Dygraph fix3 (#457)

* update readme

* update demo

* + 160G model

* qa model bugfix: models inherits docstrings

* Update README.zh.md

* Update README.en.md

* Update README.zh.md

* reorganize binaries

* Update README.zh.md

* Update README.en.md

* Update README.zh.md

* Update README.en.md

README.en.md

 English|[简体中文](./README.zh.md)
 
-![./.metas/ERNIE_milestone.png](./.metas/ERNIE_milestone.png)
+![./.metas/ERNIE_milestone.png](./.metas/ERNIE_milestone_en.png)
 
 
 **Remind： This repo has been refactored, for paper re-production or backward compatibility; plase checkout to [repro branch](https://github.com/PaddlePaddle/ERNIE/tree/repro)**
@@ -89,23 +89,23 @@ pip install paddle-ernie
 or 
 
 ```shell
-git clone -b dygraph https://github.com/PaddlePaddle/ERNIE.git --single-branch
+git clone https://github.com/PaddlePaddle/ERNIE.git --depth 1
 cd ERNIE
 pip install -r requirements.txt
-pip setup.py -e .
-
+pip install -e .
 ```
 
 ##### 3. download pretrained models (optional)
 
-| Model                                              | Description                                                  |
-| :------------------------------------------------- | :----------------------------------------------------------- |
-| [ERNIE 1.0 Base for Chinese](https://ernie-github.cdn.bcebos.com/model-ernie1.0.1.tar.gz)           | L12H768A12  |
-| [ERNIE Tiny](https://ernie-github.cdn.bcebos.com/model-ernie_tiny.1.tar.gz)                         | L3H1024A16  |
-| [ERNIE 2.0 Base for English](https://ernie-github.cdn.bcebos.com/model-ernie2.0-en.1.tar.gz)        | L12H768A12  |
-| [ERNIE 2.0 Large for English](https://ernie-github.cdn.bcebos.com/model-ernie2.0-large-en.1.tar.gz) | L24H1024A16 |
-| [ERNIE Gen base for English](https://ernie-github.cdn.bcebos.com/model-ernie-gen-base-en.1.tar.gz)  | L12H768A12  |
-| [ERNIE Gen Large for English](https://ernie-github.cdn.bcebos.com/model-ernie-gen-large-en.1.tar.gz)| L24H1024A16 |
+| Model                                              | Description                                                  |abbreviation|
+| :------------------------------------------------- | :----------------------------------------------------------- |:-----------|
+| [ERNIE 1.0 Base for Chinese](https://ernie-github.cdn.bcebos.com/model-ernie1.0.1.tar.gz)           | L12H768A12  |ernie-1.0|
+| [ERNIE Tiny](https://ernie-github.cdn.bcebos.com/model-ernie_tiny.1.tar.gz)                         | L3H1024A16  |ernie-tiny|
+| [ERNIE 2.0 Base for English](https://ernie-github.cdn.bcebos.com/model-ernie2.0-en.1.tar.gz)        | L12H768A12  |ernie-2.0-en|
+| [ERNIE 2.0 Large for English](https://ernie-github.cdn.bcebos.com/model-ernie2.0-large-en.1.tar.gz) | L24H1024A16 |ernie-2.0-large-en|
+| [ERNIE Gen base for English](https://ernie-github.cdn.bcebos.com/model-ernie-gen-base-en.1.tar.gz)  | L12H768A12  |ernie-gen-base-en|
+| [ERNIE Gen Large for English](https://ernie-github.cdn.bcebos.com/model-ernie-gen-large-en.1.tar.gz)| L24H1024A16 | ernie-gen-large-en |
+| [ERNIE Gen Large 160G for English](https://ernie-github.cdn.bcebos.com/model-ernie-gen-large-en.1.tar.gz)| Layer:24, Hidden:1024, Heads:16 + 160G pretrain corpus | ernie-gen-large-160g-en |
 
 ##### 4. download datasets
  
@@ -143,26 +143,31 @@ see [demo](https://ernie-github.cdn.bcebos.com/data-mnli-m.tar.gz) data for MNLI
 - try eager execution with `dygraph model` :
 
 ```script
-python3 ./demo/finetune_classifier_dygraph.py \
-    --from_pretrained ernie_1.0 \
-    --data_dir ./data/xnli 
+python3 ./ernie_d/demo/finetune_classifier_dygraph.py \
+       --from_pretrained ernie-1.0 \
+       --data_dir ./data/xnli  
 ```
 
