test=develop (#4255)

dygraph/bert/README.md

+# BERT on PaddlePaddle
+
+[BERT](https://arxiv.org/abs/1810.04805) 是一个迁移能力很强的通用语义表示模型， 以 [Transformer](https://arxiv.org/abs/1706.03762) 为网络基本组件，以双向 `Masked Language Model`  
+和 `Next Sentence Prediction` 为训练目标，通过预训练得到通用语义表示，再结合简单的输出层，应用到下游的 NLP 任务，在多个任务上取得了 SOTA 的结果。本项目是 BERT 在 Paddle Fluid 上的开源实现。
+
+同时推荐用户参考[ IPython Notebook demo](https://aistudio.baidu.com/aistudio/projectDetail/122282)
+
+### 发布要点
+
+1) 动态图BERT模型
+
+2）目前仅支持fine-tuning任务，后续会开展对pre-training任务的支持
+
+3）数据集目前验证了glue上的部分任务，squad上的任务后续会进行验证
+
+4）目前暂不支持FP16/FP32混合精度训练。
+
+| Model | Layers | Hidden size | Heads |Parameters |
+| :------| :------: | :------: |:------: |:------: |
+| [BERT-Base, Uncased](https://baidu-nlp.bj.bcebos.com/DYGRAPH_models%2FBERT%2Fdata.tar.gz) | 12 | 768 |12 |110M |
+
+每个压缩包都包含了模型配置文件 `bert_config.json`、参数文件夹 `params`、动态图参数文件夹`dygraph_params` 和词汇表 `vocab.txt`；
+
+## 内容速览
+- [**安装**](#安装)
+- [**Fine-Tuning**: 预训练模型如何应用到特定 NLP 任务上](#nlp-任务的-fine-tuning)
+  - [语句和句对分类任务](#语句和句对分类任务)
+
+## 目录结构
+```text
+.
+├── data                                        # 示例数据
+├── model                                       # 模型定义
+├── reader                                      # 数据读取
+├── utils                                       # 辅助文件
+├── batching.py                                 # 构建 batch 脚本
+├── optimization.py                             # 优化方法定义
+|── run_classifier.py                           # 分类任务的 fine tuning
+|── tokenization.py                             # 原始文本的 token 化
+|── train.py                                    # 预训练过程的定义
+|── run_classifier_multi_gpu.sh                 # 预训练任务的启动脚本
+|── run_classifier_single_gpu.sh                # 预训练任务的启动脚本
+```
+
+## 安装
+本项目依赖于 Paddle Fluid **1.7.0** 及以上版本，请参考[安装指南](http://www.paddlepaddle.org/#quick-start)进行安装。
+
+## NLP 任务的 Fine-tuning
+
+在完成 BERT 模型的预训练后，即可利用预训练参数在特定的 NLP 任务上做 Fine-tuning。以下利用开源的预训练模型，示例如何进行分类任务和阅读理解任务的 Fine-tuning，如果要运行这些任务，请通过 [发布要点](#发布要点) 一节提供的链接预先下载好对应的预训练模型。
+
+### 语句和句对分类任务
+
+对于 [GLUE 数据](https://gluebenchmark.com/tasks)，请下载[文件](https://baidu-nlp.bj.bcebos.com/DYGRAPH_models%2FBERT%2Fdata.tar.gz)，并解压到同一个目录。以 GLUE/MNLI 任务为例，启动 Fine-tuning 的方式如下（也可以直接运行run_classifier_single_gpu.sh）：
+
+```shell
+#!/bin/bash
+
+BERT_BASE_PATH="./data/pretrained_models/uncased_L-12_H-768_A-12/"
+TASK_NAME='MNLI'
+DATA_PATH="./data/glue_data/MNLI/"
+CKPT_PATH="./data/saved_model/mnli_models"
+
+export CUDA_VISIBLE_DEVICES=0
+
+# start fine-tuning
+python run_classifier.py\
+    --task_name ${TASK_NAME} \
+    --use_cuda true \
+    --do_train true \
+    --do_test true \
+    --batch_size 64 \
+    --init_pretraining_params ${BERT_BASE_PATH}/dygraph_params/ \
+    --data_dir ${DATA_PATH} \
+    --vocab_path ${BERT_BASE_PATH}/vocab.txt \
+    --checkpoints ${CKPT_PATH} \
+    --save_steps 1000 \
+    --weight_decay  0.01 \
+    --warmup_proportion 0.1 \
+    --validation_steps 100 \
+    --epoch 3 \
+    --max_seq_len 128 \
+    --bert_config_path ${BERT_BASE_PATH}/bert_config.json \
+    --learning_rate 5e-5 \
+    --skip_steps 10 \
+    --shuffle true
+
+```
+
+这里的 `uncased_L-12_H-768_A-12/` 即是转换后的英文预训练模型，程序会将模型存储在`CKPT_PATH`指定的位置里。
+
+### 使用单机多卡进行fine-tuning
+
+飞桨动态图使用多进程方式进行数据并行和梯度同步，可以参考`run_classifier_multi_gpu.sh`脚本进行单机多卡fine-tuning：
+
+```shell
+#!/bin/bash
+
+BERT_BASE_PATH="./data/pretrained_models/uncased_L-12_H-768_A-12/"
+TASK_NAME='MNLI'
+DATA_PATH="./data/glue_data/MNLI/"
+CKPT_PATH="./data/saved_model/mnli_models"
+GPU_TO_USE="0,1,2,3"
+
+export CUDA_VISIBLE_DEVICES=$GPU_TO_USE
+
+# start fine-tuning
+python -m paddle.distributed.launch --selected_gpus=$GPU_TO_USE --log_dir ./cls_log run_classifier.py \
+    --task_name ${TASK_NAME} \
+    --use_cuda true \
+    --use_data_parallel true \
+    --do_train true \
+    --do_test true \
+    --batch_size 64 \
+    --in_tokens false \
+    --init_pretraining_params ${BERT_BASE_PATH}/dygraph_params/ \
+    --data_dir ${DATA_PATH} \
+    --vocab_path ${BERT_BASE_PATH}/vocab.txt \
+    --checkpoints ${CKPT_PATH} \
+    --save_steps 1000 \
+    --weight_decay  0.01 \
+    --warmup_proportion 0.1 \
+    --validation_steps 100 \
+    --epoch 3 \
+    --max_seq_len 128 \
+    --bert_config_path ${BERT_BASE_PATH}/bert_config.json \
+    --learning_rate 5e-5 \
+    --skip_steps 10 \
+    --shuffle true
+```
+
+### 读取训练好的模型进行预测
+
+可以参考`run_classifier_prediction.sh`脚本，读取训练好的模型进行预测，可参考以下命令：
+
+```shell
+#!/bin/bash
+
+BERT_BASE_PATH="./data/pretrained_models/uncased_L-12_H-768_A-12/"
+TASK_NAME='MNLI'
+DATA_PATH="./data/glue_data/MNLI/"
+CKPT_PATH="./data/saved_model/mnli_models"
+
+export CUDA_VISIBLE_DEVICES=0
+
+# start testing
+python run_classifier.py\
+    --task_name ${TASK_NAME} \
+    --use_cuda true \
+    --do_train false \
+    --do_test true \
+    --batch_size 64 \
+    --in_tokens false \
+    --data_dir ${DATA_PATH} \
+    --vocab_path ${BERT_BASE_PATH}/vocab.txt \
+    --checkpoints ${CKPT_PATH} \
+    --save_steps 1000 \
+    --weight_decay  0.01 \
+    --warmup_proportion 0.1 \
+    --validation_steps 100 \
+    --epoch 3 \
+    --max_seq_len 128 \
+    --bert_config_path ${BERT_BASE_PATH}/bert_config.json \
+    --learning_rate 5e-5 \
+    --skip_steps 10 \
+    --shuffle false
+```

dygraph/bert/batching.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Mask, padding and batching."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+
+def mask(batch_tokens, total_token_num, vocab_size, CLS=1, SEP=2, MASK=3):
+    """
+    Add mask for batch_tokens, return out, mask_label, mask_pos;
+    Note: mask_pos responding the batch_tokens after padded;
+    """
+    max_len = max([len(sent) for sent in batch_tokens])
+    mask_label = []
+    mask_pos = []
+    prob_mask = np.random.rand(total_token_num)
+    # Note: the first token is [CLS], so [low=1]
+    replace_ids = np.random.randint(1, high=vocab_size, size=total_token_num)
+    pre_sent_len = 0
+    prob_index = 0
+    for sent_index, sent in enumerate(batch_tokens):
+        mask_flag = False
+        prob_index += pre_sent_len
+        for token_index, token in enumerate(sent):
+            prob = prob_mask[prob_index + token_index]
+            if prob > 0.15:
+                continue
+            elif 0.03 < prob <= 0.15:
+                # mask
+                if token != SEP and token != CLS:
+                    mask_label.append(sent[token_index])
+                    sent[token_index] = MASK
+                    mask_flag = True
+                    mask_pos.append(sent_index * max_len + token_index)
+            elif 0.015 < prob <= 0.03:
+                # random replace
+                if token != SEP and token != CLS:
+                    mask_label.append(sent[token_index])
+                    sent[token_index] = replace_ids[prob_index + token_index]
+                    mask_flag = True
+                    mask_pos.append(sent_index * max_len + token_index)
+            else:
+                # keep the original token
+                if token != SEP and token != CLS:
+                    mask_label.append(sent[token_index])
+                    mask_pos.append(sent_index * max_len + token_index)
+        pre_sent_len = len(sent)
+
+        # ensure at least mask one word in a sentence
+        while not mask_flag:
+            token_index = int(np.random.randint(1, high=len(sent) - 1, size=1))
+            if sent[token_index] != SEP and sent[token_index] != CLS:
+                mask_label.append(sent[token_index])
+                sent[token_index] = MASK
+                mask_flag = True
+                mask_pos.append(sent_index * max_len + token_index)
+    mask_label = np.array(mask_label).astype("int64").reshape([-1, 1])
+    mask_pos = np.array(mask_pos).astype("int64").reshape([-1, 1])
+    return batch_tokens, mask_label, mask_pos
+
+
+def prepare_batch_data(insts,
+                       total_token_num,
+                       voc_size=0,
+                       pad_id=None,
+                       cls_id=None,
+                       sep_id=None,
+                       mask_id=None,
+                       return_input_mask=True,
+                       return_max_len=True,
+                       return_num_token=False):
+    """
+    1. generate Tensor of data
+    2. generate Tensor of position
+    3. generate self attention mask, [shape: batch_size *  max_len * max_len]
+    """
+
+    batch_src_ids = [inst[0] for inst in insts]
+    batch_sent_ids = [inst[1] for inst in insts]
+    batch_pos_ids = [inst[2] for inst in insts]
+    labels_list = []
+    # compatible with squad, whose example includes start/end positions, 
+    # or unique id
+
+    for i in range(3, len(insts[0]), 1):
+        labels = [inst[i] for inst in insts]
+        labels = np.array(labels).astype("int64").reshape([-1, 1])
+        labels_list.append(labels)
+
+    # First step: do mask without padding
+    if mask_id >= 0:
+        out, mask_label, mask_pos = mask(
+            batch_src_ids,
+            total_token_num,
+            vocab_size=voc_size,
+            CLS=cls_id,
+            SEP=sep_id,
+            MASK=mask_id)
+    else:
+        out = batch_src_ids
+    # Second step: padding
+    src_id, self_input_mask = pad_batch_data(
+        out, pad_idx=pad_id, return_input_mask=True)
+    pos_id = pad_batch_data(
+        batch_pos_ids,
+        pad_idx=pad_id,
+        return_pos=False,
+        return_input_mask=False)
+    sent_id = pad_batch_data(
+        batch_sent_ids,
+        pad_idx=pad_id,
+        return_pos=False,
+        return_input_mask=False)
+
+    if mask_id >= 0:
+        return_list = [
+            src_id, pos_id, sent_id, self_input_mask, mask_label, mask_pos
+        ] + labels_list
+    else:
+        return_list = [src_id, pos_id, sent_id, self_input_mask] + labels_list
+
+    return return_list if len(return_list) > 1 else return_list[0]
+
+
+def pad_batch_data(insts,
+                   pad_idx=0,
+                   return_pos=False,
+                   return_input_mask=False,
+                   return_max_len=False,
+                   return_num_token=False):
+    """
+    Pad the instances to the max sequence length in batch, and generate the
+    corresponding position data and input mask.
+    """
+    return_list = []
+    max_len = max(len(inst) for inst in insts)
+    # Any token included in dict can be used to pad, since the paddings' loss
+    # will be masked out by weights and make no effect on parameter gradients.
+
+    inst_data = np.array([
+        list(inst) + list([pad_idx] * (max_len - len(inst))) for inst in insts
+    ])
+    return_list += [inst_data.astype("int64").reshape([-1, max_len])]
+
+    # position data
+    if return_pos:
+        inst_pos = np.array([
+            list(range(0, len(inst))) + [pad_idx] * (max_len - len(inst))
+            for inst in insts
+        ])
+
+        return_list += [inst_pos.astype("int64").reshape([-1, max_len])]
+
+    if return_input_mask:
+        # This is used to avoid attention on paddings.
+        input_mask_data = np.array([[1] * len(inst) + [0] *
+                                    (max_len - len(inst)) for inst in insts])
+        input_mask_data = np.expand_dims(input_mask_data, axis=-1)
+        return_list += [input_mask_data.astype("float32")]
+
+    if return_max_len:
+        return_list += [max_len]
+
+    if return_num_token:
+        num_token = 0
+        for inst in insts:
+            num_token += len(inst)
+        return_list += [num_token]
+
+    return return_list if len(return_list) > 1 else return_list[0]
+
+
+if __name__ == "__main__":
+    pass

