Contains a few fixes to update notebooks to AllenNLP 1.0.0.

Mostly, updates some of the data loaders to properly import from the new allennlp-models package as well as fix up their instantiation and use.

• POS notebook output
• LM notebook output
• NER notebook output
• CNN notebook output

I thought it might be helpful to update some of these to use AllenNLP >= 1.0.0. For most of them, I followed the example set in the updated sentiment notebook. I ran into a bit of trouble with the LM notebook where there the output data point seemed to be "wrapped" twice, nested as {"tokens": {"tokens": <vector>...

I figured I'd make a PR to help aid the process of transforming these. I'm also happy to close this if it's redundant.

I reviewed the code: examples/sentiment/sst_classifier.py, and found a bug.

    self.f1_measure = F1Measure(4)

I think this code is intended to measure precision/recall/f1 for the label '4' which is the most positive sentiment. However, the integer 4 here is considered as index in the array representation. It must be converted using label mapping stored in vocab.

• chapter 1, should be text generation

Finally, a third class of text classification is unconditional text generation, where natural language text is generated stochastically from a model. You can train models so that they can generate some random academic papers, Linux source code, or even some poems and play scripts. For example, Andrej Karpathy trained an RNN model form all works of Shakespeare and succeeded in generation pieces of text that look exactly like his work (http://realworldnlpbook.com/ch1.html#karpathy15):

• 4.2.3 typo swtich in the pseudocode

def update_gru(state, word):
    new_state = update_hidden(state, word)
 
    switch = get_switch(state, word)
 
    state = swtich * new_state + (1 – switch) * state
 
    return state

• chapter 5: micro and macro should be switched.

original text:

If these metrics are computed while ignoring entity types, it’s called a micro average. For example, the micro-averaged precision is the total number of true positives of all types divided by the total number of retrieved named entities regardless of the type. On the other hand, if these metrics are computed per entity type and then get averaged, it’s called a macro average. For example, if the precision for PER and GPE is 80% and 90%, respectively, its macro average is 85%. What AllenNLP computes in the following is the micro average.

• 5.6.1

The language detection in a previous chapter used RNN and character as input.

original text:

In the first half of this section, we are going to build an English language model and train it using a generic English corpus. Before we start, we note that the RNN language model we build in this chapter operates on characters, not on words or tokens. All the RNN models we’ve seen so far operate on words, which means the input to the RNN was always sequences of words. On the other hand, the RNN we are going to use in this section takes sequences of characters as the input.

I'm now reading the book and notice a lot of bugs related to allennlp2.0. Does the author consider upgrading the code to allennlp2.0 to make it comply more with the title real world nlp?

It's a pity because this book is I think the only book using allennlp to tackle a range of general nlp tasks and I like it very much.

Some examples:

In the sst_classifier.ipynb one can note:

vocab = Vocabulary.from_instances(train_dataset + dev_dataset,
                                  min_count={'tokens': 3})

gives

unsupported operand type(s) for +: 'generator' and 'generator'

(easily fixable using list(reader.read('train.txt')))

The following two lines

train_dataset.index_with(vocab)
dev_dataset.index_with(vocab)

give

'generator' object has no attribute 'index_with'

and also not specific to allen2.0,

predictor = SentenceClassifierPredictor(model, dataset_reader=reader)

gives

AttributeError: 'StanfordSentimentTreeBankDatasetReader' object has no attribute '_tokenizer'

Hi,

While trying

predictor = SentenceClassifierPredictor(model, dataset_reader=reader)

in Sec 2.8.1 I'm getting error

AttributeError: 'StanfordSentimentTreeBankDatasetReader' object has no attribute '_tokenizer'

I see you have made some changes in this commit.

ModuleNotFoundError: No module named 'allennlp.data.dataset_readers.seq2seq' I'm trying to run the code examples/mt/mt.py with allennlp==1.0.0 and got this error, no code changes direct clone of the repo and tried to run it.

I try reproduce examples/mt/mt.py but I have CPU/CUDA error:

File "/opt/conda/lib/python3.6/site-packages/allennlp/models/encoder_decoders/simple_seq2seq.py", line 212, in forward state = self._encode(source_tokens) File "/opt/conda/lib/python3.6/site-packages/allennlp/models/encoder_decoders/simple_seq2seq.py", line 268, in _encode embedded_input = self._source_embedder(source_tokens) File "/opt/conda/lib/python3.6/site-packages/torch/nn/modules/module.py", line 493, in __call__ result = self.forward(*input, **kwargs) File "/opt/conda/lib/python3.6/site-packages/allennlp/modules/text_field_embedders/basic_text_field_embedder.py", line 123, in forward token_vectors = embedder(*tensors) File "/opt/conda/lib/python3.6/site-packages/torch/nn/modules/module.py", line 493, in __call__ result = self.forward(*input, **kwargs) File "/opt/conda/lib/python3.6/site-packages/allennlp/modules/token_embedders/embedding.py", line 143, in forward sparse=self.sparse) File "/opt/conda/lib/python3.6/site-packages/torch/nn/functional.py", line 1506, in embedding return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse) RuntimeError: Expected object of backend CPU but got backend CUDA for argument #3 'index'

I try play in kaggle enveroment