 - Distributed finetune
 
 `paddle.distributed.launch` is a process manager, we use it to launch python processes on each avalible GPU devices:
 
-when in distributed training, `max_steps` is used as stopping criteria rather than `epoch` to prevent dead block.
-also notice than we shard the train data according to device id to prevent over fitting.
+When in distributed training, `max_steps` is used as stopping criteria rather than `epoch` to prevent dead block.
+You could calculate `max_steps` with `EPOCH * NUM_TRAIN_EXAMPLES / TOTAL_BATCH`.
+Also notice than we shard the train data according to device id to prevent over fitting.
 
 demo: 
+(make sure you have more than 2 GPUs, 
+online model download can not work in `paddle.distributed.launch`, 
+you need to run single card finetuning first to get pretrained model, or donwload and extract one manualy from [here](#section-pretrained-models)): 
+
 
 ```script
 python3 -m paddle.distributed.launch \
 ./demo/finetune_classifier_dygraph_distributed.py \
     --data_dir data/mnli \
     --max_steps 10000 \
-    --from_pretrained ernie2.0-en
+    --from_pretrained ernie-2.0-en
 ```
 
 

README.zh.md

 [English](./README.en.md)|简体中文
 
-![./.metas/ERNIE_milestone.png](./.metas/ERNIE_milestone.png)
+![./.metas/ERNIE_milestone.png](./.metas/ERNIE_milestone_zh.png)
 
 ERNIE是百度开创性提出的基于知识增强的持续学习语义理解框架，该框架将大数据预训练与多源丰富知识相结合，通过持续学习技术，不断吸收海量文本数据中词汇、结构、语义等方面的知识，实现模型效果不断进化。ERNIE在情感分析、文本匹配、自然语言推理、词法分析、阅读理解、智能问答等16个公开数据集上全面显著超越世界领先技术，在国际权威的通用语言理解评估基准GLUE上，得分首次突破90分，获得全球第一。在今年3月落下帷幕的全球最大语义评测SemEval 2020上，ERNIE摘得5项世界冠军， 该技术也被全球顶级科技商业杂志《麻省理工科技评论》官方网站报道，相关创新成果也被国际顶级学术会议AAAI、IJCAI收录。ERNIE在工业界得到了大规模应用，如搜索引擎、新闻推荐、广告系统、语音交互、智能客服等。
 
@@ -87,24 +87,25 @@ pip install paddle-ernie
 或者
 
 ```shell
-git clone -b dygraph https://github.com/PaddlePaddle/ERNIE.git --single-branch
+git clone https://github.com/PaddlePaddle/ERNIE.git --depth 1
 cd ERNIE
 pip install -r requirements.txt
-pip setup.py -e .
+pip install -e .
 
 ```
 
-##### 3. 下载预训练模型（可选）
+##### 3. 下载预训练模型（可选）<a name="section-pretrained-models"></a>
 