dygraph/bert/model/bert.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"dygraph transformer layers"
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import six
+import json
+import numpy as np
+
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid.dygraph import Embedding, LayerNorm, Linear, to_variable, Layer, guard
+
+from model.transformer_encoder import EncoderLayer, PrePostProcessLayer
+
+
+class BertConfig(object):
+    def __init__(self, config_path):
+        self._config_dict = self._parse(config_path)
+
+    def _parse(self, config_path):
+        try:
+            with open(config_path) as json_file:
+                config_dict = json.load(json_file)
+        except Exception:
+            raise IOError("Error in parsing bert model config file '%s'" %
+                          config_path)
+        else:
+            return config_dict
+
+    def __getitem__(self, key):
+        return self._config_dict[key]
+
+    def print_config(self):
+        for arg, value in sorted(six.iteritems(self._config_dict)):
+            print('%s: %s' % (arg, value))
+        print('------------------------------------------------')
+
+
+class BertModelLayer(Layer):
+    """
+    bert
+    """
+
+    def __init__(self, config, return_pooled_out=True, use_fp16=False):
+        super(BertModelLayer, self).__init__()
+
+        self._emb_size = config['hidden_size']
+        self._n_layer = config['num_hidden_layers']
+        self._n_head = config['num_attention_heads']
+        self._voc_size = config['vocab_size']
+        self._max_position_seq_len = config['max_position_embeddings']
+        self._sent_types = config['type_vocab_size']
+        self._hidden_act = config['hidden_act']
+        self._prepostprocess_dropout = config['hidden_dropout_prob']
+        self._attention_dropout = config['attention_probs_dropout_prob']
+        self.return_pooled_out = return_pooled_out
+
+        self._word_emb_name = "word_embedding"
+        self._pos_emb_name = "pos_embedding"
+        self._sent_emb_name = "sent_embedding"
+        self._dtype = "float16" if use_fp16 else "float32"
+
+        self._param_initializer = fluid.initializer.TruncatedNormal(
+            scale=config['initializer_range'])
+
+        self._src_emb = Embedding(
+            size=[self._voc_size, self._emb_size],
+            param_attr=fluid.ParamAttr(
+                name=self._word_emb_name, initializer=self._param_initializer),
+            dtype=self._dtype)
+
+        self._pos_emb = Embedding(
+            size=[self._max_position_seq_len, self._emb_size],
+            param_attr=fluid.ParamAttr(
+                name=self._pos_emb_name, initializer=self._param_initializer),
+            dtype=self._dtype)
+
+        self._sent_emb = Embedding(
+            size=[self._sent_types, self._emb_size],
+            param_attr=fluid.ParamAttr(
+                name=self._sent_emb_name, initializer=self._param_initializer),
+            dtype=self._dtype)
+
+        self.pooled_fc = Linear(
+            input_dim=self._emb_size,
+            output_dim=self._emb_size,
+            param_attr=fluid.ParamAttr(
+                name="pooled_fc.w_0", initializer=self._param_initializer),
+            bias_attr="pooled_fc.b_0",
+            act="tanh")
+
+        self.pre_process_layer = PrePostProcessLayer(
+            "nd", self._emb_size, self._prepostprocess_dropout, "")
+
+        self._encoder = EncoderLayer(
+            hidden_act=self._hidden_act,
+            n_layer=self._n_layer,
+            n_head=self._n_head,
+            d_key=self._emb_size // self._n_head,
+            d_value=self._emb_size // self._n_head,
+            d_model=self._emb_size,
+            d_inner_hid=self._emb_size * 4,
+            prepostprocess_dropout=self._prepostprocess_dropout,
+            attention_dropout=self._attention_dropout,
+            relu_dropout=0,
+            preprocess_cmd="",
+            postprocess_cmd="dan",
+            param_initializer=self._param_initializer)
+
+    def forward(self, src_ids, position_ids, sentence_ids, input_mask):
+        """
+        forward
+        """
+        src_emb = self._src_emb(src_ids)
+        pos_emb = self._pos_emb(position_ids)
+        sent_emb = self._sent_emb(sentence_ids)
+
+        emb_out = src_emb + pos_emb
+        emb_out = emb_out + sent_emb
+
+        emb_out = self.pre_process_layer(emb_out)
+
+        self_attn_mask = fluid.layers.matmul(
+            x=input_mask, y=input_mask, transpose_y=True)
+        self_attn_mask = fluid.layers.scale(
+            x=self_attn_mask, scale=10000.0, bias=-1.0, bias_after_scale=False)
+        n_head_self_attn_mask = fluid.layers.stack(
+            x=[self_attn_mask] * self._n_head, axis=1)
+        n_head_self_attn_mask.stop_gradient = True
+
+        enc_output = self._encoder(emb_out, n_head_self_attn_mask)
+
+        if not self.return_pooled_out:
+            return enc_output
+        next_sent_feat = fluid.layers.slice(
+            input=enc_output, axes=[1], starts=[0], ends=[1])
+        next_sent_feat = self.pooled_fc(next_sent_feat)
+        next_sent_feat = fluid.layers.reshape(
+            next_sent_feat, shape=[-1, self._emb_size])
+
+        return enc_output, next_sent_feat
+
+
+class PretrainModelLayer(Layer):
+    """
+    pretrain model
+    """
+
+    def __init__(self,
+                 config,
+                 return_pooled_out=True,
+                 weight_sharing=True,
+                 use_fp16=False):
+        super(PretrainModelLayer, self).__init__()
+        self.config = config
+        self._voc_size = config['vocab_size']
+        self._emb_size = config['hidden_size']
+        self._hidden_act = config['hidden_act']
+        self._prepostprocess_dropout = config['hidden_dropout_prob']
+
+        self._word_emb_name = "word_embedding"
+        self._param_initializer = fluid.initializer.TruncatedNormal(
+            scale=config['initializer_range'])
+        self._weight_sharing = weight_sharing
+        self.use_fp16 = use_fp16
+        self._dtype = "float16" if use_fp16 else "float32"
+
+        self.bert_layer = BertModelLayer(
+            config=self.config, return_pooled_out=True, use_fp16=self.use_fp16)
+
+        self.pre_process_layer = PrePostProcessLayer(
+            "n", self._emb_size, self._prepostprocess_dropout, "pre_encoder")
+
+        self.pooled_fc = Linear(
+            input_dim=self._emb_size,
+            output_dim=self._emb_size,
+            param_attr=fluid.ParamAttr(
+                name="mask_lm_trans_fc.w_0",
+                initializer=self._param_initializer),
+            bias_attr="mask_lm_trans_fc.b_0",
+            act="tanh")
+
+        self.mask_lm_out_bias_attr = fluid.ParamAttr(
+            name="mask_lm_out_fc.b_0",
+            initializer=fluid.initializer.Constant(value=0.0))
+
+        if not self._weight_sharing:
+            self.out_fc = Linear(
+                input_dim=self._emb_size,
+                output_dim=self._voc_size,
+                param_attr=fluid.ParamAttr(
+                    name="mask_lm_out_fc.w_0",
+                    initializer=self._param_initializer),
+                bias_attr=self.mask_lm_out_bias_attr)
+        else:
+            self.fc_create_params = self.create_parameter(
+                shape=[self._voc_size],
+                dtype=self._dtype,
+                attr=self.mask_lm_out_bias_attr,
+                is_bias=True)
+
+        self.next_sent_fc = Linear(
+            input_dim=self._emb_size,
+            output_dim=2,
+            param_attr=fluid.ParamAttr(
+                name="next_sent_fc.w_0", initializer=self._param_initializer),
+            bias_attr="next_sent_fc.b_0")
+
+    def forward(self, src_ids, position_ids, sentence_ids, input_mask,
+                mask_label, mask_pos, labels):
+        """
+        forward
+        """
+        mask_pos = fluid.layers.cast(x=mask_pos, dtype='int32')
+
+        enc_output, next_sent_feat = self.bert_layer(src_ids, position_ids,
+                                                     sentence_ids, input_mask)
+        reshaped_emb_out = fluid.layers.reshape(
+            x=enc_output, shape=[-1, self._emb_size])
+
+        mask_feat = fluid.layers.gather(input=reshaped_emb_out, index=mask_pos)
+
+        mask_trans_feat = self.pooled_fc(mask_feat)
+        mask_trans_feat = self.pre_process_layer(None, mask_trans_feat, "n",
+                                                 self._prepostprocess_dropout)
+
+        if self._weight_sharing:
+            fc_out = fluid.layers.matmul(
+                x=mask_trans_feat,
+                y=self.bert_layer._src_emb._w,
+                transpose_y=True)
+            fc_out += self.fc_create_params
+        else:
+            fc_out = self.out_fc(mask_trans_feat)
+
+        mask_lm_loss = fluid.layers.softmax_with_cross_entropy(
+            logits=fc_out, label=mask_label)
+        mean_mask_lm_loss = fluid.layers.mean(mask_lm_loss)
+
+        next_sent_fc_out = self.next_sent_fc(next_sent_feat)
+
+        next_sent_loss, next_sent_softmax = fluid.layers.softmax_with_cross_entropy(
+            logits=next_sent_fc_out, label=labels, return_softmax=True)
+
+        next_sent_acc = fluid.layers.accuracy(
+            input=next_sent_softmax, label=labels)
+
+        mean_next_sent_loss = fluid.layers.mean(next_sent_loss)
+
+        loss = mean_next_sent_loss + mean_mask_lm_loss
+        return next_sent_acc, mean_mask_lm_loss, loss

dygraph/bert/model/cls.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"dygraph transformer layers"
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import six
+import json
+import numpy as np
+
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid.dygraph import Linear, Layer
+
+from model.bert import BertModelLayer
+
+
+class ClsModelLayer(Layer):
+    """
+    classify model
+    """
+
+    def __init__(self,
+                 args,
+                 config,
+                 num_labels,
+                 is_training=True,
+                 return_pooled_out=True,
+                 use_fp16=False):
+        super(ClsModelLayer, self).__init__()
+        self.config = config
+        self.is_training = is_training
+        self.use_fp16 = use_fp16
+        self.loss_scaling = args.loss_scaling
+
+        self.bert_layer = BertModelLayer(
+            config=self.config, return_pooled_out=True, use_fp16=self.use_fp16)
+
+        self.cls_fc = Linear(
+            input_dim=self.config["hidden_size"],
+            output_dim=num_labels,
+            param_attr=fluid.ParamAttr(
+                name="cls_out_w",
+                initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
+            bias_attr=fluid.ParamAttr(
+                name="cls_out_b", initializer=fluid.initializer.Constant(0.)))
+
+    def forward(self, data_ids):
+        """
+        forward
+        """
+        src_ids = data_ids[0]
+        position_ids = data_ids[1]
+        sentence_ids = data_ids[2]
+        input_mask = data_ids[3]
+        labels = data_ids[4]
+
+        enc_output, next_sent_feat = self.bert_layer(src_ids, position_ids,
+                                                     sentence_ids, input_mask)
+
+        cls_feats = fluid.layers.dropout(
+            x=next_sent_feat,
+            dropout_prob=0.1,
+            dropout_implementation="upscale_in_train")
+
+        logits = self.cls_fc(cls_feats)
+
+        ce_loss, probs = fluid.layers.softmax_with_cross_entropy(
+            logits=logits, label=labels, return_softmax=True)
+        loss = fluid.layers.mean(x=ce_loss)
+
+        if self.use_fp16 and self.loss_scaling > 1.0:
+            loss *= self.loss_scaling
+
+        num_seqs = fluid.layers.create_tensor(dtype='int64')
+        accuracy = fluid.layers.accuracy(
+            input=probs, label=labels, total=num_seqs)
+
+        return loss, accuracy, num_seqs

