merge develop branch

applications/text_classification/hierarchical/few-shot/README.md

@@ -65,9 +65,9 @@
 
    内存: 630 GB
 
-3. PaddlePaddle 版本：2.3.1
+3. PaddlePaddle 版本：2.4rc
 
-4. PaddleNLP 版本：2.3.5 (develop)
+4. PaddleNLP 版本：2.4.3
 
 5. 评估设置
 
@@ -91,7 +91,7 @@
    | model_name | 训练方式 | Micro F1分数 | Macro F1分数 |
    | ---------- | ------- | ----------- | ----------- |
    | ernie-3.0-base-zh | 微调学习 | 0.7172 | 0.3821 |
-   | ernie-3.0-base-zh | 提示学习 | 0.8855 | 0.8443 |
+   | ernie-3.0-base-zh | 提示学习 | 0.8945 | 0.8516 |
 
 
 <a name="定制训练"></a>
@@ -102,10 +102,10 @@
 <a name="运行环境"></a>
 ### 3.1 运行环境
 
-- python >= 3.6
-- paddlepaddle > 2.3（2.4版本发布前推荐安装[develop版本](https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/zh/develop/install/pip/linux-pip.html)）
-- paddlenlp >= 2.3.5
-- paddle2onnx >= 1.0.0rc3
+- python >= 3.7
+- paddlepaddle >= 2.4rc
+- paddlenlp >= 2.4.3
+- paddle2onnx >= 1.0.3
 
 <a name="代码结构"></a>
 ### 3.2 代码结构
@@ -222,12 +222,12 @@ python train.py \
 --do_export \
 --num_train_epochs 100 \
 --logging_steps 5 \
+--save_total_limit 1 \
 --per_device_eval_batch_size 32 \
 --per_device_train_batch_size 8 \
 --metric_for_best_model macro_f1_score \
 --load_best_model_at_end \
---evaluation_strategy epoch \
---save_strategy epoch
+--eval_steps 100
 ```
 **多卡训练**
 
@@ -247,12 +247,12 @@ python -u -m paddle.distributed.launch --gpus 0,1,2,3 train.py \
 --do_export \
 --num_train_epochs 100 \
 --logging_steps 5 \
+--save_total_limit 1 \
 --per_device_eval_batch_size 32 \
 --per_device_train_batch_size 8 \
 --metric_for_best_model macro_f1_score \
 --load_best_model_at_end \
---evaluation_strategy epoch \
---save_strategy epoch
+--eval_steps 100
 ```
 
 可配置参数说明：
@@ -273,6 +273,7 @@ python -u -m paddle.distributed.launch --gpus 0,1,2,3 train.py \
 - `do_export`: 是否在运行结束时将模型导出为静态图，保存路径为`output_dir/export`。
 - `num_train_epochs`: 训练的最大轮数。
 - `max_steps`: 训练的最大步数。此设置将会覆盖`num_train_epochs`。
+- `save_total_limit`: 模型检查点保存数量。
 - `device`: 使用的设备，默认为`gpu`。
 - `eval_steps`: 评估模型的间隔步数。
 - `logging_steps`: 打印日志的间隔步数。
@@ -352,9 +353,9 @@ python infer.py --model_path_prefix checkpoints/export/model --data_dir ./data -
 可配置参数说明：
 
 - `model_path_prefix`: 导出的静态图模型路径及文件前缀。
-- `model_name_or_path`: 内置预训练模型名，或者模型参数配置目录路径，用于加载tokenizer。默认为`ernie-3.0-base-zh`。
+- `model_name`: 内置预训练模型名，用于加载tokenizer。默认为`ernie-3.0-base-zh`。
 - `data_dir`: 待推理数据所在路径，数据应存放在该目录下的`data.txt`文件。
-- `max_seq_length`: 最大句子长度，超过该长度的文本将被截断，不足的以Pad补全。提示文本不会被截断。
+- `max_length`: 最大句子长度，超过该长度的文本将被截断，不足的以Pad补全。提示文本不会被截断。
 - `batch_size`: 每次预测的样本数量。
 - `device`: 选择推理设备，包括`cpu`和`gpu`。默认为`gpu`。
 - `device_id`: 指定GPU设备ID。

applications/text_classification/hierarchical/few-shot/infer.py

@@ -14,23 +14,24 @@
 
 import os
 import six
+import json
 import psutil
 import argparse
 
 import numpy as np
 
 from paddlenlp.utils.log import logger
-from paddlenlp.prompt import AutoTemplate, Verbalizer, InputExample
-from paddlenlp.transformers import AutoTokenizer
+from paddlenlp.prompt import AutoTemplate, PromptDataCollatorWithPadding
+from paddlenlp.transformers import AutoTokenizer, AutoModelForMaskedLM
 import paddle2onnx
 import onnxruntime as ort
 