 
-| Model                                              | Description                                                  |
-| :------------------------------------------------- | :----------------------------------------------------------- |
-| [ERNIE 1.0 Base 中文](https://ernie-github.cdn.bcebos.com/model-ernie1.0.1.tar.gz)           | L12H768A12  |
-| [ERNIE Tiny](https://ernie-github.cdn.bcebos.com/model-ernie_tiny.1.tar.gz)                 | L3H1024A16      |
-| [ERNIE 2.0 Base 英文](https://ernie-github.cdn.bcebos.com/model-ernie2.0-en.1.tar.gz)        | base: L12H768A12  |
-| [ERNIE 2.0 Large 英文](https://ernie-github.cdn.bcebos.com/model-ernie2.0-large-en.1.tar.gz) | large: L24H1024A16|
-| [ERNIE Gen base 英文](https://ernie-github.cdn.bcebos.com/model-ernie-gen-base-en.1.tar.gz)  | L12H768A12  |
-| [ERNIE Gen Large 英文](https://ernie-github.cdn.bcebos.com/model-ernie-gen-large-en.1.tar.gz)| L24H1024A16 |
+| Model                                              | 细节参数                                                                  |下载简写|
+| :------------------------------------------------- |:------------------------------------------------------------------------- |:-------|
+| [ERNIE 1.0 Base 中文](https://ernie-github.cdn.bcebos.com/model-ernie1.0.1.tar.gz)           | Layer:12, Hidden:768, Heads:12  |ernie-1.0|
+| [ERNIE Tiny](https://ernie-github.cdn.bcebos.com/model-ernie_tiny.1.tar.gz)                  | Layer:3, Hdden:1024, Heads:16   |ernie-tiny|
+| [ERNIE 2.0 Base 英文](https://ernie-github.cdn.bcebos.com/model-ernie2.0-en.1.tar.gz)        | Layer:12, Hidden:768, Heads:12  |ernie-2.0-en|
+| [ERNIE 2.0 Large 英文](https://ernie-github.cdn.bcebos.com/model-ernie2.0-large-en.1.tar.gz) | Layer:24, Hidden:1024, Heads16  |ernie-2.0-large-en|
+| [ERNIE Gen Base 英文](https://ernie-github.cdn.bcebos.com/model-ernie-gen-base-en.1.tar.gz)  | Layer:12, Hidden:768, Heads:12  |ernie-gen-base-en|
+| [ERNIE Gen Large 英文](https://ernie-github.cdn.bcebos.com/model-ernie-gen-large-en.1.tar.gz)| Layer:24, Hidden:1024, Heads:16 |ernie-gen-large-en|
+| [ERNIE Gen Large 160G英文](https://ernie-github.cdn.bcebos.com/model-ernie-gen-large-en.1.tar.gz)| Layer:24, Hidden:1024, Heads:16 + 额外160G 预训练语料 | ernie-gen-large-160g-en |
 
 ##### 4. 下载数据集
 
@@ -144,9 +145,9 @@ data/xnli
 - 使用 `动态图` 模型进行finetune:
 
 ```script
-python3 ./demo/finetune_classifier_dygraph.py \
-    --from_pretrained ernie_1.0 \
-    --data_dir ./data/xnli
+python3 ./ernie_d/demo/finetune_classifier_dygraph.py \
+       --from_pretrained ernie-1.0 \
+       --data_dir ./data/xnli  
 ```
 
 - 分布式 finetune
@@ -154,9 +155,11 @@ python3 ./demo/finetune_classifier_dygraph.py \
 `paddle.distributed.launch` 是一个进程管理器，我们采用它在每一张GPU上启动一个python进程，并配置相应的环境变量以进行分布式训练:
 
 当采用分布式训练时，我们采用`max_steps`做为终止条件而非`epoch`, 这样处理是为了避免进程间死锁。
+你可以通过`EPOCH * NUM_TRAIN_EXAMPLES / TOTAL_BATCH`的方式计算出所需执行的`max_steps`.
 另外值得注意的是训练集需要在不同的进程间进行切分；以避免所有进程训练同一份数据造成的过拟合。
 
-示例脚本（请确保你有两张以上GPU卡）:
+示例脚本（请确保你有两张以上GPU卡, 在线模型下载功能在`paddle.distributed.launch`下无法工作，
+你可能需要一个先通过单卡finetune方式下载预训练模型，或者根据[这里](#section-pretrained-models)手动下载并解压预训练模型）:
 
 ```script
 python3 -m paddle.distributed.launch \
@@ -227,7 +230,7 @@ sids = np.expand_dims(sids, 0)
 result = client(ids, sids)
 ```
 
-你也可从[此处]((https://ernie.bj.bcebos.com/ernie1.0_zh_inference_model.tar.gz).)下载一个预先制作好的ernie-1.0 base模型的 `inference_model`.
+你也可从[此处]((https://ernie.bj.bcebos.com/ernie1.0_zh_inference_model.tar.gz.)下载一个预先制作好的ernie-1.0 base模型的 `inference_model`.
 该模型没有经过finetune，一般可以用做上层模型结构的 feature-base finetune或者做为一个文本特征抽取器。
 因为该模行由老版API 产出，在进行客户端请求时需要在输入tensor后面追加一个维度：
 