dygraph/bert/model/transformer_encoder.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"dygraph transformer layers"
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid.dygraph import Embedding, LayerNorm, Linear, Layer
+
+
+class PrePostProcessLayer(Layer):
+    """
+    PrePostProcessLayer
+    """
+
+    def __init__(self, process_cmd, d_model, dropout_rate, name):
+        super(PrePostProcessLayer, self).__init__()
+        self.process_cmd = process_cmd
+        self.functors = []
+        self.exec_order = ""
+
+        for cmd in self.process_cmd:
+            if cmd == "a":  # add residual connection
+                self.functors.append(lambda x, y: x + y if y else x)
+                self.exec_order += "a"
+            elif cmd == "n":  # add layer normalization
+                self.functors.append(
+                    self.add_sublayer(
+                        "layer_norm_%d" % len(
+                            self.sublayers(include_sublayers=False)),
+                        LayerNorm(
+                            normalized_shape=d_model,
+                            param_attr=fluid.ParamAttr(
+                                name=name + "_layer_norm_scale",
+                                initializer=fluid.initializer.Constant(1.)),
+                            bias_attr=fluid.ParamAttr(
+                                name=name + "_layer_norm_bias",
+                                initializer=fluid.initializer.Constant(0.)))))
+                self.exec_order += "n"
+            elif cmd == "d":  # add dropout
+                if dropout_rate:
+                    self.functors.append(lambda x: fluid.layers.dropout(
+                        x, dropout_prob=dropout_rate, is_test=False))
+                    self.exec_order += "d"
+
+    def forward(self, x, residual=None):
+        for i, cmd in enumerate(self.exec_order):
+            if cmd == "a":
+                x = self.functors[i](x, residual)
+            else:
+                x = self.functors[i](x)
+        return x
+
+
+class PositionwiseFeedForwardLayer(Layer):
+    """
+    PositionwiseFeedForwardLayer
+    """
+
+    def __init__(self,
+                 hidden_act,
+                 d_inner_hid,
+                 d_model,
+                 dropout_rate,
+                 param_initializer=None,
+                 name=""):
+        super(PositionwiseFeedForwardLayer, self).__init__()
+
+        self._i2h = Linear(
+            input_dim=d_model,
+            output_dim=d_inner_hid,
+            param_attr=fluid.ParamAttr(
+                name=name + '_fc_0.w_0', initializer=param_initializer),
+            bias_attr=name + '_fc_0.b_0',
+            act=hidden_act)
+
+        self._h2o = Linear(
+            input_dim=d_inner_hid,
+            output_dim=d_model,
+            param_attr=fluid.ParamAttr(
+                name=name + '_fc_1.w_0', initializer=param_initializer),
+            bias_attr=name + '_fc_1.b_0')
+
+        self._dropout_rate = dropout_rate
+
+    def forward(self, x):
+        """
+        forward
+        :param x:
+        :return:
+        """
+        hidden = self._i2h(x)
+        if self._dropout_rate:
+            hidden = fluid.layers.dropout(
+                hidden,
+                dropout_prob=self._dropout_rate,
+                upscale_in_train="upscale_in_train",
+                is_test=False)
+        out = self._h2o(hidden)
+        return out
+
+
+class MultiHeadAttentionLayer(Layer):
+    """
+    MultiHeadAttentionLayer
+    """
+
+    def __init__(self,
+                 d_key,
+                 d_value,
+                 d_model,
+                 n_head=1,
+                 dropout_rate=0.,
+                 cache=None,
+                 gather_idx=None,
+                 static_kv=False,
+                 param_initializer=None,
+                 name=""):
+        super(MultiHeadAttentionLayer, self).__init__()
+        self._n_head = n_head
+        self._d_key = d_key
+        self._d_value = d_value
+        self._d_model = d_model
+        self._dropout_rate = dropout_rate
+
+        self._q_fc = Linear(
+            input_dim=d_model,
+            output_dim=d_key * n_head,
+            param_attr=fluid.ParamAttr(
+                name=name + '_query_fc.w_0', initializer=param_initializer),
+            bias_attr=name + '_query_fc.b_0')
+
+        self._k_fc = Linear(
+            input_dim=d_model,
+            output_dim=d_key * n_head,
+            param_attr=fluid.ParamAttr(
+                name=name + '_key_fc.w_0', initializer=param_initializer),
+            bias_attr=name + '_key_fc.b_0')
+
+        self._v_fc = Linear(
+            input_dim=d_model,
+            output_dim=d_value * n_head,
+            param_attr=fluid.ParamAttr(
+                name=name + '_value_fc.w_0', initializer=param_initializer),
+            bias_attr=name + '_value_fc.b_0')
+
+        self._proj_fc = Linear(
+            input_dim=d_value * n_head,
+            output_dim=d_model,
+            param_attr=fluid.ParamAttr(
+                name=name + '_output_fc.w_0', initializer=param_initializer),
+            bias_attr=name + '_output_fc.b_0')
+
+    def forward(self, queries, keys, values, attn_bias):
+        """
+        forward
+        :param queries:
+        :param keys:
+        :param values:
+        :param attn_bias:
+        :return:
+        """
+        # compute q ,k ,v
+        keys = queries if keys is None else keys
+        values = keys if values is None else values
+
+        q = self._q_fc(queries)
+        k = self._k_fc(keys)
+        v = self._v_fc(values)
+
+        # split head
+
+        q_hidden_size = q.shape[-1]
+        reshaped_q = fluid.layers.reshape(
+            x=q,
+            shape=[0, 0, self._n_head, q_hidden_size // self._n_head],
+            inplace=False)
+        transpose_q = fluid.layers.transpose(x=reshaped_q, perm=[0, 2, 1, 3])
+
+        k_hidden_size = k.shape[-1]
+        reshaped_k = fluid.layers.reshape(
+            x=k,
+            shape=[0, 0, self._n_head, k_hidden_size // self._n_head],
+            inplace=False)
+        transpose_k = fluid.layers.transpose(x=reshaped_k, perm=[0, 2, 1, 3])
+
+        v_hidden_size = v.shape[-1]
+        reshaped_v = fluid.layers.reshape(
+            x=v,
+            shape=[0, 0, self._n_head, v_hidden_size // self._n_head],
+            inplace=False)
+        transpose_v = fluid.layers.transpose(x=reshaped_v, perm=[0, 2, 1, 3])
+
+        scaled_q = fluid.layers.scale(x=transpose_q, scale=self._d_key**-0.5)
+        # scale dot product attention
+        product = fluid.layers.matmul(
+            #x=transpose_q,
+            x=scaled_q,
+            y=transpose_k,
+            transpose_y=True)
+        #alpha=self._d_model**-0.5)
+        if attn_bias:
+            product += attn_bias
+        weights = fluid.layers.softmax(product)
+        if self._dropout_rate:
+            weights_droped = fluid.layers.dropout(
+                weights,
+                dropout_prob=self._dropout_rate,
+                dropout_implementation="upscale_in_train",
+                is_test=False)
+            out = fluid.layers.matmul(weights_droped, transpose_v)
+        else:
+            out = fluid.layers.matmul(weights, transpose_v)
+
+        # combine heads
+        if len(out.shape) != 4:
+            raise ValueError("Input(x) should be a 4-D Tensor.")
+        trans_x = fluid.layers.transpose(out, perm=[0, 2, 1, 3])
+        final_out = fluid.layers.reshape(
+            x=trans_x,
+            shape=[0, 0, trans_x.shape[2] * trans_x.shape[3]],
+            inplace=False)
+
+        # fc to output
+        proj_out = self._proj_fc(final_out)
+        return proj_out
+
+
+class EncoderSubLayer(Layer):
+    """
+    EncoderSubLayer
+    """
+
+    def __init__(self,
+                 hidden_act,
+                 n_head,
+                 d_key,
+                 d_value,
+                 d_model,
+                 d_inner_hid,
+                 prepostprocess_dropout,
+                 attention_dropout,
+                 relu_dropout,
+                 preprocess_cmd="n",
+                 postprocess_cmd="da",
+                 param_initializer=None,
+                 name=""):
+
+        super(EncoderSubLayer, self).__init__()
+        self.name = name
+        self._preprocess_cmd = preprocess_cmd
+        self._postprocess_cmd = postprocess_cmd
+        self._prepostprocess_dropout = prepostprocess_dropout
+
+        self._preprocess_layer = PrePostProcessLayer(
+            self._preprocess_cmd,
+            d_model,
+            prepostprocess_dropout,
+            name=name + "_pre_att")
+
+        self._multihead_attention_layer = MultiHeadAttentionLayer(
+            d_key,
+            d_value,
+            d_model,
+            n_head,
+            attention_dropout,
+            None,
+            None,
+            False,
+            param_initializer,
+            name=name + "_multi_head_att")
+
+        self._postprocess_layer = PrePostProcessLayer(
+            self._postprocess_cmd,
+            d_model,
+            self._prepostprocess_dropout,
+            name=name + "_post_att")
+        self._preprocess_layer2 = PrePostProcessLayer(
+            self._preprocess_cmd,
+            d_model,
+            self._prepostprocess_dropout,
+            name=name + "_pre_ffn")
+
+        self._positionwise_feed_forward = PositionwiseFeedForwardLayer(
+            hidden_act,
+            d_inner_hid,
+            d_model,
+            relu_dropout,
+            param_initializer,
+            name=name + "_ffn")
+
+        self._postprocess_layer2 = PrePostProcessLayer(
+            self._postprocess_cmd,
+            d_model,
+            self._prepostprocess_dropout,
+            name=name + "_post_ffn")
+
+    def forward(self, enc_input, attn_bias):
+        """
+        forward
+        :param enc_input:
+        :param attn_bias:
+        :return:
+        """
+        pre_process_multihead = self._preprocess_layer(enc_input)
+
+        attn_output = self._multihead_attention_layer(pre_process_multihead,
+                                                      None, None, attn_bias)
+        attn_output = self._postprocess_layer(attn_output, enc_input)
+
+        pre_process2_output = self._preprocess_layer2(attn_output)
+
+        ffd_output = self._positionwise_feed_forward(pre_process2_output)
+
+        return self._postprocess_layer2(ffd_output, attn_output)
+
+
+class EncoderLayer(Layer):
+    """
+    encoder
+    """
+
+    def __init__(self,
+                 hidden_act,
+                 n_layer,
+                 n_head,
+                 d_key,
+                 d_value,
+                 d_model,
+                 d_inner_hid,
+                 prepostprocess_dropout,
+                 attention_dropout,
+                 relu_dropout,
+                 preprocess_cmd="n",
+                 postprocess_cmd="da",
+                 param_initializer=None,
+                 name=""):
+
+        super(EncoderLayer, self).__init__()
+        self._preprocess_cmd = preprocess_cmd
+        self._encoder_sublayers = list()
+        self._prepostprocess_dropout = prepostprocess_dropout
+        self._n_layer = n_layer
+        self._hidden_act = hidden_act
+        self._preprocess_layer = PrePostProcessLayer(
+            self._preprocess_cmd, 3, self._prepostprocess_dropout,
+            "post_encoder")
+
+        for i in range(n_layer):
+            self._encoder_sublayers.append(
+                self.add_sublayer(
+                    'esl_%d' % i,
+                    EncoderSubLayer(
+                        hidden_act,
+                        n_head,
+                        d_key,
+                        d_value,
+                        d_model,
+                        d_inner_hid,
+                        prepostprocess_dropout,
+                        attention_dropout,
+                        relu_dropout,
+                        preprocess_cmd,
+                        postprocess_cmd,
+                        param_initializer,
+                        name=name + '_layer_' + str(i))))
+
+    def forward(self, enc_input, attn_bias):
+        """
+        forward
+        :param enc_input:
+        :param attn_bias:
+        :return:
+        """
+        for i in range(self._n_layer):
+            enc_output = self._encoder_sublayers[i](enc_input, attn_bias)
+            enc_input = enc_output
+
+        return self._preprocess_layer(enc_output)