 # yapf: disable
 parser = argparse.ArgumentParser()
 parser.add_argument("--model_path_prefix", type=str, required=True, help="The path prefix of inference model to be used.")
-parser.add_argument("--model_name_or_path", default="ernie-3.0-base-zh", type=str, help="The directory or name of model.")
+parser.add_argument("--model_name", default="ernie-3.0-base-zh", type=str, help="The name of pretrained model.")
 parser.add_argument("--data_dir", default=None, type=str, help="The path to the prediction data, including label.txt and data.txt.")
-parser.add_argument("--max_seq_length", default=128, type=int, help="The maximum total input sequence length after tokenization.")
+parser.add_argument("--max_length", default=128, type=int, help="The maximum total input sequence length after tokenization.")
 parser.add_argument("--use_fp16", action='store_true', help="Whether to use fp16 inference, only takes effect when deploying on gpu.")
 parser.add_argument("--batch_size", default=200, type=int, help="Batch size per GPU/CPU for predicting.")
 parser.add_argument("--num_threads", default=psutil.cpu_count(logical=False), type=int, help="num_threads for cpu.")
@@ -103,12 +104,6 @@ def __init__(self,
                                                   'device_id':
                                                   device_id
                                               }])
-        self.input_handles = [
-            self.predictor.get_inputs()[0].name,
-            self.predictor.get_inputs()[1].name,
-            self.predictor.get_inputs()[2].name
-        ]
-
         if device == "gpu":
             try:
                 assert 'CUDAExecutionProvider' in self.predictor.get_providers()
@@ -122,27 +117,52 @@ def __init__(self,
         logger.info(">>> [InferBackend] Engine Created ...")
 
     def infer(self, input_dict: dict):
-        input_dict = {
-            k: v
-            for k, v in input_dict.items() if k in self.input_handles
-        }
         result = self.predictor.run(None, input_dict)
         return result
 
 
 class HierachicalPredictor(object):
 
-    def __init__(self, args, label_list):
-        self._label_list = label_list
-        self._tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
-        self._max_seq_length = args.max_seq_length
-        self._batch_size = args.batch_size
-        self.inference_backend = InferBackend(args.model_path_prefix,
-                                              args.device, args.device_id,
-                                              args.use_fp16, args.num_threads)
-        self._template = AutoTemplate.load_from(
-            os.path.dirname(args.model_path_prefix), self._tokenizer,
-            args.max_seq_length)
+    def __init__(self, args):
+        self.args = args
+        self.tokenizer = AutoTokenizer.from_pretrained(args.model_name)
+        self.model = AutoModelForMaskedLM.from_pretrained(args.model_name)
+        self.template, self.labels, self.input_handles = self.post_init()
+        self.collate_fn = PromptDataCollatorWithPadding(
+            self.tokenizer,
+            padding=True,
+            return_tensors="np",
+            return_attention_mask=True)
+
+        self.inference_backend = InferBackend(self.args.model_path_prefix,
+                                              self.args.device,
+                                              self.args.device_id,
+                                              self.args.use_fp16,
+                                              self.args.num_threads)
+
+    def post_init(self):
+        export_path = os.path.dirname(self.args.model_path_prefix)
+        template_path = os.path.join(export_path, "template_config.json")
+        with open(template_path, "r") as fp:
+            prompt = json.load(fp)
+            template = AutoTemplate.create_from(prompt, self.tokenizer,
+                                                self.args.max_length,
+                                                self.model)
+        keywords = template.extract_template_keywords(template.prompt)
+        inputs = [
+            "input_ids", "token_type_ids", "position_ids", "attention_mask",
+            "masked_positions"
+        ]
+        if "soft" in keywords:
+            inputs.append("soft_token_ids")
+        if "encoder" in keywords:
+            inputs.append("encoder_ids")
+        verbalizer_path = os.path.join(export_path, "verbalizer_config.json")
+        with open(verbalizer_path, "r") as fp:
+            label_words = json.load(fp)
+            labels = sorted(list(label_words.keys()))
+
+        return template, labels, inputs
 
     def predict(self, input_data: list):
         encoded_inputs = self.preprocess(input_data)
@@ -155,14 +175,27 @@ def _infer(self, input_dict):
         infer_data = self.inference_backend.infer(input_dict)
         return infer_data
 
