!5 修复了一些错误

Merge pull request !5 from WillQvQ/dev

.Jenkinsfile

@@ -29,7 +29,7 @@ pipeline {
                     steps {
                         sh 'python -m spacy download en'
                         sh 'pip install fitlog'
-                        sh 'pytest ./test --html=test_results.html --self-contained-html'
+                        sh 'pytest ./tests --html=test_results.html --self-contained-html'
                     }
                 }
             }

.travis.yml

@@ -14,7 +14,7 @@ install:
 # command to run tests
 script:
   - python -m spacy download en
-  - pytest --cov=fastNLP test/
+  - pytest --cov=fastNLP tests/
 
 after_success:
   - bash <(curl -s https://codecov.io/bash)

MANIFEST.in

@@ -1,7 +1,7 @@
 include requirements.txt
 include LICENSE
 include README.md
-prune test/
+prune tests/
 prune reproduction/
 prune fastNLP/api
 prune fastNLP/automl

fastNLP/embeddings/bert_embedding.py

@@ -93,7 +93,7 @@ class BertEmbedding(ContextualEmbedding):
         """
         super(BertEmbedding, self).__init__(vocab, word_dropout=word_dropout, dropout=dropout)
 
-        if word_dropout>0:
+        if word_dropout > 0:
             assert vocab.unknown != None, "When word_drop>0, Vocabulary must contain the unknown token."
 
         if model_dir_or_name.lower() in PRETRAINED_BERT_MODEL_DIR:
@@ -370,17 +370,29 @@ class _BertWordModel(nn.Module):
                  include_cls_sep: bool = False, pooled_cls: bool = False, auto_truncate: bool = False, min_freq=2):
         super().__init__()
 
-        self.tokenzier = BertTokenizer.from_pretrained(model_dir_or_name)
-        self.encoder = BertModel.from_pretrained(model_dir_or_name)
-        self._max_position_embeddings = self.encoder.config.max_position_embeddings
-        #  检查encoder_layer_number是否合理
-        encoder_layer_number = len(self.encoder.encoder.layer)
         if isinstance(layers, list):
             self.layers = [int(l) for l in layers]
         elif isinstance(layers, str):
             self.layers = list(map(int, layers.split(',')))
         else:
             raise TypeError("`layers` only supports str or list[int]")
+        assert len(self.layers) > 0, "There is no layer selected!"
+
+        neg_num_output_layer = -16384
+        pos_num_output_layer = 0
+        for layer in self.layers:
+            if layer < 0:
+                neg_num_output_layer = max(layer, neg_num_output_layer)
+            else:
+                pos_num_output_layer = max(layer, pos_num_output_layer)
+
+        self.tokenzier = BertTokenizer.from_pretrained(model_dir_or_name)
+        self.encoder = BertModel.from_pretrained(model_dir_or_name,
+                                                 neg_num_output_layer=neg_num_output_layer,
+                                                 pos_num_output_layer=pos_num_output_layer)
+        self._max_position_embeddings = self.encoder.config.max_position_embeddings
+        #  检查encoder_layer_number是否合理
+        encoder_layer_number = len(self.encoder.encoder.layer)
         for layer in self.layers:
             if layer < 0:
                 assert -layer <= encoder_layer_number, f"The layer index:{layer} is out of scope for " \

fastNLP/embeddings/roberta_embedding.py

@@ -196,20 +196,30 @@ class _RobertaWordModel(nn.Module):
                  include_cls_sep: bool = False, pooled_cls: bool = False, auto_truncate: bool = False, min_freq=2):
         super().__init__()
 
-        self.tokenizer = RobertaTokenizer.from_pretrained(model_dir_or_name)
-        self.encoder = RobertaModel.from_pretrained(model_dir_or_name)
-        # 由于RobertaEmbedding中设置了padding_idx为1, 且使用了非常神奇的position计算方式，所以-2
-        self._max_position_embeddings = self.encoder.config.max_position_embeddings - 2
-        #  检查encoder_layer_number是否合理
-        encoder_layer_number = len(self.encoder.encoder.layer)
-
         if isinstance(layers, list):
             self.layers = [int(l) for l in layers]
         elif isinstance(layers, str):
             self.layers = list(map(int, layers.split(',')))
         else:
             raise TypeError("`layers` only supports str or list[int]")
+        assert len(self.layers) > 0, "There is no layer selected!"
+
+        neg_num_output_layer = -16384
+        pos_num_output_layer = 0
+        for layer in self.layers:
+            if layer < 0:
+                neg_num_output_layer = max(layer, neg_num_output_layer)
+            else:
+                pos_num_output_layer = max(layer, pos_num_output_layer)
 