dygraph/bert/optimization.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Optimization and learning rate scheduling."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import paddle.fluid as fluid
+
+from paddle.fluid.dygraph.learning_rate_scheduler import LearningRateDecay
+
+
+class ConstantLR(LearningRateDecay):
+    def __init__(self, learning_rate, begin=0, step=1, dtype='float32'):
+        super(ConstantLR, self).__init__(begin, step, dtype)
+        self.learning_rate = learning_rate
+
+    def step(self):
+        return self.learning_rate
+
+
+class LinearDecay(LearningRateDecay):
+    def __init__(self,
+                 learning_rate,
+                 warmup_steps,
+                 decay_steps,
+                 end_learning_rate=0.0001,
+                 power=1.0,
+                 cycle=False,
+                 begin=0,
+                 step=1,
+                 dtype='float32'):
+        super(LinearDecay, self).__init__(begin, step, dtype)
+        self.learning_rate = learning_rate
+        self.warmup_steps = warmup_steps
+        self.decay_steps = decay_steps
+        self.end_learning_rate = end_learning_rate
+        self.power = power
+        self.cycle = cycle
+
+    def step(self):
+        if self.step_num < self.warmup_steps:
+            decayed_lr = self.learning_rate * (self.step_num /
+                                               self.warmup_steps)
+            decayed_lr = self.create_lr_var(decayed_lr)
+        else:
+            tmp_step_num = self.step_num
+            tmp_decay_steps = self.decay_steps
+            if self.cycle:
+                div_res = fluid.layers.ceil(
+                    self.create_lr_var(tmp_step_num / float(self.decay_steps)))
+                if tmp_step_num == 0:
+                    div_res = self.create_lr_var(1.0)
+                tmp_decay_steps = self.decay_steps * div_res
+            else:
+                tmp_step_num = self.create_lr_var(
+                    tmp_step_num
+                    if tmp_step_num < self.decay_steps else self.decay_steps)
+                decayed_lr = (self.learning_rate - self.end_learning_rate) * \
+                    ((1 - tmp_step_num / tmp_decay_steps) ** self.power) + self.end_learning_rate
+
+        return decayed_lr
+
+
+class Optimizer(object):
+    def __init__(self,
+                 warmup_steps,
+                 num_train_steps,
+                 learning_rate,
+                 model_cls,
+                 weight_decay,
+                 scheduler='linear_warmup_decay',
+                 loss_scaling=1.0,
+                 parameter_list=None):
+        self.warmup_steps = warmup_steps
+        self.num_train_steps = num_train_steps
+        self.learning_rate = learning_rate
+        self.model_cls = model_cls
+        self.weight_decay = weight_decay
+        self.scheduler = scheduler
+        self.loss_scaling = loss_scaling
+        self.parameter_list = parameter_list
+
+        self.scheduled_lr = 0.0
+        self.optimizer = self.lr_schedule()
+
+    def lr_schedule(self):
+        if self.warmup_steps > 0:
+            if self.scheduler == 'noam_decay':
+                self.scheduled_lr = fluid.dygraph.NoamDecay(1 / (
+                    self.warmup_steps * (self.learning_rate**2)),
+                                                            self.warmup_steps)
+            elif self.scheduler == 'linear_warmup_decay':
+                self.scheduled_lr = LinearDecay(self.learning_rate,
+                                                self.warmup_steps,
+                                                self.num_train_steps, 0.0)
+            else:
+                raise ValueError("Unkown learning rate scheduler, should be "
+                                 "'noam_decay' or 'linear_warmup_decay'")
+            optimizer = fluid.optimizer.Adam(
+                learning_rate=self.scheduled_lr,
+                parameter_list=self.parameter_list)
+        else:
+            self.scheduled_lr = ConstantLR(self.learning_rate)
+            optimizer = fluid.optimizer.Adam(
+                learning_rate=self.scheduled_lr,
+                parameter_list=self.parameter_list)
+
+        return optimizer
+
+    def exclude_from_weight_decay(self, name):
+        if name.find("layer_norm") > -1:
+            return True
+        bias_suffix = ["_bias", "_b", ".b_0"]
+        for suffix in bias_suffix:
+            if name.endswith(suffix):
+                return True
+        return False
+
+    def optimization(self, loss, use_data_parallel=False, model=None):
+        param_list = dict()
+
+        clip_norm_thres = 1.0
+        #grad_clip = fluid.dygraph_grad_clip.GradClipByGlobalNorm(clip_norm_thres)
+
+        if use_data_parallel:
+            loss = model.scale_loss(loss)
+
+        loss.backward()
+
+        if self.weight_decay > 0:
+            for param in self.model_cls.parameters():
+                param_list[param.name] = param * 1.0
+                param_list[param.name].stop_gradient = True
+
+        if use_data_parallel:
+            assert model is not None
+            model.apply_collective_grads()
+
+        #_, param_grads = self.optimizer.minimize(loss, grad_clip=grad_clip)
+        _, param_grads = self.optimizer.minimize(loss)
+
+        if self.weight_decay > 0:
+            for param, grad in param_grads:
+                if self.exclude_from_weight_decay(param.name):
+                    continue
+                if isinstance(self.scheduled_lr.step(), float):
+                    updated_param = param.numpy() - param_list[
+                        param.name].numpy(
+                        ) * self.weight_decay * self.scheduled_lr.step()
+                else:
+                    updated_param = param.numpy() - param_list[
+                        param.name].numpy(
+                        ) * self.weight_decay * self.scheduled_lr.step().numpy()
+                updated_param_var = fluid.dygraph.to_variable(updated_param)
+                param = updated_param_var
+                #param = fluid.layers.reshape(x=updated_param_var, shape=list(updated_param_var.shape))

dygraph/bert/reader/pretraining.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+from __future__ import division
+
+import os
+import numpy as np
+import types
+import gzip
+import logging
+import re
+import six
+import collections
+import tokenization
+
+import paddle
+import paddle.fluid as fluid
+
+from batching import prepare_batch_data
+
+
+class DataReader(object):
+    def __init__(self,
+                 data_dir,
+                 vocab_path,
+                 batch_size=4096,
+                 in_tokens=True,
+                 max_seq_len=512,
+                 shuffle_files=True,
+                 epoch=100,
+                 voc_size=0,
+                 is_test=False,
+                 generate_neg_sample=False):
+
+        self.vocab = self.load_vocab(vocab_path)
+        self.data_dir = data_dir
+        self.batch_size = batch_size
+        self.in_tokens = in_tokens
+        self.shuffle_files = shuffle_files
+        self.epoch = epoch
+        self.current_epoch = 0
+        self.current_file_index = 0
+        self.total_file = 0
+        self.current_file = None
+        self.voc_size = voc_size
+        self.max_seq_len = max_seq_len
+        self.pad_id = self.vocab["[PAD]"]
+        self.cls_id = self.vocab["[CLS]"]
+        self.sep_id = self.vocab["[SEP]"]
+        self.mask_id = self.vocab["[MASK]"]
+        self.is_test = is_test
+        self.generate_neg_sample = generate_neg_sample
+        if self.in_tokens:
+            assert self.batch_size >= self.max_seq_len, "The number of " \
+                   "tokens in batch should not be smaller than max seq length."
+
+        if self.is_test:
+            self.epoch = 1
+            self.shuffle_files = False
+
+    def get_progress(self):
+        """return current progress of traning data
+        """
+        return self.current_epoch, self.current_file_index, self.total_file, self.current_file
+
+    def parse_line(self, line, max_seq_len=512):
+        """ parse one line to token_ids, sentence_ids, pos_ids, label
+        """
+        line = line.strip().decode().split(";")
+        assert len(line) == 4, "One sample must have 4 fields!"
+        (token_ids, sent_ids, pos_ids, label) = line
+        token_ids = [int(token) for token in token_ids.split(" ")]
+        sent_ids = [int(token) for token in sent_ids.split(" ")]
+        pos_ids = [int(token) for token in pos_ids.split(" ")]
+        assert len(token_ids) == len(sent_ids) == len(
+            pos_ids
+        ), "[Must be true]len(token_ids) == len(sent_ids) == len(pos_ids)"
+        label = int(label)
+        if len(token_ids) > max_seq_len:
+            return None
+        return [token_ids, sent_ids, pos_ids, label]
+
+    def read_file(self, file):
+        assert file.endswith('.gz'), "[ERROR] %s is not a gzip file" % file
+        file_path = self.data_dir + "/" + file
+        with gzip.open(file_path, "rb") as f:
+            for line in f:
+                parsed_line = self.parse_line(
+                    line, max_seq_len=self.max_seq_len)
+                if parsed_line is None:
+                    continue
+                yield parsed_line
+
+    def convert_to_unicode(self, text):
+        """Converts `text` to Unicode (if it's not already), assuming utf-8 input."""
+        if six.PY3:
+            if isinstance(text, str):
+                return text
+            elif isinstance(text, bytes):
+                return text.decode("utf-8", "ignore")
+            else:
+                raise ValueError("Unsupported string type: %s" % (type(text)))
+        elif six.PY2:
+            if isinstance(text, str):
+                return text.decode("utf-8", "ignore")
+            elif isinstance(text, unicode):
+                return text
+            else:
+                raise ValueError("Unsupported string type: %s" % (type(text)))
+        else:
+            raise ValueError("Not running on Python2 or Python 3?")
+
+    def load_vocab(self, vocab_file):
+        """Loads a vocabulary file into a dictionary."""
+        vocab = collections.OrderedDict()
+        fin = open(vocab_file)
+        for num, line in enumerate(fin):
+            items = self.convert_to_unicode(line.strip()).split("\t")
+            if len(items) > 2:
+                break
+            token = items[0]
+            index = items[1] if len(items) == 2 else num
+            token = token.strip()
+            vocab[token] = int(index)
+        return vocab
+
+    def random_pair_neg_samples(self, pos_samples):
+        """ randomly generate negtive samples using pos_samples
+
+            Args:
+                pos_samples: list of positive samples
+            
+            Returns:
+                neg_samples: list of negtive samples
+        """
+        np.random.shuffle(pos_samples)
+        num_sample = len(pos_samples)
+        neg_samples = []
+        miss_num = 0
+
+        for i in range(num_sample):
+            pair_index = (i + 1) % num_sample
+            origin_src_ids = pos_samples[i][0]
+            origin_sep_index = origin_src_ids.index(2)
+            pair_src_ids = pos_samples[pair_index][0]
+            pair_sep_index = pair_src_ids.index(2)
+
+            src_ids = origin_src_ids[:origin_sep_index + 1] + pair_src_ids[
+                pair_sep_index + 1:]
+            if len(src_ids) >= self.max_seq_len:
+                miss_num += 1
+                continue
+            sent_ids = [0] * len(origin_src_ids[:origin_sep_index + 1]) + [
+                1
+            ] * len(pair_src_ids[pair_sep_index + 1:])
+            pos_ids = list(range(len(src_ids)))
+            neg_sample = [src_ids, sent_ids, pos_ids, 0]
+            assert len(src_ids) == len(sent_ids) == len(
+                pos_ids
+            ), "[ERROR]len(src_id) == lne(sent_id) == len(pos_id) must be True"
+            neg_samples.append(neg_sample)
+        return neg_samples, miss_num
+
+    def mixin_negtive_samples(self, pos_sample_generator, buffer=1000):
+        """ 1. generate negtive samples by randomly group sentence_1 and sentence_2 of positive samples
+            2. combine negtive samples and positive samples
+            
+            Args:
+                pos_sample_generator: a generator producing a parsed positive sample, which is a list: [token_ids, sent_ids, pos_ids, 1]
+
+            Returns:
+                sample: one sample from shuffled positive samples and negtive samples
+        """
+        pos_samples = []
+        num_total_miss = 0
+        pos_sample_num = 0
+        try:
+            while True:
+                while len(pos_samples) < buffer:
+                    pos_sample = next(pos_sample_generator)
+                    label = pos_sample[3]
+                    assert label == 1, "positive sample's label must be 1"
+                    pos_samples.append(pos_sample)
+                    pos_sample_num += 1
+
+                neg_samples, miss_num = self.random_pair_neg_samples(
+                    pos_samples)
+                num_total_miss += miss_num
+                samples = pos_samples + neg_samples
+                pos_samples = []
+                np.random.shuffle(samples)
+                for sample in samples:
+                    yield sample
+        except StopIteration:
+            print("stopiteration: reach end of file")
+            if len(pos_samples) == 1:
+                yield pos_samples[0]
+            elif len(pos_samples) == 0:
+                yield None
+            else:
+                neg_samples, miss_num = self.random_pair_neg_samples(
+                    pos_samples)
+                num_total_miss += miss_num
+                samples = pos_samples + neg_samples
+                pos_samples = []
+                np.random.shuffle(samples)
+                for sample in samples:
+                    yield sample
+            print("miss_num:%d\tideal_total_sample_num:%d\tmiss_rate:%f" %
+                  (num_total_miss, pos_sample_num * 2,
+                   num_total_miss / (pos_sample_num * 2)))
+
+    def data_generator(self):
+        """
+        data_generator
+        """
+        files = os.listdir(self.data_dir)
+        self.total_file = len(files)
+        assert self.total_file > 0, "[Error] data_dir is empty"
+
+        def wrapper():
+            def reader():
+                for epoch in range(self.epoch):
+                    self.current_epoch = epoch + 1
+                    if self.shuffle_files:
+                        np.random.shuffle(files)
+                    for index, file in enumerate(files):
+                        self.current_file_index = index + 1
+                        self.current_file = file
+                        sample_generator = self.read_file(file)
+                        if not self.is_test and self.generate_neg_sample:
+                            sample_generator = self.mixin_negtive_samples(
+                                sample_generator)
+                        for sample in sample_generator:
+                            if sample is None:
+                                continue
+                            yield sample
+
+            def batch_reader(reader, batch_size, in_tokens):
+                batch, total_token_num, max_len = [], 0, 0
+                for parsed_line in reader():
+                    token_ids, sent_ids, pos_ids, label = parsed_line
+                    max_len = max(max_len, len(token_ids))
+                    if in_tokens:
+                        to_append = (len(batch) + 1) * max_len <= batch_size
+                    else:
+                        to_append = len(batch) < batch_size
+                    if to_append:
+                        batch.append(parsed_line)
+                        total_token_num += len(token_ids)
+                    else:
+                        yield batch, total_token_num
+                        batch, total_token_num, max_len = [parsed_line], len(
+                            token_ids), len(token_ids)
+
+                if len(batch) > 0:
+                    yield batch, total_token_num
+
+            for batch_data, total_token_num in batch_reader(
+                    reader, self.batch_size, self.in_tokens):
+                yield prepare_batch_data(
+                    batch_data,
+                    total_token_num,
+                    voc_size=self.voc_size,
+                    pad_id=self.pad_id,
+                    cls_id=self.cls_id,
+                    sep_id=self.sep_id,
+                    mask_id=self.mask_id,
+                    return_input_mask=True,
+                    return_max_len=False,
+                    return_num_token=False)
+
+        return wrapper
+
+
+if __name__ == "__main__":
+    pass

