added wordpiece - updated readme

README.md

 # pytorch-pretrained-BERT
-A PyTorch version of Google's pretrained BERT model
+A PyTorch version of Google's pretrained BERT model as described in
+
+No bells and whitles, just:
+- [one class](bert_model.py) with a clean commented version of Google's BERT model that can load the weights pre-trained by Google's authors,
+- [another class](data_processor.py) with all you need to pre- and post-process text data for the model (tokenize and encode),
+- and [a script](download_weigths.sh) to download Google's pre-trained weights.
+
+Here is how to use these:
+
+```python
+from .bert_model import BERT
+from .data_processor import DataProcessor
+
+bert_model = BERT(bert_model_path='.')
+data_processor = DataProcessor(bert_vocab_path='.')
+
+input_sentence = "We are playing with the BERT model."
+
+tensor_input = data_processor.encode(input_sentence)
+tensor_output = bert_model(prepared_input)
+output_sentence = data_processor.decode(tensor_output)
+```

bert_model.py

@@ -13,7 +13,6 @@ from typing import NamedTuple, List
 import copy
 import io
 import json
-import logging
 import math
 import pathlib
 import re
@@ -273,10 +272,7 @@ class BERT(torch.nn.Module):
         config = BERTConfig(
             embedding_dim,
             num_heads,
-                embedding_dropout_probability,
-                attention_dropout_probability,
-                residual_dropout_probability,
-                activation_function,
+            dropout_probability,
         )
 
         # the embedding size is vocab_size + n_special embeddings + n_ctx
@@ -288,7 +284,7 @@ class BERT(torch.nn.Module):
         self.num_output_layers = 1 + num_layers
 
         self.embed = torch.nn.Embedding(embedding_size, embedding_dim)
-        self.drop = torch.nn.Dropout(embedding_dropout_probability)
+        self.drop = torch.nn.Dropout(dropout_probability)
 
         block = Block(n_ctx, config, scale=True)
         self.h = torch.nn.ModuleList([copy.deepcopy(block) for _ in range(num_layers)])
@@ -332,16 +328,13 @@ class BERT(torch.nn.Module):
                      names: List[str] = _PARAMETER_NAMES) -> None:
         # pylint: disable=dangerous-default-value
 
-        logger.info(f"loading weights from {bert_model_path}")
-        # if `file_path` is a URL, redirect to the cache
-
         with tarfile.open(bert_model_path) as tmp:
             num_params_files = len([member for member in tmp.getmembers() if member.name.endswith('.npy')])
             shapesfile = tmp.extractfile('model/params_shapes.json')
             if shapesfile:
                 shapes = json.loads(shapesfile.read())
             else:
-                raise ConfigurationError("unable to find model/params_shapes.json in the archive")
+                raise Exception("unable to find model/params_shapes.json in the archive")
 
             # numpy can't read from a tarfile directly, so we need a workaround
             # https://github.com/numpy/numpy/issues/7989#issuecomment-341656702

example.py

-"""
-Show how to use HuggingFace's PyTorch implementation of Google's BERT Model.
-"""
-from .bert_model import BERT
-from .prepare_inputs import DataPreprocessor
-
-bert_model = BERT()
-bert_model.load_from('.')
-data_processor = DataProcessor(encoder_file_path='.')
-
-input_sentence = "We are playing with the BERT model."
-print("BERT inputs: {}".format(input_sentence))
-
-tensor_input = data_processor.encode(input_sentence)
-tensor_output = bert_model(prepared_input)
-output_sentence = data_processor.decode(tensor_output)
-
-print("BERT predicted: {}".format(output_sentence))