+        self.tokenizer = RobertaTokenizer.from_pretrained(model_dir_or_name)
+        self.encoder = RobertaModel.from_pretrained(model_dir_or_name,
+                                                    neg_num_output_layer=neg_num_output_layer,
+                                                    pos_num_output_layer=pos_num_output_layer)
+        # 由于RobertaEmbedding中设置了padding_idx为1, 且使用了非常神奇的position计算方式，所以-2
+        self._max_position_embeddings = self.encoder.config.max_position_embeddings - 2
+        #  检查encoder_layer_number是否合理
+        encoder_layer_number = len(self.encoder.encoder.layer)
         for layer in self.layers:
             if layer < 0:
                 assert -layer <= encoder_layer_number, f"The layer index:{layer} is out of scope for " \

fastNLP/modules/encoder/bert.py

@@ -366,19 +366,28 @@ class BertLayer(nn.Module):
 
 
 class BertEncoder(nn.Module):
-    def __init__(self, config):
+    def __init__(self, config, num_output_layer=-1):
         super(BertEncoder, self).__init__()
         layer = BertLayer(config)
         self.layer = nn.ModuleList([copy.deepcopy(layer) for _ in range(config.num_hidden_layers)])
+        num_output_layer = num_output_layer if num_output_layer >= 0 else (len(self.layer) + num_output_layer)
+        self.num_output_layer = max(min(num_output_layer, len(self.layer)), 0)
+        if self.num_output_layer + 1 < len(self.layer):
+            logger.info(f'The transformer encoder will early exit after layer-{self.num_output_layer} '
+                        f'(start from 0)!')
 
     def forward(self, hidden_states, attention_mask, output_all_encoded_layers=True):
         all_encoder_layers = []
-        for layer_module in self.layer:
+        for idx, layer_module in enumerate(self.layer):
+            if idx > self.num_output_layer:
+                break
             hidden_states = layer_module(hidden_states, attention_mask)
             if output_all_encoded_layers:
                 all_encoder_layers.append(hidden_states)
         if not output_all_encoded_layers:
             all_encoder_layers.append(hidden_states)
+        if len(all_encoder_layers) == 0:
+            all_encoder_layers.append(hidden_states)
         return all_encoder_layers
 
 
@@ -435,6 +444,9 @@ class BertModel(nn.Module):
         self.config = config
         self.hidden_size = self.config.hidden_size
         self.model_type = 'bert'
+        neg_num_output_layer = kwargs.get('neg_num_output_layer', -1)
+        pos_num_output_layer = kwargs.get('pos_num_output_layer', self.config.num_hidden_layers - 1)
+        self.num_output_layer = max(neg_num_output_layer + self.config.num_hidden_layers, pos_num_output_layer)
         if hasattr(config, 'sinusoidal_pos_embds'):
             self.model_type = 'distilbert'
         elif 'model_type' in kwargs:
@@ -445,7 +457,7 @@ class BertModel(nn.Module):
         else:
             self.embeddings = BertEmbeddings(config)
 
-        self.encoder = BertEncoder(config)
+        self.encoder = BertEncoder(config, num_output_layer=self.num_output_layer)
         if self.model_type != 'distilbert':
             self.pooler = BertPooler(config)
         else:

fastNLP/modules/encoder/roberta.py

@@ -64,8 +64,8 @@ class RobertaModel(BertModel):
     undocumented
     """
 
-    def __init__(self, config):
-        super().__init__(config)
+    def __init__(self, config, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
 
         self.embeddings = RobertaEmbeddings(config)
         self.apply(self.init_bert_weights)

fastNLP/modules/encoder/seq2seq_encoder.py

@@ -132,7 +132,7 @@ class TransformerSeq2SeqEncoder(Seq2SeqEncoder):
         x = self.input_fc(x)
         x = F.dropout(x, p=self.dropout, training=self.training)
 