dygraph/bert/run_classifier.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""BERT fine-tuning in Paddle Dygraph Mode."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import six
+import sys
+if six.PY2:
+    reload(sys)
+    sys.setdefaultencoding('utf8')
+import ast
+import time
+import argparse
+import numpy as np
+import multiprocessing
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid.dygraph import to_variable
+import reader.cls as reader
+from model.bert import BertConfig
+from model.cls import ClsModelLayer
+from optimization import Optimizer
+from utils.args import ArgumentGroup, print_arguments, check_cuda
+from utils.init import init_from_static_model
+
+# yapf: disable
+parser = argparse.ArgumentParser(__doc__)
+model_g = ArgumentGroup(parser, "model", "model configuration and paths.")
+model_g.add_arg("bert_config_path",      str,  "./config/bert_config.json",  "Path to the json file for bert model config.")
+model_g.add_arg("init_checkpoint",       str,  None,                         "Init checkpoint to resume training from.")
+model_g.add_arg("init_pretraining_params",  str,  None,
+                "Init pre-training params which preforms fine-tuning from. If the "
+                "arg 'init_checkpoint' has been set, this argument wouldn't be valid.")
+model_g.add_arg("checkpoints",           str,  "checkpoints",                "Path to save checkpoints.")
+
+train_g = ArgumentGroup(parser, "training", "training options.")
+train_g.add_arg("epoch",             int,    100,     "Number of epoches for training.")
+train_g.add_arg("learning_rate",     float,  0.0001,  "Learning rate used to train with warmup.")
+train_g.add_arg("lr_scheduler",      str,    "linear_warmup_decay",
+                "scheduler of learning rate.", choices=['linear_warmup_decay', 'noam_decay'])
+train_g.add_arg("weight_decay",      float,  0.01,    "Weight decay rate for L2 regularizer.")
+train_g.add_arg("warmup_proportion",     float,  0.1,                         "Proportion of training steps to perform linear learning rate warmup for.")
+train_g.add_arg("save_steps",        int,    10000,   "The steps interval to save checkpoints.")
+train_g.add_arg("validation_steps",  int,    1000,    "The steps interval to evaluate model performance.")
+train_g.add_arg("loss_scaling",      float,  1.0,
+                "Loss scaling factor for mixed precision training, only valid when use_fp16 is enabled.")
+
+log_g = ArgumentGroup(parser,     "logging", "logging related.")
+log_g.add_arg("skip_steps",          int,    10,    "The steps interval to print loss.")
+
+data_g = ArgumentGroup(parser, "data", "Data paths, vocab paths and data processing options")
+data_g.add_arg("data_dir",            str,  None,       "Path to training data.")
+data_g.add_arg("vocab_path",          str,  None,       "Vocabulary path.")
+data_g.add_arg("max_seq_len",         int,  512,                   "Tokens' number of the longest seqence allowed.")
+data_g.add_arg("batch_size",          int,  32,
+               "The total number of examples in one batch for training, see also --in_tokens.")
+data_g.add_arg("in_tokens",           bool, False,
+               "If set, the batch size will be the maximum number of tokens in one batch. "
+               "Otherwise, it will be the maximum number of examples in one batch.")
+data_g.add_arg("do_lower_case", bool, True,
+               "Whether to lower case the input text. Should be True for uncased models and False for cased models.")
+data_g.add_arg("random_seed",   int,  5512,     "Random seed.")
+
+run_type_g = ArgumentGroup(parser, "run_type", "running type options.")
+run_type_g.add_arg("use_cuda",                     bool,   True,   "If set, use GPU for training.")
+run_type_g.add_arg("shuffle",                      bool,   True,  "")
+run_type_g.add_arg("task_name",                    str,    None,
+                   "The name of task to perform fine-tuning, should be in {'xnli', 'mnli', 'cola', 'mrpc'}.")
+run_type_g.add_arg("do_train",                     bool,   True,  "Whether to perform training.")
+run_type_g.add_arg("do_test",                      bool,   False,  "Whether to perform evaluation on test data set.")
+run_type_g.add_arg("use_data_parallel", bool, False,  "The flag indicating whether to shuffle instances in each pass.")
+
+args = parser.parse_args()
+
+def create_data(batch):
+    """
+    convert data to variable
+    """
+    src_ids = to_variable(batch[0], "src_ids")
+    position_ids = to_variable(batch[1], "position_ids")
+    sentence_ids = to_variable(batch[2], "sentence_ids")
+    input_mask = to_variable(batch[3], "input_mask")
+    labels = to_variable(batch[4], "labels")
+    labels.stop_gradient = True
+    return src_ids, position_ids, sentence_ids, input_mask, labels
+
+if args.use_cuda:
+    place = fluid.CUDAPlace(fluid.dygraph.parallel.Env().dev_id)
+    dev_count = fluid.core.get_cuda_device_count()
+else:
+    place = fluid.CPUPlace()
+    dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+
+
+def train(args):
+    bert_config = BertConfig(args.bert_config_path)
+    bert_config.print_config()
+
+    if not (args.do_train or args.do_test):
+        raise ValueError("For args `do_train`, `do_test`, at "
+                        "least one of them must be True.")
+
+    trainer_count = fluid.dygraph.parallel.Env().nranks
+
+    task_name = args.task_name.lower()
+    processors = {
+        'xnli': reader.XnliProcessor,
+        'cola': reader.ColaProcessor,
+        'mrpc': reader.MrpcProcessor,
+        'mnli': reader.MnliProcessor,
+    }
+
+    processor = processors[task_name](data_dir=args.data_dir,
+                                      vocab_path=args.vocab_path,
+                                      max_seq_len=args.max_seq_len,
+                                      do_lower_case=args.do_lower_case,
+                                      in_tokens=args.in_tokens,
+                                      random_seed=args.random_seed)
+    num_labels = len(processor.get_labels())
+    shuffle_seed = 1 if trainer_count > 1 else None
+
+    train_data_generator = processor.data_generator(
+                                      batch_size=args.batch_size,
+                                      phase='train',
+                                      epoch=args.epoch,
+                                      dev_count=trainer_count,
+                                      shuffle=args.shuffle,
+                                      shuffle_seed=shuffle_seed)
+    num_train_examples = processor.get_num_examples(phase='train')
+    max_train_steps = args.epoch * num_train_examples // args.batch_size // trainer_count
+    warmup_steps = int(max_train_steps * args.warmup_proportion)
+
+    print("Device count: %d" % dev_count)
+    print("Trainer count: %d" % trainer_count)
+    print("Num train examples: %d" % num_train_examples)
+    print("Max train steps: %d" % max_train_steps)
+    print("Num warmup steps: %d" % warmup_steps)
+
+    with fluid.dygraph.guard(place):
+
+        if args.use_data_parallel:
+            strategy = fluid.dygraph.parallel.prepare_context()
+
+        cls_model = ClsModelLayer(
+                            args,
+                            bert_config,
+                            num_labels,
+                            is_training=True,
+                            return_pooled_out=True)
+
+        optimizer = Optimizer(
+                    warmup_steps=warmup_steps,
+                    num_train_steps=max_train_steps,
+                    learning_rate=args.learning_rate,
+                    model_cls=cls_model,
+                    weight_decay=args.weight_decay,
+                    scheduler=args.lr_scheduler,
+                    loss_scaling=args.loss_scaling,
+                    parameter_list=cls_model.parameters())
+
+        if args.init_pretraining_params:
+            print("Load pre-trained model from %s" % args.init_pretraining_params)
+            init_from_static_model(args.init_pretraining_params, cls_model, bert_config)
+
+        if args.use_data_parallel:
+            cls_model = fluid.dygraph.parallel.DataParallel(cls_model, strategy)
+            train_data_generator = fluid.contrib.reader.distributed_batch_reader(train_data_generator)
+
+        steps = 0
+        time_begin = time.time()
+
+        for batch in train_data_generator():
+            data_ids = create_data(batch)
+            loss, accuracy, num_seqs = cls_model(data_ids)
+
+            optimizer.optimization(loss, use_data_parallel = args.use_data_parallel, model = cls_model)
+            cls_model.clear_gradients()
+
+            if steps != 0 and steps % args.skip_steps == 0:
+                time_end = time.time()
+                used_time = time_end - time_begin
+                current_example, current_epoch = processor.get_train_progress()
+                localtime = time.asctime(time.localtime(time.time()))
+                print("%s, epoch: %s, steps: %s, dy_graph loss: %f, acc: %f, speed: %f steps/s" % (localtime, current_epoch, steps, loss.numpy(), accuracy.numpy(), args.skip_steps / used_time))
+                time_begin = time.time()
+
+            if steps != 0 and steps % args.save_steps == 0 and fluid.dygraph.parallel.Env().local_rank == 0:
+                save_path = os.path.join(args.checkpoints, "steps" + "_" + str(steps))
+                fluid.save_dygraph(
+                    cls_model.state_dict(),
+                    save_path)
+                fluid.save_dygraph(
+                    optimizer.optimizer.state_dict(),
+                    save_path)
+                print("Save model parameters and optimizer status at %s" % save_path)
+
+            steps += 1
+
+        if fluid.dygraph.parallel.Env().local_rank == 0:
+            save_path = os.path.join(args.checkpoints, "final")
+            fluid.save_dygraph(
+                cls_model.state_dict(),
+                save_path)
+            fluid.save_dygraph(
+                optimizer.optimizer.state_dict(),
+                save_path)
+            print("Save model parameters and optimizer status at %s" % save_path)
+        return cls_model
+
+def predict(args, cls_model = None):
+
+    bert_config = BertConfig(args.bert_config_path)
+    bert_config.print_config()
+
+    task_name = args.task_name.lower()
+    processors = {
+        'xnli': reader.XnliProcessor,
+        'cola': reader.ColaProcessor,
+        'mrpc': reader.MrpcProcessor,
+        'mnli': reader.MnliProcessor,
+    }
+
+    processor = processors[task_name](data_dir=args.data_dir,
+            vocab_path=args.vocab_path,
+            max_seq_len=args.max_seq_len,
+            do_lower_case=args.do_lower_case,
+            in_tokens=False)
+
+    test_data_generator = processor.data_generator(
+                                batch_size=args.batch_size,
+                                phase='dev',
+                                epoch=1,
+                                shuffle=False)
+
+    num_labels = len(processor.get_labels())
+
+    with fluid.dygraph.guard(place):
+        if cls_model is None:
+            cls_model = ClsModelLayer(
+                args,
+                bert_config,
+                num_labels,
+                is_training=False,
+                return_pooled_out=True)
+
+            #restore the model
+            save_path = os.path.join(args.checkpoints, "final")
+            print("Load params from %s" % save_path)
+            model_dict,_ = fluid.load_dygraph(save_path)
+            cls_model.load_dict(model_dict)
+
+        print('Do predicting ...... ')
+        cls_model.eval()
+
+        total_cost, total_acc, total_num_seqs = [], [], []
+
+        for batch in test_data_generator():
+            data_ids = create_data(batch)
+            np_loss, np_acc, np_num_seqs = cls_model(data_ids)
+
+            np_loss = np_loss.numpy()
+            np_acc = np_acc.numpy()
+            np_num_seqs = np_num_seqs.numpy()
+
+            total_cost.extend(np_loss * np_num_seqs)
+            total_acc.extend(np_acc * np_num_seqs)
+            total_num_seqs.extend(np_num_seqs)
+
+        print("[evaluation] average acc: %f" % (np.sum(total_acc) / np.sum(total_num_seqs)))
+
+
+if __name__ == '__main__':
+
+    print_arguments(args)
+    check_cuda(args.use_cuda)
+
+    if args.do_train:
+        cls_model = train(args)
+        if args.do_test:
+            predict(args, cls_model = cls_model)
+
+    elif args.do_test:
+        predict(args)