-    def infer_batch(self, encoded_inputs):
-        num_sample = len(encoded_inputs["input_ids"])
+    def infer_batch(self, inputs):
+        num_sample = len(inputs)
         infer_data = None
         num_infer_data = None
-        for idx in range(0, num_sample, self._batch_size):
-            l, r = idx, idx + self._batch_size
-            keys = encoded_inputs.keys()
-            input_dict = {k: encoded_inputs[k][l:r] for k in keys}
+        for index in range(0, num_sample, self.args.batch_size):
+            left, right = index, index + self.args.batch_size
+            batch_dict = self.collate_fn(inputs[left:right])
+            input_dict = {}
+            for key in self.input_handles:
+                value = batch_dict[key]
+                if key == "attention_mask":
+                    if value.ndim == 2:
+                        value = (1 - value[:, np.newaxis, np.newaxis, :]) * -1e4
+                    elif value.ndim != 4:
+                        raise ValueError(
+                            "Expect attention mask with ndim=2 or 4, but get ndim={}"
+                            .format(value.ndim))
+                    value = value.astype("float32")
+                else:
+                    value = value.astype("int64")
+                input_dict[key] = value
             results = self._infer(input_dict)
             if infer_data is None:
                 infer_data = [[x] for x in results]
@@ -175,16 +208,8 @@ def infer_batch(self, encoded_inputs):
         return infer_data
 
     def preprocess(self, input_data: list):
-        text = [InputExample(text_a=x) for x in input_data]
-        inputs = [self._template.wrap_one_example(x) for x in text]
-        inputs = {
-            "input_ids":
-            np.array([x["input_ids"] for x in inputs], dtype="int64"),
-            "mask_ids":
-            np.array([x["mask_ids"] for x in inputs], dtype="int64"),
-            "soft_token_ids":
-            np.array([x["soft_token_ids"] for x in inputs], dtype="int64")
-        }
+        text = [{"text_a": x} for x in input_data]
+        inputs = [self.template(x) for x in text]
         return inputs
 
     @staticmethod
@@ -197,7 +222,7 @@ def postprocess(self, infer_data):
         label_ids = np.argwhere(probs > threshold)
         labels = [[] for _ in range(probs.shape[0])]
         for idx, label_id in label_ids:
-            labels[idx].append(self._label_list[label_id])
+            labels[idx].append(self.labels[label_id])
         return {"label": labels}
 
     def printer(self, result, input_data):
@@ -212,12 +237,10 @@ def printer(self, result, input_data):
     for arg_name, arg_value in vars(args).items():
         logger.info("{:20}: {}".format(arg_name, arg_value))
 
-    export_path = os.path.dirname(args.model_path_prefix)
-    labels, _ = Verbalizer.load_from(export_path)
+    predictor = HierachicalPredictor(args)
 
     text_dir = os.path.join(args.data_dir, "data.txt")
     with open(text_dir, "r", encoding="utf-8") as f:
         text_list = [x.strip() for x in f.readlines()]
 
-    predictor = HierachicalPredictor(args, labels)
     predictor.predict(text_list)

applications/text_classification/hierarchical/few-shot/requirements_cpu.txt

@@ -0,0 +1,5 @@
+psutil
+paddlepaddle>=2.4rc
+paddlenlp>=2.4.3
+paddle2onnx>=1.0.3
+onnxruntime

applications/text_classification/hierarchical/few-shot/requirements_gpu.txt

@@ -0,0 +1,7 @@
+psutil
+paddlepaddle-gpu>=2.4rc
+paddlenlp>=2.4.3
+paddle2onnx>=1.0.3
+onnxruntime-gpu
+onnx
+onnxconverter-common

applications/text_classification/hierarchical/few-shot/train.py

@@ -15,6 +15,7 @@
 from dataclasses import dataclass, field
 import os
 import sys
+from collections import defaultdict
 
 import paddle
 import paddle.nn.functional as F
@@ -41,9 +42,6 @@
 class DataArguments:
     data_dir: str = field(default="./data", metadata={"help": "The dataset dictionary includes train.txt, dev.txt, test.txt, label.txt and data.txt (optional) files."})
     prompt: str = field(default=None, metadata={"help": "The input prompt for tuning."})
-    soft_encoder: str = field(default="lstm", metadata={"help": "The encoder type of soft template, `lstm`, `mlp` or None."})
-    encoder_hidden_size: int = field(default=200, metadata={"help": "The dimension of soft embeddings."})
-
 
