Merge branch 'GH-2400-release' of https://github.com/flairNLP/flair i…

…nto GH-2400-release

resources/docs/TUTORIAL_6_CORPUS.md

@@ -178,8 +178,7 @@ The `MultiCorpus` inherits from `Corpus`, so you can use it like any other corpu
 Flair supports many datasets out of the box. It automatically downloads and sets up the 
 data the first time you call the corresponding constructor ID. 
 
-The following datasets are supported (click category to 
-expand): 
+The following datasets are supported (**click category to expand**): 
 
 <details>
   <summary>Named Entity Recognition (NER) datasets</summary>

resources/docs/TUTORIAL_7_TRAINING_A_MODEL.md

@@ -211,9 +211,7 @@ tutorials on both for difference). The rest is exactly the same as before!
 The best results in text classification use fine-tuned transformers with `TransformerDocumentEmbeddings` as shown in the
 code below:
 
-(If you don't have a big GPU to fine-tune transformers, try `DocumentPoolEmbeddings` or `DocumentRNNEmbeddings` instead
-
-- sometimes they work just as well!)
+(If you don't have a big GPU to fine-tune transformers, try `DocumentPoolEmbeddings` or `DocumentRNNEmbeddings` instead; sometimes they work just as well!)
 
 ```python
 import torch

resources/docs/embeddings/TRANSFORMER_EMBEDDINGS.md

@@ -64,21 +64,21 @@ from flair.embeddings import TransformerWordEmbeddings
 sentence = Sentence('The grass is green.')
 
 # use only last layers
-embeddings = TransformerWordEmbeddings('bert-base-uncased', layers='-1')
+embeddings = TransformerWordEmbeddings('bert-base-uncased', layers='-1', layer_mean=False)
 embeddings.embed(sentence)
 print(sentence[0].embedding.size())
 
 sentence.clear_embeddings()
 
 # use last two layers
-embeddings = TransformerWordEmbeddings('bert-base-uncased', layers='-1,-2')
+embeddings = TransformerWordEmbeddings('bert-base-uncased', layers='-1,-2', layer_mean=False)
 embeddings.embed(sentence)
 print(sentence[0].embedding.size())
 
 sentence.clear_embeddings()
 
 # use ALL layers
-embeddings = TransformerWordEmbeddings('bert-base-uncased', layers='all')
+embeddings = TransformerWordEmbeddings('bert-base-uncased', layers='all', layer_mean=False)
 embeddings.embed(sentence)
 print(sentence[0].embedding.size())
 ```
@@ -90,7 +90,7 @@ torch.Size([1536])
 torch.Size([9984])
 ```
 
-I.e. the size of the embedding increases the mode layers we use.
+I.e. the size of the embedding increases the mode layers we use (but ONLY if layer_mean is set to False, otherwise the length is always the same).
 
 
 ### Pooling operation