dygraph/bert/run_classifier_multi_gpu.sh

+#!/bin/bash
+
+BERT_BASE_PATH="./data/pretrained_models/uncased_L-12_H-768_A-12/"
+TASK_NAME='MNLI'
+DATA_PATH="./data/glue_data/MNLI/"
+CKPT_PATH="./data/saved_model/mnli_models"
+GPU_TO_USE="0,1,2,3"
+
+export CUDA_VISIBLE_DEVICES=$GPU_TO_USE
+
+# start fine-tuning
+python -m paddle.distributed.launch --selected_gpus=$GPU_TO_USE --log_dir ./cls_log run_classifier.py \
+    --task_name ${TASK_NAME} \
+    --use_cuda true \
+    --use_data_parallel true \
+    --do_train true \
+    --do_test true \
+    --batch_size 64 \
+    --in_tokens false \
+    --init_pretraining_params ${BERT_BASE_PATH}/dygraph_params/ \
+    --data_dir ${DATA_PATH} \
+    --vocab_path ${BERT_BASE_PATH}/vocab.txt \
+    --checkpoints ${CKPT_PATH} \
+    --save_steps 1000 \
+    --weight_decay  0.01 \
+    --warmup_proportion 0.1 \
+    --validation_steps 100 \
+    --epoch 3 \
+    --max_seq_len 128 \
+    --bert_config_path ${BERT_BASE_PATH}/bert_config.json \
+    --learning_rate 5e-5 \
+    --skip_steps 10 \
+    --shuffle true
+
+

dygraph/bert/run_classifier_predict.sh

+#!/bin/bash
+
+BERT_BASE_PATH="./data/pretrained_models/uncased_L-12_H-768_A-12/"
+TASK_NAME='MNLI'
+DATA_PATH="./data/glue_data/MNLI/"
+CKPT_PATH="./data/saved_model/mnli_models"
+
+export CUDA_VISIBLE_DEVICES=0
+
+# start testing
+python run_classifier.py\
+    --task_name ${TASK_NAME} \
+    --use_cuda true \
+    --do_train false \
+    --do_test true \
+    --batch_size 64 \
+    --in_tokens false \
+    --data_dir ${DATA_PATH} \
+    --vocab_path ${BERT_BASE_PATH}/vocab.txt \
+    --checkpoints ${CKPT_PATH} \
+    --save_steps 1000 \
+    --weight_decay  0.01 \
+    --warmup_proportion 0.1 \
+    --validation_steps 100 \
+    --epoch 3 \
+    --max_seq_len 128 \
+    --bert_config_path ${BERT_BASE_PATH}/bert_config.json \
+    --learning_rate 5e-5 \
+    --skip_steps 10 \
+    --shuffle false
+

dygraph/bert/run_classifier_single_gpu.sh

+#!/bin/bash
+
+BERT_BASE_PATH="./data/pretrained_models/uncased_L-12_H-768_A-12/"
+TASK_NAME='MNLI'
+DATA_PATH="./data/glue_data/MNLI/"
+CKPT_PATH="./data/saved_model/mnli_models"
+
+export CUDA_VISIBLE_DEVICES=0
+
+# start fine-tuning
+python run_classifier.py\
+    --task_name ${TASK_NAME} \
+    --use_cuda true \
+    --do_train true \
+    --do_test true \
+    --batch_size 64 \
+    --init_pretraining_params ${BERT_BASE_PATH}/dygraph_params/ \
+    --data_dir ${DATA_PATH} \
+    --vocab_path ${BERT_BASE_PATH}/vocab.txt \
+    --checkpoints ${CKPT_PATH} \
+    --save_steps 1000 \
+    --weight_decay  0.01 \
+    --warmup_proportion 0.1 \
+    --validation_steps 100 \
+    --epoch 3 \
+    --max_seq_len 128 \
+    --bert_config_path ${BERT_BASE_PATH}/bert_config.json \
+    --learning_rate 5e-5 \
+    --skip_steps 10 \
+    --shuffle true
+

dygraph/bert/tokenization.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#         http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization classes."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import unicodedata
+import six
+import io
+
+
+def convert_to_unicode(text):
+    """Converts `text` to Unicode (if it's not already), assuming utf-8 input."""
+    if six.PY3:
+        if isinstance(text, str):
+            return text
+        elif isinstance(text, bytes):
+            return text.decode("utf-8", "ignore")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    elif six.PY2:
+        if isinstance(text, str):
+            return text.decode("utf-8", "ignore")
+        elif isinstance(text, unicode):
+            return text
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    else:
+        raise ValueError("Not running on Python2 or Python 3?")
+
+
+def printable_text(text):
+    """Returns text encoded in a way suitable for print or `tf.logging`."""
+
+    # These functions want `str` for both Python2 and Python3, but in one case
+    # it's a Unicode string and in the other it's a byte string.
+    if six.PY3:
+        if isinstance(text, str):
+            return text
+        elif isinstance(text, bytes):
+            return text.decode("utf-8", "ignore")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    elif six.PY2:
+        if isinstance(text, str):
+            return text
+        elif isinstance(text, unicode):
+            return text.encode("utf-8")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    else:
+        raise ValueError("Not running on Python2 or Python 3?")
+
+
+def load_vocab(vocab_file):
+    """Loads a vocabulary file into a dictionary."""
+    vocab = collections.OrderedDict()
+    fin = io.open(vocab_file, encoding="utf8")
+    for num, line in enumerate(fin):
+        items = convert_to_unicode(line.strip()).split("\t")
+        if len(items) > 2:
+            break
+        token = items[0]
+        index = items[1] if len(items) == 2 else num
+        token = token.strip()
+        vocab[token] = int(index)
+    return vocab
+
+
+def convert_by_vocab(vocab, items):
+    """Converts a sequence of [tokens|ids] using the vocab."""
+    output = []
+    for item in items:
+        output.append(vocab[item])
+    return output
+
+
+def convert_tokens_to_ids(vocab, tokens):
+    return convert_by_vocab(vocab, tokens)
+
+
+def convert_ids_to_tokens(inv_vocab, ids):
+    return convert_by_vocab(inv_vocab, ids)
+
+
+def whitespace_tokenize(text):
+    """Runs basic whitespace cleaning and splitting on a peice of text."""
+    text = text.strip()
+    if not text:
+        return []
+    tokens = text.split()
+    return tokens
+
+
+class FullTokenizer(object):
+    """Runs end-to-end tokenziation."""
+
+    def __init__(self, vocab_file, do_lower_case=True):
+        self.vocab = load_vocab(vocab_file)
+        self.inv_vocab = {v: k for k, v in self.vocab.items()}
+        self.basic_tokenizer = BasicTokenizer(do_lower_case=do_lower_case)
+        self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab)
+
+    def tokenize(self, text):
+        split_tokens = []
+        for token in self.basic_tokenizer.tokenize(text):
+            for sub_token in self.wordpiece_tokenizer.tokenize(token):
+                split_tokens.append(sub_token)
+
+        return split_tokens
+
+    def convert_tokens_to_ids(self, tokens):
+        return convert_by_vocab(self.vocab, tokens)
+
+    def convert_ids_to_tokens(self, ids):
+        return convert_by_vocab(self.inv_vocab, ids)
+
+
+class CharTokenizer(object):
+    """Runs end-to-end tokenziation."""
+
+    def __init__(self, vocab_file, do_lower_case=True):
+        self.vocab = load_vocab(vocab_file)
+        self.inv_vocab = {v: k for k, v in self.vocab.items()}
+        self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab)
+
+    def tokenize(self, text):
+        split_tokens = []
+        for token in text.lower().split(" "):
+            for sub_token in self.wordpiece_tokenizer.tokenize(token):
+                split_tokens.append(sub_token)
+
+        return split_tokens
+
+    def convert_tokens_to_ids(self, tokens):
+        return convert_by_vocab(self.vocab, tokens)
+
+    def convert_ids_to_tokens(self, ids):
+        return convert_by_vocab(self.inv_vocab, ids)
+
+
+class BasicTokenizer(object):
+    """Runs basic tokenization (punctuation splitting, lower casing, etc.)."""
+
+    def __init__(self, do_lower_case=True):
+        """Constructs a BasicTokenizer.
+
+        Args:
+            do_lower_case: Whether to lower case the input.
+        """
+        self.do_lower_case = do_lower_case
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text."""
+        text = convert_to_unicode(text)
+        text = self._clean_text(text)
+
+        # This was added on November 1st, 2018 for the multilingual and Chinese
+        # models. This is also applied to the English models now, but it doesn't
+        # matter since the English models were not trained on any Chinese data
+        # and generally don't have any Chinese data in them (there are Chinese
+        # characters in the vocabulary because Wikipedia does have some Chinese
+        # words in the English Wikipedia.).
+        text = self._tokenize_chinese_chars(text)
+
+        orig_tokens = whitespace_tokenize(text)
+        split_tokens = []
+        for token in orig_tokens:
+            if self.do_lower_case:
+                token = token.lower()
+                token = self._run_strip_accents(token)
+            split_tokens.extend(self._run_split_on_punc(token))
+
+        output_tokens = whitespace_tokenize(" ".join(split_tokens))
+        return output_tokens
+
+    def _run_strip_accents(self, text):
+        """Strips accents from a piece of text."""
+        text = unicodedata.normalize("NFD", text)
+        output = []
+        for char in text:
+            cat = unicodedata.category(char)
+            if cat == "Mn":
+                continue
+            output.append(char)
+        return "".join(output)
+
+    def _run_split_on_punc(self, text):
+        """Splits punctuation on a piece of text."""
+        chars = list(text)
+        i = 0
+        start_new_word = True
+        output = []
+        while i < len(chars):
+            char = chars[i]
+            if _is_punctuation(char):
+                output.append([char])
+                start_new_word = True
+            else:
+                if start_new_word:
+                    output.append([])
+                start_new_word = False
+                output[-1].append(char)
+            i += 1
+
+        return ["".join(x) for x in output]
+
+    def _tokenize_chinese_chars(self, text):
+        """Adds whitespace around any CJK character."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if self._is_chinese_char(cp):
+                output.append(" ")
+                output.append(char)
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+    def _is_chinese_char(self, cp):
+        """Checks whether CP is the codepoint of a CJK character."""
+        # This defines a "chinese character" as anything in the CJK Unicode block:
+        #     https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
+        #
+        # Note that the CJK Unicode block is NOT all Japanese and Korean characters,
+        # despite its name. The modern Korean Hangul alphabet is a different block,
+        # as is Japanese Hiragana and Katakana. Those alphabets are used to write
+        # space-separated words, so they are not treated specially and handled
+        # like the all of the other languages.
+        if ((cp >= 0x4E00 and cp <= 0x9FFF) or  #
+            (cp >= 0x3400 and cp <= 0x4DBF) or  #
+            (cp >= 0x20000 and cp <= 0x2A6DF) or  #
+            (cp >= 0x2A700 and cp <= 0x2B73F) or  #
+            (cp >= 0x2B740 and cp <= 0x2B81F) or  #
+            (cp >= 0x2B820 and cp <= 0x2CEAF) or
+            (cp >= 0xF900 and cp <= 0xFAFF) or  #
+            (cp >= 0x2F800 and cp <= 0x2FA1F)):  #
+            return True
+
+        return False
+
+    def _clean_text(self, text):
+        """Performs invalid character removal and whitespace cleanup on text."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if cp == 0 or cp == 0xfffd or _is_control(char):
+                continue
+            if _is_whitespace(char):
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+
+class WordpieceTokenizer(object):
+    """Runs WordPiece tokenziation."""
+
+    def __init__(self, vocab, unk_token="[UNK]", max_input_chars_per_word=100):
+        self.vocab = vocab
+        self.unk_token = unk_token
+        self.max_input_chars_per_word = max_input_chars_per_word
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text into its word pieces.
+
+        This uses a greedy longest-match-first algorithm to perform tokenization
+        using the given vocabulary.
+
+        For example:
+            input = "unaffable"
+            output = ["un", "##aff", "##able"]
+
+        Args:
+            text: A single token or whitespace separated tokens. This should have
+                already been passed through `BasicTokenizer.
+
+        Returns:
+            A list of wordpiece tokens.
+        """
+
+        text = convert_to_unicode(text)
+
+        output_tokens = []
+        for token in whitespace_tokenize(text):
+            chars = list(token)
+            if len(chars) > self.max_input_chars_per_word:
+                output_tokens.append(self.unk_token)
+                continue
+
+            is_bad = False
+            start = 0
+            sub_tokens = []
+            while start < len(chars):
+                end = len(chars)
+                cur_substr = None
+                while start < end:
+                    substr = "".join(chars[start:end])
+                    if start > 0:
+                        substr = "##" + substr
+                    if substr in self.vocab:
+                        cur_substr = substr
+                        break
+                    end -= 1
+                if cur_substr is None:
+                    is_bad = True
+                    break
+                sub_tokens.append(cur_substr)
+                start = end
+
+            if is_bad:
+                output_tokens.append(self.unk_token)
+            else:
+                output_tokens.extend(sub_tokens)
+        return output_tokens
+
+
+def _is_whitespace(char):
+    """Checks whether `chars` is a whitespace character."""
+    # \t, \n, and \r are technically contorl characters but we treat them
+    # as whitespace since they are generally considered as such.
+    if char == " " or char == "\t" or char == "\n" or char == "\r":
+        return True
+    cat = unicodedata.category(char)
+    if cat == "Zs":
+        return True
+    return False
+
+
+def _is_control(char):
+    """Checks whether `chars` is a control character."""
+    # These are technically control characters but we count them as whitespace
+    # characters.
+    if char == "\t" or char == "\n" or char == "\r":
+        return False
+    cat = unicodedata.category(char)
+    if cat.startswith("C"):
+        return True
+    return False
+
+
+def _is_punctuation(char):
+    """Checks whether `chars` is a punctuation character."""
+    cp = ord(char)
+    # We treat all non-letter/number ASCII as punctuation.
+    # Characters such as "^", "$", and "`" are not in the Unicode
+    # Punctuation class but we treat them as punctuation anyways, for
+    # consistency.
+    if ((cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or
+        (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126)):
+        return True
+    cat = unicodedata.category(char)
+    if cat.startswith("P"):
+        return True
+    return False