 @dataclass
 class ModelArguments:
@@ -67,17 +65,20 @@ def main():
     tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path)
 
     # Define the template for preprocess and the verbalizer for postprocess.
-    template = AutoTemplate.create_from(
-        data_args.prompt,
-        tokenizer,
-        training_args.max_seq_length,
-        model=model,
-        prompt_encoder=data_args.soft_encoder,
-        encoder_hidden_size=data_args.encoder_hidden_size)
-    logger.info("Using template: {}".format(template.template))
+    template = AutoTemplate.create_from(data_args.prompt,
+                                        tokenizer,
+                                        training_args.max_seq_length,
+                                        model=model)
+    logger.info("Using template: {}".format(template.prompt))
 
     label_file = os.path.join(data_args.data_dir, "label.txt")
-    verbalizer = SoftVerbalizer.from_file(tokenizer, model, label_file)
+    with open(label_file, "r", encoding="utf-8") as fp:
+        label_words = defaultdict(list)
+        for line in fp:
+            data = line.strip().split("==")
+            word = data[1] if len(data) > 1 else data[0].split("##")[-1]
+            label_words[data[0]].append(word)
+    verbalizer = SoftVerbalizer(label_words, tokenizer, model)
 
     # Load the few-shot datasets.
     train_ds, dev_ds, test_ds = load_local_dataset(
@@ -139,11 +140,24 @@ def compute_metrics(eval_preds):
 
     # Export static model.
     if training_args.do_export:
+        template = prompt_model.template
+        template_keywords = template.extract_template_keywords(template.prompt)
         input_spec = [
-            InputSpec(shape=[None, None], dtype="int64"),  # input_ids
-            InputSpec(shape=[None, None], dtype="int64"),  # mask_ids
-            InputSpec(shape=[None, None], dtype="int64"),  # soft_token_ids
+            InputSpec(shape=[None, None], dtype="int64"),  # input_ids,
+            InputSpec(shape=[None, None], dtype="int64"),  # token_type_ids
+            InputSpec(shape=[None, None], dtype="int64"),  # position_ids
+            InputSpec(shape=[None, None, None, None],
+                      dtype="float32")  # attention_mask
         ]
+        if "mask" in template_keywords:
+            input_spec.append(InputSpec(shape=[None],
+                                        dtype="int64"))  # masked_positions
+        if "soft" in template_keywords:
+            input_spec.append(InputSpec(shape=[None, None],
+                                        dtype="int64"))  # soft_token_ids
+        if "encoder" in template_keywords:
+            input_spec.append(InputSpec(shape=[None, None],
+                                        dtype="int64"))  # encoder_ids
         export_path = os.path.join(training_args.output_dir, 'export')
         trainer.export_model(export_path,
                              input_spec=input_spec,

applications/text_classification/hierarchical/few-shot/utils.py

@@ -15,7 +15,6 @@
 import os
 
 from paddlenlp.datasets import load_dataset
-from paddlenlp.prompt import InputExample
 
 
 def load_local_dataset(data_path, splits, label_list):
@@ -39,14 +38,14 @@ def _reader(data_file, label_list):
             for idx, line in enumerate(fp):
                 data = line.strip().split("\t")
                 if len(data) == 1:
-                    yield InputExample(text_a=data[0])
+                    yield {"text_a": data[0]}
                 else:
                     text, label = data
                     label = label.strip().split(",")
                     label = [
                         float(1) if x in label else float(0) for x in label_list
                     ]
-                    yield InputExample(text_a=text, labels=label)
+                    yield {"text_a": text, "labels": label}
 
     split_map = {"train": "train.txt", "dev": "dev.txt", "test": "test.txt"}
     datasets = []