Every example I've looked at so far seems to use a shared vocabulary between source and target languages, and I'm wondering if that is a hard-coded constraint of the Huggingface models, or my misunderstanding, or I've just not looked in the right place yet?
To take a random example, when I look at the files here, https://huggingface.co/Helsinki-NLP/opus-mt-en-zls/tree/main, I see separate "spm" (sentience piece model) files for source and target languages, and they are of different sizes (792kb vs. 850kb). But there is only a single "vocab.json" file. And the config.json file only mentions a single "vocab_size": 57680.
I've also been experimenting, e.g. tokenizer(inputs, text_target=inputs, return_tensors="pt"). If source and target used different vocabulary I would expect the returned input_ids and labels to use different numbers. But every model I've tried so far the numbers are identical (NO, my mistake - see update below).
Can a Huggingface tokenizer even support two vocabularies? If not then a model would need two tokenizers, which seems to clash with the way AutoTokenizer works.
UPDATE
Here is a test script to show the above model is actually using two spm vocabs with AutoTokenizer.
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
model_name = 'Helsinki-NLP/opus-mt-en-zls'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
inputs = ['Filter all items from same host']
targets = ['Filtriraj sve stavke s istog hosta']
x=tokenizer(inputs, text_target=targets, return_tensors="pt")
print(x)
print(tokenizer.decode(x['input_ids'][0]))
print(tokenizer.decode(x['labels'][0]))
print("\nGiving inputs on both sides")
x=tokenizer(inputs, text_target=inputs, return_tensors="pt")
print(x) ## Expecting to see different numbers if they use different vocabs
print(tokenizer.decode(x['input_ids'][0]))
print(tokenizer.decode(x['labels'][0]))
print("\nGiving targets on both sides")
x=tokenizer(targets, text_target=targets, return_tensors="pt") ## Expecting to see different numbers if they use different vocabs
print(x)
print(tokenizer.decode(x['input_ids'][0]))
print(tokenizer.decode(x['labels'][0]))
print(model)
The output is:
{'input_ids': tensor([[10373, 90, 8255, 98, 605, 6276, 0]]), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1]]), 'labels': tensor([[11638, 1392, 7636, 386, 35861, 95, 2130, 218, 6276, 27,
0]])}
▁Filter all▁items from same host</s>
Filtriraj sve stavke s istog hosta</s>
Giving inputs on both sides
{'input_ids': tensor([[10373, 90, 8255, 98, 605, 6276, 0]]), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1]]), 'labels': tensor([[11638, 911, 90, 3188, 7, 98, 605, 6276, 0]])}
▁Filter all▁items from same host</s>
Filter all items from same host</s>
Giving targets on both sides
{'input_ids': tensor([[11638, 1392, 7636, 95, 120, 914, 465, 478, 95, 29,
25, 897, 6276, 27, 0]]), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]), 'labels': tensor([[11638, 1392, 7636, 386, 35861, 95, 2130, 218, 6276, 27,
0]])}
Filtriraj sve stavke s istog hosta</s>
Filtriraj sve stavke s istog hosta</s>
When I choose identical strings in English or Croatian it gives slightly different numbers, showing that different tokenizers are involved. You can then see that the different ids sometimes map back to an identical string, sometimes not.
But when I print out the model we see it is actually a shared vocabulary, which makes the two spm models a bit pointless.
(encoder): MarianEncoder(
(embed_tokens): Embedding(57680, 512, padding_idx=57679)
...
(decoder): MarianDecoder(
(embed_tokens): Embedding(57680, 512, padding_idx=57679)
...
(lm_head): Linear(in_features=512, out_features=57680, bias=False)
I haven't got as far as finding out if a non-shared vocabulary is possible, but still yet to see evidence of one.