dygraph/bert/utils/args.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Arguments for configuration."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import six
+import argparse
+
+import paddle.fluid as fluid
+
+
+def str2bool(v):
+    # because argparse does not support to parse "true, False" as python
+    # boolean directly
+    return v.lower() in ("true", "t", "1")
+
+
+class ArgumentGroup(object):
+    def __init__(self, parser, title, des):
+        self._group = parser.add_argument_group(title=title, description=des)
+
+    def add_arg(self, name, type, default, help, **kwargs):
+        type = str2bool if type == bool else type
+        self._group.add_argument(
+            "--" + name,
+            default=default,
+            type=type,
+            help=help + ' Default: %(default)s.',
+            **kwargs)
+
+
+def print_arguments(args):
+    print('-----------  Configuration Arguments -----------')
+    for arg, value in sorted(six.iteritems(vars(args))):
+        print('%s: %s' % (arg, value))
+    print('------------------------------------------------')
+
+def check_cuda(use_cuda, err = \
+    "\nYou can not set use_cuda = True in the model because you are using paddlepaddle-cpu.\n \
+    Please: 1. Install paddlepaddle-gpu to run your models on GPU or 2. Set use_cuda = False to run models on CPU.\n"
+                                                                                                                     ):
+    try:
+        if use_cuda == True and fluid.is_compiled_with_cuda() == False:
+            print(err)
+            sys.exit(1)
+    except Exception as e:
+        pass

dygraph/bert/utils/cards.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+
+
+def get_cards():
+    """
+    get gpu cards number
+    """
+    num = 0
+    cards = os.environ.get('CUDA_VISIBLE_DEVICES', '')
+    if cards != '':
+        num = len(cards.split(","))
+    return num

dygraph/bert/utils/convert_static_to_dygraph.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import shutil
+import sys
+import os
+
+
+def usage():
+    """
+    usage information
+    """
+    print
+    print("please use command: ")
+    print(
+        "python convert_static_to_dygraph.py input_params_dir output_params_dir")
+    print
+
+
+def convert_static_to_dygraph(static_model_path, dygraph_model_path):
+    """
+    convert paddle static bert model to dygraph model 
+    """
+
+    def mkdir(path):
+        if not os.path.isdir(path):
+            if os.path.split(path)[0]:
+                mkdir(os.path.split(path)[0])
+        else:
+            return
+        os.mkdir(path)
+
+    if os.path.exists(dygraph_model_path):
+        shutil.rmtree(dygraph_model_path)
+    mkdir(dygraph_model_path)
+
+    if not os.path.exists(static_model_path):
+        print("paddle static model path doesn't exist.....")
+        return -1
+
+    file_list = []
+    for root, dirs, files in os.walk(static_model_path):
+        file_list.extend(files)
+
+    os.makedirs(os.path.join(dygraph_model_path, "PretrainModelLayer_0"))
+    os.makedirs(
+        os.path.join(dygraph_model_path,
+                     "PretrainModelLayer_0/BertModelLayer_0"))
+    os.makedirs(
+        os.path.join(dygraph_model_path,
+                     "PretrainModelLayer_0/PrePostProcessLayer_0"))
+    os.makedirs(
+        os.path.join(
+            dygraph_model_path,
+            "PretrainModelLayer_0/BertModelLayer_0/PrePostProcessLayer_0"))
+
+    #os.chdir(static_model_path)
+    #convert embedding file
+    embedding_type = ["word", "pos", "sent"]
+    for i in range(3):
+        src_name = embedding_type[i] + "_embedding"
+        trg_name = "Embedding_" + str(i) + "." + src_name
+        shutil.copyfile(
+            os.path.join(static_model_path, src_name),
+            os.path.join(dygraph_model_path,
+                         "PretrainModelLayer_0/BertModelLayer_0/" + trg_name))
+
+    #convert pre_encoder file
+    shutil.copyfile(
+        os.path.join(static_model_path, "pre_encoder_layer_norm_scale"),
+        os.path.join(
+            dygraph_model_path,
+            "PretrainModelLayer_0/BertModelLayer_0/PrePostProcessLayer_0/LayerNorm_0._layer_norm_scale"
+        ))
+    shutil.copyfile(
+        os.path.join(static_model_path, "pre_encoder_layer_norm_bias"),
+        os.path.join(
+            dygraph_model_path,
+            "PretrainModelLayer_0/BertModelLayer_0/PrePostProcessLayer_0/LayerNorm_0._layer_norm_bias"
+        ))
+
+    #convert mask lm params file
+    shutil.copyfile(
+        os.path.join(static_model_path, "mask_lm_out_fc.b_0"),
+        os.path.join(dygraph_model_path,
+                     "PretrainModelLayer_0/Layer_0.mask_lm_out_fc.b_0"))
+    shutil.copyfile(
+        os.path.join(static_model_path, "mask_lm_trans_fc.b_0"),
+        os.path.join(dygraph_model_path,
+                     "PretrainModelLayer_0/FC_0.mask_lm_trans_fc.b_0"))
+    shutil.copyfile(
+        os.path.join(static_model_path, "mask_lm_trans_fc.w_0"),
+        os.path.join(dygraph_model_path,
+                     "PretrainModelLayer_0/FC_0.mask_lm_trans_fc.w_0"))
+    shutil.copyfile(
+        os.path.join(static_model_path, "mask_lm_trans_layer_norm_bias"),
+        os.path.join(
+            dygraph_model_path,
+            "PretrainModelLayer_0/PrePostProcessLayer_0/LayerNorm_0._layer_norm_bias"
+        ))
+    shutil.copyfile(
+        os.path.join(static_model_path, "mask_lm_trans_layer_norm_scale"),
+        os.path.join(
+            dygraph_model_path,
+            "PretrainModelLayer_0/PrePostProcessLayer_0/LayerNorm_0._layer_norm_scale"
+        ))
+    shutil.copyfile(
+        os.path.join(static_model_path, "next_sent_fc.b_0"),
+        os.path.join(dygraph_model_path,
+                     "PretrainModelLayer_0/FC_1.next_sent_fc.b_0"))
+    shutil.copyfile(
+        os.path.join(static_model_path, "next_sent_fc.w_0"),
+        os.path.join(dygraph_model_path,
+                     "PretrainModelLayer_0/FC_1.next_sent_fc.w_0"))
+    shutil.copyfile(
+        os.path.join(static_model_path, "pooled_fc.b_0"),
+        os.path.join(
+            dygraph_model_path,
+            "PretrainModelLayer_0/BertModelLayer_0/FC_0.pooled_fc.b_0"))
+    shutil.copyfile(
+        os.path.join(static_model_path, "pooled_fc.w_0"),
+        os.path.join(
+            dygraph_model_path,
+            "PretrainModelLayer_0/BertModelLayer_0/FC_0.pooled_fc.w_0"))
+
+    encoder_num = 0
+    for f in file_list:
+        if not f.startswith("encoder_layer"):
+            continue
+        layer_num = f.split('_')[2]
+        if int(layer_num) > encoder_num:
+            encoder_num = int(layer_num)
+
+    encoder_num += 1
+    for i in range(encoder_num):
+        encoder_dir = "EncoderSubLayer_" + str(i)
+        os.makedirs(
+            os.path.join(dygraph_model_path,
+                         "PretrainModelLayer_0/BertModelLayer_0/" +
+                         "EncoderLayer_0/", encoder_dir))
+        os.makedirs(
+            os.path.join(dygraph_model_path,
+                         "PretrainModelLayer_0/BertModelLayer_0/" +
+                         "EncoderLayer_0/", encoder_dir +
+                         "/PositionwiseFeedForwardLayer_0"))
+        os.makedirs(
+            os.path.join(
+                dygraph_model_path, "PretrainModelLayer_0/BertModelLayer_0/" +
+                "EncoderLayer_0/", encoder_dir + "/MultiHeadAttentionLayer_0"))
+        os.makedirs(
+            os.path.join(
+                dygraph_model_path, "PretrainModelLayer_0/BertModelLayer_0/" +
+                "EncoderLayer_0/", encoder_dir + "/PrePostProcessLayer_1"))
+        os.makedirs(
+            os.path.join(
+                dygraph_model_path, "PretrainModelLayer_0/BertModelLayer_0/" +
+                "EncoderLayer_0/", encoder_dir + "/PrePostProcessLayer_3"))
+
+    encoder_map_dict = {
+        "ffn_fc_0.b_0": ("PositionwiseFeedForwardLayer_0", "FC_0.ffn_fc_0.b_0"),
+        "ffn_fc_0.w_0": ("PositionwiseFeedForwardLayer_0", "FC_0.ffn_fc_0.w_0"),
+        "ffn_fc_1.b_0": ("PositionwiseFeedForwardLayer_0", "FC_1.ffn_fc_1.b_0"),
+        "ffn_fc_1.w_0": ("PositionwiseFeedForwardLayer_0", "FC_1.ffn_fc_1.w_0"),
+        "multi_head_att_key_fc.b_0":
+        ("MultiHeadAttentionLayer_0", "FC_1.key_fc.b_0"),
+        "multi_head_att_key_fc.w_0":
+        ("MultiHeadAttentionLayer_0", "FC_1.key_fc.w_0"),
+        "multi_head_att_output_fc.b_0":
+        ("MultiHeadAttentionLayer_0", "FC_3.output_fc.b_0"),
+        "multi_head_att_output_fc.w_0":
+        ("MultiHeadAttentionLayer_0", "FC_3.output_fc.w_0"),
+        "multi_head_att_query_fc.b_0":
+        ("MultiHeadAttentionLayer_0", "FC_0.query_fc.b_0"),
+        "multi_head_att_query_fc.w_0":
+        ("MultiHeadAttentionLayer_0", "FC_0.query_fc.w_0"),
+        "multi_head_att_value_fc.b_0":
+        ("MultiHeadAttentionLayer_0", "FC_2.value_fc.b_0"),
+        "multi_head_att_value_fc.w_0":
+        ("MultiHeadAttentionLayer_0", "FC_2.value_fc.w_0"),
+        "post_att_layer_norm_bias":
+        ("PrePostProcessLayer_1", "LayerNorm_0.post_att_layer_norm_bias"),
+        "post_att_layer_norm_scale":
+        ("PrePostProcessLayer_1", "LayerNorm_0.post_att_layer_norm_scale"),
+        "post_ffn_layer_norm_bias":
+        ("PrePostProcessLayer_3", "LayerNorm_0.post_ffn_layer_norm_bias"),
+        "post_ffn_layer_norm_scale":
+        ("PrePostProcessLayer_3", "LayerNorm_0.post_ffn_layer_norm_scale")
+    }
+
+    for f in file_list:
+        if not f.startswith("encoder_layer"):
+            continue
+        layer_num = f.split('_')[2]
+        suffix_name = "_".join(f.split('_')[3:])
+        in_dir = encoder_map_dict[suffix_name][0]
+        rename = encoder_map_dict[suffix_name][1]
+        encoder_layer = "EncoderSubLayer_" + layer_num
+        shutil.copyfile(
+            os.path.join(static_model_path, f),
+            os.path.join(dygraph_model_path,
+                         "PretrainModelLayer_0/BertModelLayer_0/EncoderLayer_0/"
+                         + encoder_layer + "/" + in_dir + "/" + rename))
+
+
+if __name__ == "__main__":
+
+    if len(sys.argv) < 3:
+        usage()
+        exit(1)
+    static_model_path = sys.argv[1]
+    dygraph_model_path = sys.argv[2]
+    convert_static_to_dygraph(static_model_path, dygraph_model_path)

