LUKE: masked language modeling with entity awareness

This page is the Transformers contract of the LUKE MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against studio-ousia/luke-base. The MIC runner runs the same blocks against each slim fixture and each checkpoint in integration.json. Lines marked # doc-builder: hide only matter there: they swap in the pinned checkpoint or variant and print the result record.

Predict the masked word

import torch
from transformers import AutoModelForMaskedLM, AutoTokenizer, set_seed
import json  # doc-builder: hide
import os  # doc-builder: hide

model_id = "studio-ousia/luke-base"
model_id = os.environ.get("MIC_MODEL", model_id)  # doc-builder: hide
variant = json.loads(os.environ.get("MIC_MODEL_KWARGS", "{}"))  # doc-builder: hide
device = "cuda" if torch.cuda.is_available() else "cpu"
set_seed(0)  # seeds torch, NumPy, and Python: loading, dropout, and data augmentation draw random numbers

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForMaskedLM.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    del model  # doc-builder: hide
    model = AutoModelForMaskedLM.from_pretrained(model_id, **variant).to(device)  # doc-builder: hide

countries = ["France", "Germany", "Italy", "the United Kingdom"]
prompts = [f"The capital of {country} is {tokenizer.mask_token}." for country in countries]
inputs = tokenizer(prompts, return_tensors="pt", padding=True).to(device)

# LUKE expects entity inputs; use a single dummy entity per example
batch_size = inputs["input_ids"].shape[0]
entity_ids = torch.zeros((batch_size, 1), dtype=torch.long, device=device)
entity_attention_mask = torch.ones((batch_size, 1), dtype=torch.long, device=device)
entity_token_type_ids = torch.zeros((batch_size, 1), dtype=torch.long, device=device)
entity_position_ids = torch.zeros((batch_size, 1, 1), dtype=torch.long, device=device)

with torch.no_grad():
    outputs = model(
        **inputs,
        entity_ids=entity_ids,
        entity_attention_mask=entity_attention_mask,
        entity_token_type_ids=entity_token_type_ids,
        entity_position_ids=entity_position_ids,
    )
    logits = outputs.logits
    entity_logits = outputs.entity_logits

masked = inputs["input_ids"] == tokenizer.mask_token_id
predicted = logits[masked].argmax(dim=-1)
print(tokenizer.convert_ids_to_tokens(predicted))

One training step

answers = ["paris", "berlin", "rome", "london"]
answer_ids = tokenizer.convert_tokens_to_ids(answers)
labels = torch.full_like(inputs["input_ids"], -100)
labels[masked] = torch.tensor(answer_ids, device=device)

model.train()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
loss = model(
    **inputs,
    entity_ids=entity_ids,
    entity_attention_mask=entity_attention_mask,
    entity_token_type_ids=entity_token_type_ids,
    entity_position_ids=entity_position_ids,
    labels=labels,
).loss
loss.backward()
before = {n: p.detach().cpu().clone() for n, p in model.named_parameters() if p.requires_grad}  # doc-builder: hide
optimizer.step()
print(f"loss: {loss.item():.4f}")

if os.environ.get("MIC_RUNNER"):  # doc-builder: hide
    named = dict(model.named_parameters())
    checked = (
        "luke.encoder.layer.0.attention.self.query.weight",
        "luke.encoder.layer.0.intermediate.dense.weight",
        "lm_head.dense.weight",
    )
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "predicted_tokens": predicted.tolist(),
        "answer_logits": logits[masked][:, answer_ids].float().tolist(),
        "entity_logits": entity_logits[:, 0, :5].float().tolist(),
        "training": {
            "loss": loss.item(),
            "gradients": {
                n: bool(
                    named[n].grad is not None
                    and torch.count_nonzero(named[n].grad) > 0
                    and torch.isfinite(named[n].grad).all()
                )
                for n in checked
            },
            "updated": any(
                not torch.equal(p.detach().cpu(), before[n])
                for n, p in model.named_parameters()
                if n in before
            ),
        },
    }
    print(json.dumps(result))