demo/finetune_classifier_dygraph.py

@@ -51,11 +51,13 @@ if __name__ == '__main__':
     parser.add_argument('--bsz', type=int, default=32, help='batchsize')
     parser.add_argument('--epoch', type=int, default=3, help='epoch')
     parser.add_argument('--data_dir', type=str, required=True, help='data directory includes train / develop data')
-    parser.add_argument('--max_steps', type=int, required=True, help='max_train_steps, set this to EPOCH * NUM_SAMPLES / BATCH_SIZE')
-    parser.add_argument('--warmup_proportion', type=float, default=0.1)
+    parser.add_argument('--use_lr_decay', action='store_true', help='if set, learning rate will decay to zero at `max_steps`')
+    parser.add_argument('--warmup_proportion', type=float, default=0.1, help='if use_lr_decay is set, '
+            'learning rate will raise to `lr` at `warmup_proportion` * `max_steps` and decay to 0. at `max_steps`')
     parser.add_argument('--lr', type=float, default=5e-5, help='learning rate')
     parser.add_argument('--inference_model_dir', type=str, default=None, help='inference model output directory')
     parser.add_argument('--save_dir', type=str, default=None, help='model output directory')
+    parser.add_argument('--max_steps', type=int, default=None, help='max_train_steps, set this to EPOCH * NUM_SAMPLES / BATCH_SIZE')
     parser.add_argument('--wd', type=float, default=0.01, help='weight decay, aka L2 regularizer')
 
 
@@ -102,7 +104,11 @@ if __name__ == '__main__':
     with FD.guard(place):
         model = ErnieModelForSequenceClassification.from_pretrained(args.from_pretrained, num_labels=3, name='')
 
-        opt = AdamW(learning_rate=LinearDecay(args.lr, int(args.warmup_proportion * args.max_steps), args.max_steps), parameter_list=model.parameters(), weight_decay=args.wd)
+        if args.use_lr_decay:
+            opt = AdamW(learning_rate=LinearDecay(args.lr, int(args.warmup_proportion * args.max_steps), args.max_steps), parameter_list=model.parameters(), weight_decay=args.wd)
+        else:
+            opt = AdamW(args.lr, parameter_list=model.parameters(), weight_decay=args.wd)
+
         g_clip = F.dygraph_grad_clip.GradClipByGlobalNorm(1.0) #experimental
         for epoch in range(args.epoch):
             for step, d in enumerate(tqdm(train_ds.start(place), desc='training')):
@@ -117,7 +123,7 @@ if __name__ == '__main__':
                     acc = []
                     with FD.base._switch_tracer_mode_guard_(is_train=False):
                         model.eval()
-                        for step, d in enumerate(tqdm(dev_ds.start(), desc='evaluating %d' % epoch)):
+                        for step, d in enumerate(tqdm(dev_ds.start(place), desc='evaluating %d' % epoch)):
                             ids, sids, label = d
                             loss, logits = model(ids, sids, labels=label)
                             #print('\n'.join(map(str, logits.numpy().tolist())))

demo/finetune_mrc_dygraph.py

@@ -44,8 +44,8 @@ from ernie.modeling_ernie import ErnieModel, ErnieModelForQuestionAnswering
 from ernie.tokenizing_ernie import ErnieTokenizer, ErnieTinyTokenizer
 from ernie.optimization import AdamW, LinearDecay
 
-from ernie.mrc import mrc_reader
-from ernie.mrc import mrc_metrics
+from demo.mrc import mrc_reader
+from demo.mrc import mrc_metrics
 
 log.setLevel(logging.DEBUG)
 logging.getLogger().addHandler(log.handlers[0])

ernie/file_utils.py

@@ -41,3 +41,10 @@ def _fetch_from_remote(url, force_download=False):
     log.debug('%s cached in %s' % (url, cached_dir))
     return cached_dir
 
+
+def add_docstring(doc):
+    def func(f):
+        f.__doc__ += ('\n======other docs from supper class ======\n%s' % doc)
+        return f
+    return func
+

ernie/modeling_ernie.py

@@ -29,7 +29,7 @@ import paddle.fluid.dygraph as D
 import paddle.fluid as F
 import paddle.fluid.layers as L
 
-from ernie.file_utils import _fetch_from_remote
+from ernie.file_utils import _fetch_from_remote, add_docstring
 
 log = logging.getLogger(__name__)
 
