How do I resolve this error "ValueError: The number of parallel adapters and the number of active heads must match."?

[charliezuo]

**
Hi, this is the code I wrote to run the parallel inference using adapters library. I have trained two adapters, and I am leveraging what you have shown in the tutorial from this notebook:
https://github.com/adapter-hub/adapters/blob/main/notebooks/05_Parallel_Adapter_Inference.ipynb

I am able to do inference with both of my adapter models separately, without issues.

I am not sure where the issue may be, I can provide additional details if you need. Based on what you showed, this code should be able to work.
**

from transformers import AutoTokenizer
from adapters import AutoAdapterModel

#Defining the configuration for the model
tokenizer = AutoTokenizer.from_pretrained("FacebookAI/xlm-roberta-large")

#Setting up the model
model = AutoAdapterModel.from_pretrained(
    "FacebookAI/xlm-roberta-large"
)
ph_adapters = model.load_adapter("/path1")
NER_adapter = model.load_adapter("/path2)

import adapters.composition as ac

model.set_active_adapters(ac.Parallel(NER_adapter, ph_adapters))
model.active_heads = ac.Parallel("ner_head", "ph_head")

import torch

def analyze_sentence(sentence):
    tokens = tokenizer.tokenize(sentence)
    input_ids = torch.tensor(tokenizer.convert_tokens_to_ids(tokens))
    outputs = model(input_ids)

    # Post-process NER output
    ner_labels_map = model.get_labels_dict(NER_adapter)
    ner_label_ids = torch.argmax(outputs[0].logits, dim=2).numpy().squeeze().tolist()
    ner_labels = [ner_labels_map[id_] for id_ in ner_label_ids]
    annotated = []
    for token, label_id in zip(tokens, ner_label_ids):
        token = token.replace('\u0120', '')
        label = ner_labels_map[label_id]
        annotated.append(f"{token}<{label}>")
    print("NER: " + " ".join(annotated))

    # Post-process sentiment output
    sentiment_labels = model.get_labels_dict(ph_adapters)
    label_id = torch.argmax(outputs[1].logits).item()
    print("PH: " + sentiment_labels[label_id])
    print()

sentences = [
  "I'm Emma and I have a chest pain and it's been making it hard to breathe. I think I might have pneumonia.",
  "John Smith's Address is: 14 HAMLET Ct, CHATHAM CANADA.",
  "Ben Emma ve göğüs ağrım var ve nefes almamı zorlaştırıyor. Zatürre olabileceğimi düşünüyorum."
]

for sentence in sentences:
  analyze_sentence(sentence)

[calpt]

Hey @charliezuo, the error message you mentioned might occur if not all loaded adapters have corresponding heads added to the model correctly. Usually, this should automatically happen with load_adapter() if a head is available in the checkpoint. You can check if this is the case like this after loading the adapters:

assert NER_adapter in model.heads
assert ph_adapters in model.heads

There's one minor error in your code: it should be active_head instead of active_heads to explicitly activate loaded heads, i.e.:

model.active_head = ac.Parallel(NER_adapter, ph_adapters)

This line is also not necessary as set_active_adapters() will automatically also activate the heads if available.

Hope this helps!

[charliezuo]

Thank you for your reply.