    "Verbalizer", "ManualVerbalizer", "SoftVerbalizer", "MaskedLMVerbalizer"
+]
 
 # Verbalizer used to be saved in a file.
 VERBALIZER_CONFIG_FILE = "verbalizer_config.json"
@@ -263,9 +265,11 @@ class ManualVerbalizer(Verbalizer):
             An instance of PretrainedTokenizer for label word tokenization.
     """
 
-    def __init__(self, label_words: Dict, tokenizer: PretrainedTokenizer):
+    def __init__(self, label_words: Dict, tokenizer: PretrainedTokenizer,
+                 **kwargs):
         super(ManualVerbalizer, self).__init__(label_words=label_words,
-                                               tokenizer=tokenizer)
+                                               tokenizer=tokenizer,
+                                               **kwargs)
 
     def create_parameters(self):
         return None
@@ -292,10 +296,7 @@ def aggregate_multiple_mask(self, outputs: Tensor, atype: str = None):
                 "tokens.".format(atype))
         return outputs
 
-    def process_outputs(self,
-                        outputs: Tensor,
-                        masked_positions: Tensor = None,
-                        **kwargs):
+    def process_outputs(self, outputs: Tensor, masked_positions: Tensor = None):
         """
         Process outputs over the vocabulary, including the following steps:
 
@@ -364,10 +365,11 @@ class SoftVerbalizer(Verbalizer):
     LAST_LINEAR = ["AlbertForMaskedLM", "RobertaForMaskedLM"]
 
     def __init__(self, label_words: Dict, tokenizer: PretrainedTokenizer,
-                 model: PretrainedModel):
+                 model: PretrainedModel, **kwargs):
         super(SoftVerbalizer, self).__init__(label_words=label_words,
                                              tokenizer=tokenizer,
-                                             model=model)
+                                             model=modeli,
+                                             **kwargs)
         del self.model
         setattr(model, self.head_name[0], MaskedLMIdentity())
 
@@ -472,3 +474,63 @@ def _create_init_weight(self, weight, is_bias=False):
                                          axis=1).reshape(word_shape)
             weight = self.aggregate(weight, token_mask, aggr_type)
             return weight
+
+
+class MaskedLMVerbalizer(Verbalizer):
+    """
+    MaskedLMVerbalizer defines mapping from labels to words manually and supports
+    multiple masks corresponding to multiple tokens in words.
+
+    Args:
+        label_words (`dict`):
+            Define the mapping from labels to a single word. Only the first word
+            is used if multiple words are defined.
+        tokenizer (`PretrainedTokenizer`):
+            An instance of PretrainedTokenizer for label word tokenization.
+    """
+
+    def __init__(self, label_words: Dict, tokenizer: PretrainedTokenizer,
+                 **kwargs):
+        super(MaskedLMVerbalizer, self).__init__(label_words=label_words,
+                                                 tokenizer=tokenizer,
+                                                 **kwargs)
+
+    def create_parameters(self):
+        return None
+
+    def aggregate_multiple_mask(self, outputs: Tensor, atype: str = "product"):
+        assert outputs.ndim == 3
+        token_ids = self.token_ids[:, 0, :].T
+        batch_size, num_token, num_pred = outputs.shape
+        results = paddle.index_select(outputs[:, 0, :], token_ids[0], axis=1)
+        if atype == "first":
+            return results
+
+        for index in range(1, num_token):
+            sub_results = paddle.index_select(outputs[:, index, :],
+                                              token_ids[index],
+                                              axis=1)
+            if atype in ("mean", "sum"):
+                results += sub_results
+            elif atype == "product":
+                results *= sub_results
+            elif atype == "max":
+                results = paddle.stack([results, sub_results], axis=-1)
+                results = results.max(axis=-1)
+            else:
+                raise ValueError(
+                    "Strategy {} is not supported to aggregate multiple "
+                    "tokens.".format(atype))
+        if atype == "mean":
+            results = results / num_token
+        return results
+
+    def process_outputs(self, outputs: Tensor, masked_positions: Tensor = None):
+        if masked_positions is None:
+            return outputs
+
+        batch_size, _, num_pred = outputs.shape
+        outputs = outputs.reshape([-1, num_pred])
+        outputs = paddle.gather(outputs, masked_positions)
+        outputs = outputs.reshape([batch_size, -1, num_pred])
+        return outputs