-        encoder_mask = seq_len_to_mask(seq_len)
+        encoder_mask = seq_len_to_mask(seq_len, max_len=max_src_len)
         encoder_mask = encoder_mask.to(device)
 
         for layer in self.layer_stacks:

test/core/test_dataset.py → tests/core/test_dataset.py

@@ -228,7 +228,7 @@ class TestDataSetMethods(unittest.TestCase):
         def split_sent(ins):
             return ins['raw_sentence'].split()
         csv_loader = CSVLoader(headers=['raw_sentence', 'label'], sep='\t')
-        data_bundle = csv_loader.load('test/data_for_tests/tutorial_sample_dataset.csv')
+        data_bundle = csv_loader.load('tests/data_for_tests/tutorial_sample_dataset.csv')
         dataset = data_bundle.datasets['train']
         dataset.drop(lambda x: len(x['raw_sentence'].split()) == 0, inplace=True)
         dataset.apply(split_sent, new_field_name='words', is_input=True)

test/core/test_utils.py → tests/core/test_utils.py

@@ -120,8 +120,8 @@ class TestCache(unittest.TestCase):
     def test_cache_save(self):
         try:
             start_time = time.time()
-            embed, vocab, d = process_data_1('test/data_for_tests/embedding/small_static_embedding/word2vec_test.txt',
-                                             'test/data_for_tests/cws_train')
+            embed, vocab, d = process_data_1('tests/data_for_tests/embedding/small_static_embedding/word2vec_test.txt',
+                                             'tests/data_for_tests/cws_train')
             end_time = time.time()
             pre_time = end_time - start_time
             with open('test/demo1.pkl', 'rb') as f:
@@ -130,8 +130,8 @@ class TestCache(unittest.TestCase):
             for i in range(embed.shape[0]):
                 self.assertListEqual(embed[i].tolist(), _embed[i].tolist())
             start_time = time.time()
-            embed, vocab, d = process_data_1('test/data_for_tests/embedding/small_static_embedding/word2vec_test.txt',
-                                             'test/data_for_tests/cws_train')
+            embed, vocab, d = process_data_1('tests/data_for_tests/embedding/small_static_embedding/word2vec_test.txt',
+                                             'tests/data_for_tests/cws_train')
             end_time = time.time()
             read_time = end_time - start_time
             print("Read using {:.3f}, while prepare using:{:.3f}".format(read_time, pre_time))
@@ -142,7 +142,7 @@ class TestCache(unittest.TestCase):
     def test_cache_save_overwrite_path(self):
         try:
             start_time = time.time()
-            embed, vocab, d = process_data_1('test/data_for_tests/embedding/small_static_embedding/word2vec_test.txt', 'test/data_for_tests/cws_train',
+            embed, vocab, d = process_data_1('tests/data_for_tests/embedding/small_static_embedding/word2vec_test.txt', 'tests/data_for_tests/cws_train',
                                              _cache_fp='test/demo_overwrite.pkl')
             end_time = time.time()
             pre_time = end_time - start_time
@@ -152,8 +152,8 @@ class TestCache(unittest.TestCase):
             for i in range(embed.shape[0]):
                 self.assertListEqual(embed[i].tolist(), _embed[i].tolist())
             start_time = time.time()
-            embed, vocab, d = process_data_1('test/data_for_tests/embedding/small_static_embedding/word2vec_test.txt',
-                                             'test/data_for_tests/cws_train',
+            embed, vocab, d = process_data_1('tests/data_for_tests/embedding/small_static_embedding/word2vec_test.txt',
+                                             'tests/data_for_tests/cws_train',
                                              _cache_fp='test/demo_overwrite.pkl')
             end_time = time.time()
             read_time = end_time - start_time
@@ -165,8 +165,8 @@ class TestCache(unittest.TestCase):
     def test_cache_refresh(self):
         try:
             start_time = time.time()
-            embed, vocab, d = process_data_1('test/data_for_tests/embedding/small_static_embedding/word2vec_test.txt',
-                                             'test/data_for_tests/cws_train',
+            embed, vocab, d = process_data_1('tests/data_for_tests/embedding/small_static_embedding/word2vec_test.txt',
+                                             'tests/data_for_tests/cws_train',
                                              _refresh=True)
             end_time = time.time()
             pre_time = end_time - start_time
@@ -176,8 +176,8 @@ class TestCache(unittest.TestCase):
             for i in range(embed.shape[0]):
                 self.assertListEqual(embed[i].tolist(), _embed[i].tolist())
             start_time = time.time()
-            embed, vocab, d = process_data_1('test/data_for_tests/embedding/small_static_embedding/word2vec_test.txt',
-                                             'test/data_for_tests/cws_train',
+            embed, vocab, d = process_data_1('tests/data_for_tests/embedding/small_static_embedding/word2vec_test.txt',
+                                             'tests/data_for_tests/cws_train',
                                              _refresh=True)
             end_time = time.time()
             read_time = end_time - start_time