@@ -288,6 +288,11 @@ class ErnieModel(D.Layer, PretrainedModel):
                 Mask to avoid performing attention on the padding token indices of the encoder input.
             attn_bias(optional, `Variable` of shape `[batch_size, seq_len, seq_len] or False`): 
                 3D version of `input_mask`, if set, overrides `input_mask`; if set not False, will not apply attention mask
+            past_cache(optional, tuple of two lists: cached key and cached value,
+                each is a list of `Variable`s of shape `[batch_size, seq_len, hidden_size]`):
+                cached key/value tensor that will be concated to generated key/value when performing self attention. 
+                if set, `attn_bias` should not be None.
+
         Returns:
             pooled (`Variable` of shape `[batch_size, hidden_size]`):
                 output logits of pooler classifier
@@ -360,6 +365,7 @@ class ErnieModelForSequenceClassification(ErnieModel):
         prob = cfg.get('classifier_dropout_prob', cfg['hidden_dropout_prob'])
         self.dropout = lambda i: L.dropout(i, dropout_prob=prob, dropout_implementation="upscale_in_train",) if self.training else i
 
+    @add_docstring(ErnieModel.forward.__doc__)
     def forward(self, *args, **kwargs):
         """
         Args:
@@ -400,6 +406,7 @@ class ErnieModelForTokenClassification(ErnieModel):
         prob = cfg.get('classifier_dropout_prob', cfg['hidden_dropout_prob'])
         self.dropout = lambda i: L.dropout(i, dropout_prob=prob, dropout_implementation="upscale_in_train",) if self.training else i
 
+    @add_docstring(ErnieModel.forward.__doc__)
     def forward(self, *args, **kwargs):
         """
         Args:
@@ -441,6 +448,7 @@ class ErnieModelForQuestionAnswering(ErnieModel):
         prob = cfg.get('classifier_dropout_prob', cfg['hidden_dropout_prob'])
         self.dropout = lambda i: L.dropout(i, dropout_prob=prob, dropout_implementation="upscale_in_train",) if self.training else i
 
+    @add_docstring(ErnieModel.forward.__doc__)
     def forward(self, *args, **kwargs):
         """
         Args:
@@ -460,7 +468,7 @@ class ErnieModelForQuestionAnswering(ErnieModel):
 
         start_pos = kwargs.pop('start_pos', None)
         end_pos = kwargs.pop('end_pos', None)
-        pooled, encoded, _ = super(ErnieModelForQuestionAnswering, self).forward(*args, **kwargs)
+        pooled, encoded = super(ErnieModelForQuestionAnswering, self).forward(*args, **kwargs)
         encoded = self.dropout(encoded)
         encoded = self.classifier(encoded)
         start_logit, end_logits = L.unstack(encoded, axis=-1)
@@ -529,6 +537,7 @@ class ErnieModelForPretraining(ErnieModel):
                 is_bias=True,
             )
 
+    @add_docstring(ErnieModel.forward.__doc__)
     def forward(self, *args, **kwargs):
         """
         Args:
@@ -550,7 +559,7 @@ class ErnieModelForPretraining(ErnieModel):
         mlm_labels = kwargs.pop('labels')
         mlm_pos = kwargs.pop('mlm_pos')
         nsp_labels = kwargs.pop('nsp_labels')
-        pooled, encoded, _ = super(ErnieModelForPretraining, self).forward(*args, **kwargs)
+        pooled, encoded = super(ErnieModelForPretraining, self).forward(*args, **kwargs)
         if len(mlm_labels.shape) == 1:
             mlm_labels = L.reshape(mlm_labels, [-1, 1])
         if len(nsp_labels.shape) == 1:

experimental/seq2seq/modeling_ernie_gen.py

@@ -32,6 +32,7 @@ class ErnieModelForGeneration(ErnieModel):
     resource_map = {
         'ernie-gen-base-en': ErnieModel.bce + 'model-ernie-gen-base-en.1.tar.gz',
         'ernie-gen-large-en': ErnieModel.bce + 'model-ernie-gen-large-en.1.tar.gz',
+        'ernie-gen-large-160g-en': ErnieModel.bce + 'model-ernie-gen-large-160g-en.1.tar.gz',
         'ernie-1.0': ErnieModel.bce + 'model-ernie1.0.1.tar.gz',
     }
     def __init__(self, cfg, name=None):