Model() got multiple values for argument 'nr_class' - SpaCy multi-classification model (BERT integration)

Question

Hi I am working on implementing a multi-classification model (5 classes) with the new SpaCy Model en_pytt_bertbaseuncased_lg. The code for the new pipe is here:

nlp = spacy.load('en_pytt_bertbaseuncased_lg')
textcat = nlp.create_pipe(
    'pytt_textcat',
    config={
        "nr_class":5,
        "exclusive_classes": True,
    }
)
nlp.add_pipe(textcat, last = True)

textcat.add_label("class1")
textcat.add_label("class2")
textcat.add_label("class3")
textcat.add_label("class4")
textcat.add_label("class5")

The code for the training is as follows and is based on the example from here(https://pypi.org/project/spacy-pytorch-transformers/):

def extract_cat(x):
    for key in x.keys():
        if x[key]:
            return key

# get names of other pipes to disable them during training
n_iter = 250 # number of epochs

train_data = list(zip(train_texts, [{"cats": cats} for cats in train_cats]))


dev_cats_single   = [extract_cat(x) for x in dev_cats]
train_cats_single = [extract_cat(x) for x in train_cats]
cats = list(set(train_cats_single))
recall = {}
for c in cats:
    if c is not None: 
        recall['dev_'+c] = []
        recall['train_'+c] = []



optimizer = nlp.resume_training()
batch_sizes = compounding(1.0, round(len(train_texts)/2), 1.001)

for i in range(n_iter):
    random.shuffle(train_data)
    losses = {}
    batches = minibatch(train_data, size=batch_sizes)
    for batch in batches:
        texts, annotations = zip(*batch)
        nlp.update(texts, annotations, sgd=optimizer, drop=0.2, losses=losses)
    print(i, losses)

So the structure of my data looks like this:

[('TEXT TEXT TEXT',
  {'cats': {'class1': False,
    'class2': False,
    'class3': False,
    'class4': True,
    'class5': False}}), ... ]

I am not sure why I get the following error:

TypeError                                 Traceback (most recent call last)
<ipython-input-32-1588a4eadc8d> in <module>
     21 
     22 
---> 23 optimizer = nlp.resume_training()
     24 batch_sizes = compounding(1.0, round(len(train_texts)/2), 1.001)
     25 

TypeError: Model() got multiple values for argument 'nr_class'

EDIT:

if I take out the nr_class argument, I get this error here:

ValueError: operands could not be broadcast together with shapes (1,2) (1,5)

I actually thought this would happen because I didn't specify the nr_class argument. Is that correct?

Oarsman

Answer 1

This is a regression in the most recent version we released of spacy-pytorch-transformers. Sorry about this!

The root cause is, this is another case of the evils of **kwargs. I'm looking forward to refining the spaCy API to prevent these issues in future.

You can see the offending line here: https://github.com/explosion/spacy-pytorch-transformers/blob/c1def95e1df783c69bff9bc8b40b5461800e9231/spacy_pytorch_transformers/pipeline/textcat.py#L71 . We provide the nr_class positional argument, which overlaps with the explicit argument you passed in during the config.

In order to workaround the problem, you can simply remove the nr_class key from your the config dict you're passing into spacy.create_pipe().

Sena

I see, but the point is that if I take out the nr_class argument, I get this error here: ValueError: operands could not be broadcast together with shapes (1,2) (1,5) I actually thought this would happen because I didn't specify the nr_class argument. Is that correct? – Oarsman

@HenrykBorzymowski I also have the same problem. You can add architecture='softmax_pooler_output' to config dict when creating the pipe, and it would probably work. However, other architectures like softmax_class_vector or softmax_last_hidden give this error in case of multi-class classification. I have inspected the source code and could not figure out where that 2 comes from in case of using softmax_class_vector. – Perturb

It would be great if you could take a look at this relevant issue on Github, @syllogism_. – Perturb