dygraph/bert/utils/fp16.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+import paddle
+import paddle.fluid as fluid
+
+
+def cast_fp16_to_fp32(i, o, prog):
+    prog.global_block().append_op(
+        type="cast",
+        inputs={"X": i},
+        outputs={"Out": o},
+        attrs={
+            "in_dtype": fluid.core.VarDesc.VarType.FP16,
+            "out_dtype": fluid.core.VarDesc.VarType.FP32
+        })
+
+
+def cast_fp32_to_fp16(i, o, prog):
+    prog.global_block().append_op(
+        type="cast",
+        inputs={"X": i},
+        outputs={"Out": o},
+        attrs={
+            "in_dtype": fluid.core.VarDesc.VarType.FP32,
+            "out_dtype": fluid.core.VarDesc.VarType.FP16
+        })
+
+
+def copy_to_master_param(p, block):
+    v = block.vars.get(p.name, None)
+    if v is None:
+        raise ValueError("no param name %s found!" % p.name)
+    new_p = fluid.framework.Parameter(
+        block=block,
+        shape=v.shape,
+        dtype=fluid.core.VarDesc.VarType.FP32,
+        type=v.type,
+        lod_level=v.lod_level,
+        stop_gradient=p.stop_gradient,
+        trainable=p.trainable,
+        optimize_attr=p.optimize_attr,
+        regularizer=p.regularizer,
+        gradient_clip_attr=p.gradient_clip_attr,
+        error_clip=p.error_clip,
+        name=v.name + ".master")
+    return new_p
+
+
+def create_master_params_grads(params_grads, main_prog, startup_prog,
+                               loss_scaling):
+    master_params_grads = []
+    tmp_role = main_prog._current_role
+    OpRole = fluid.core.op_proto_and_checker_maker.OpRole
+    main_prog._current_role = OpRole.Backward
+    for p, g in params_grads:
+        # create master parameters
+        master_param = copy_to_master_param(p, main_prog.global_block())
+        startup_master_param = startup_prog.global_block()._clone_variable(
+            master_param)
+        startup_p = startup_prog.global_block().var(p.name)
+        cast_fp16_to_fp32(startup_p, startup_master_param, startup_prog)
+        # cast fp16 gradients to fp32 before apply gradients
+        if g.name.find("layer_norm") > -1:
+            if loss_scaling > 1:
+                scaled_g = g / float(loss_scaling)
+            else:
+                scaled_g = g
+            master_params_grads.append([p, scaled_g])
+            continue
+        master_grad = fluid.layers.cast(g, "float32")
+        if loss_scaling > 1:
+            master_grad = master_grad / float(loss_scaling)
+        master_params_grads.append([master_param, master_grad])
+    main_prog._current_role = tmp_role
+    return master_params_grads
+
+
+def master_param_to_train_param(master_params_grads, params_grads, main_prog):
+    for idx, m_p_g in enumerate(master_params_grads):
+        train_p, _ = params_grads[idx]
+        if train_p.name.find("layer_norm") > -1:
+            continue
+        with main_prog._optimized_guard([m_p_g[0], m_p_g[1]]):
+            cast_fp32_to_fp16(m_p_g[0], train_p, main_prog)

dygraph/bert/utils/init.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import os
+import six
+import ast
+import copy
+
+import numpy as np
+import paddle.fluid as fluid
+
+
+def cast_fp32_to_fp16(exe, main_program):
+    print("Cast parameters to float16 data format.")
+    for param in main_program.global_block().all_parameters():
+        if not param.name.endswith(".master"):
+            param_t = fluid.global_scope().find_var(param.name).get_tensor()
+            data = np.array(param_t)
+            if param.name.find("layer_norm") == -1:
+                param_t.set(np.float16(data).view(np.uint16), exe.place)
+            master_param_var = fluid.global_scope().find_var(param.name +
+                                                             ".master")
+            if master_param_var is not None:
+                master_param_var.get_tensor().set(data, exe.place)
+
+
+def init_checkpoint(exe, init_checkpoint_path, main_program, use_fp16=False):
+    assert os.path.exists(
+        init_checkpoint_path), "[%s] cann't be found." % init_checkpoint_path
+
+    def existed_persitables(var):
+        if not fluid.io.is_persistable(var):
+            return False
+        return os.path.exists(os.path.join(init_checkpoint_path, var.name))
+
+    fluid.io.load_vars(
+        exe,
+        init_checkpoint_path,
+        main_program=main_program,
+        predicate=existed_persitables)
+    print("Load model from {}".format(init_checkpoint_path))
+
+    if use_fp16:
+        cast_fp32_to_fp16(exe, main_program)
+
+
+def init_pretraining_params(exe,
+                            pretraining_params_path,
+                            main_program,
+                            use_fp16=False):
+    assert os.path.exists(pretraining_params_path
+                          ), "[%s] cann't be found." % pretraining_params_path
+
+    def existed_params(var):
+        if not isinstance(var, fluid.framework.Parameter):
+            return False
+        return os.path.exists(os.path.join(pretraining_params_path, var.name))
+
+    fluid.io.load_vars(
+        exe,
+        pretraining_params_path,
+        main_program=main_program,
+        predicate=existed_params)
+    print("Load pretraining parameters from {}.".format(
+        pretraining_params_path))
+
+    if use_fp16:
+        cast_fp32_to_fp16(exe, main_program)
+
+
+def init_from_static_model(dir_path, cls_model, bert_config):
+    def load_numpy_weight(file_name):
+        res = np.load(os.path.join(dir_path, file_name), allow_pickle=True)
+        assert res is not None
+        return res
+
+    # load word embedding
+    _param = load_numpy_weight("word_embedding")
+    cls_model.bert_layer._src_emb.set_dict({"weight": _param})
+    print("INIT word embedding")
+
+    _param = load_numpy_weight("pos_embedding")
+    cls_model.bert_layer._pos_emb.set_dict({"weight": _param})
+    print("INIT pos embedding")
+
+    _param = load_numpy_weight("sent_embedding")
+    cls_model.bert_layer._sent_emb.set_dict({"weight": _param})
+    print("INIT sent embedding")
+
+    _param0 = load_numpy_weight("pooled_fc.w_0")
+    _param1 = load_numpy_weight("pooled_fc.b_0")
+    cls_model.bert_layer.pooled_fc.set_dict({
+        "weight": _param0,
+        "bias": _param1
+    })
+    print("INIT pooled_fc")
+
+    _param0 = load_numpy_weight("pre_encoder_layer_norm_scale")
+    _param1 = load_numpy_weight("pre_encoder_layer_norm_bias")
+    cls_model.bert_layer.pre_process_layer._sub_layers["layer_norm_0"].set_dict(
+        {
+            "weight": _param0,
+            "bias": _param1
+        })
+    print("INIT pre_encoder layer norm")
+
+    for _i in range(bert_config["num_hidden_layers"]):
+        _param_weight = "encoder_layer_%d_multi_head_att_query_fc.w_0" % _i
+        _param_bias = "encoder_layer_%d_multi_head_att_query_fc.b_0" % _i
+
+        _param_weight = load_numpy_weight(_param_weight)
+        _param_bias = load_numpy_weight(_param_bias)
+
+        cls_model.bert_layer._encoder._sub_layers[
+            "esl_%d" % _i]._multihead_attention_layer._q_fc.set_dict({
+                "weight": _param_weight,
+                "bias": _param_bias
+            })
+        print("INIT multi_head_att_query_fc %d" % _i)
+
+        _param_weight = "encoder_layer_%d_multi_head_att_key_fc.w_0" % _i
+        _param_bias = "encoder_layer_%d_multi_head_att_key_fc.b_0" % _i
+
+        _param_weight = load_numpy_weight(_param_weight)
+        _param_bias = load_numpy_weight(_param_bias)
+
+        cls_model.bert_layer._encoder._sub_layers[
+            "esl_%d" % _i]._multihead_attention_layer._k_fc.set_dict({
+                "weight": _param_weight,
+                "bias": _param_bias
+            })
+        print("INIT multi_head_att_key_fc %d" % _i)
+
+        _param_weight = "encoder_layer_%d_multi_head_att_value_fc.w_0" % _i
+        _param_bias = "encoder_layer_%d_multi_head_att_value_fc.b_0" % _i
+
+        _param_weight = load_numpy_weight(_param_weight)
+        _param_bias = load_numpy_weight(_param_bias)
+
+        cls_model.bert_layer._encoder._sub_layers[
+            "esl_%d" % _i]._multihead_attention_layer._v_fc.set_dict({
+                "weight": _param_weight,
+                "bias": _param_bias
+            })
+        print("INIT multi_head_att_value_fc %d" % _i)
+
+        # init output fc
+        _param_weight = "encoder_layer_%d_multi_head_att_output_fc.w_0" % _i
+        _param_bias = "encoder_layer_%d_multi_head_att_output_fc.b_0" % _i
+
+        _param_weight = load_numpy_weight(_param_weight)
+        _param_bias = load_numpy_weight(_param_bias)
+
+        cls_model.bert_layer._encoder._sub_layers[
+            "esl_%d" % _i]._multihead_attention_layer._proj_fc.set_dict({
+                "weight": _param_weight,
+                "bias": _param_bias
+            })
+        print("INIT multi_head_att_output_fc %d" % _i)
+
+        # init layer_norm 1
+        _param_weight = "encoder_layer_%d_post_att_layer_norm_scale" % _i
+        _param_bias = "encoder_layer_%d_post_att_layer_norm_bias" % _i
+
+        _param_weight = load_numpy_weight(_param_weight)
+        _param_bias = load_numpy_weight(_param_bias)
+
+        cls_model.bert_layer._encoder._sub_layers[
+            "esl_%d" % _i]._postprocess_layer.layer_norm_0.set_dict({
+                "weight": _param_weight,
+                "bias": _param_bias
+            })
+        print("INIT layer norm in attention at %d layer" % _i)
+
+        # init layer_norm 2
+        _param_weight = "encoder_layer_%d_post_ffn_layer_norm_scale" % _i
+        _param_bias = "encoder_layer_%d_post_ffn_layer_norm_bias" % _i
+
+        _param_weight = load_numpy_weight(_param_weight)
+        _param_bias = load_numpy_weight(_param_bias)
+
+        cls_model.bert_layer._encoder._sub_layers[
+            "esl_%d" % _i]._postprocess_layer2.layer_norm_0.set_dict({
+                "weight": _param_weight,
+                "bias": _param_bias
+            })
+        print("INIT layer norm in FFN at %d layer" % _i)
+
+        # init FFN 1
+        _param_weight = "encoder_layer_%d_ffn_fc_0.w_0" % _i
+        _param_bias = "encoder_layer_%d_ffn_fc_0.b_0" % _i
+
+        _param_weight = load_numpy_weight(_param_weight)
+        _param_bias = load_numpy_weight(_param_bias)
+
+        cls_model.bert_layer._encoder._sub_layers[
+            "esl_%d" % _i]._positionwise_feed_forward._i2h.set_dict({
+                "weight": _param_weight,
+                "bias": _param_bias
+            })
+        print("INIT FFN-1 at %d layer" % _i)
+
+        # init FFN 2
+        _param_weight = "encoder_layer_%d_ffn_fc_1.w_0" % _i
+        _param_bias = "encoder_layer_%d_ffn_fc_1.b_0" % _i
+
+        _param_weight = load_numpy_weight(_param_weight)
+        _param_bias = load_numpy_weight(_param_bias)
+
+        cls_model.bert_layer._encoder._sub_layers[
+            "esl_%d" % _i]._positionwise_feed_forward._h2o.set_dict({
+                "weight": _param_weight,
+                "bias": _param_bias
+            })
+        print("INIT FFN-2 at %d layer" % _i)
+
+    # init cls fc
+    #_param_weight = "cls_out_w"
+    #_param_bias = "cls_out_b"
+
+    #_param_weight = load_numpy_weight(_param_weight)
+    #_param_bias = load_numpy_weight(_param_bias)
+
+    #cls_model.cls_fc.set_dict({"weight":_param_weight, "bias":_param_bias})
+    #print("INIT CLS FC layer")
+    return True