ConvBERT: base encoder with Transformers

This page is the Transformers contract of the ConvBERT MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against YituTech/conv-bert-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.

YituTech/conv-bert-base is a pretraining-only checkpoint: it ships ConvBertModel (no task head, no tokenizer), so this page runs the encoder and trains a small linear probe on its pooled hidden states.

Encode a batch

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

model_id = "YituTech/conv-bert-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 = AutoModel.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    del model  # doc-builder: hide
    model = AutoModel.from_pretrained(model_id, **variant).to(device)  # doc-builder: hide
model.eval()

sentences = [
    "The quick brown fox jumps over the lazy dog.",
    "ConvBERT mixes convolution with self-attention in every encoder layer.",
]
inputs = tokenizer(sentences, return_tensors="pt", padding=True).to(device)

with torch.no_grad():
    outputs = model(**inputs)

hidden = outputs.last_hidden_state
pooled = hidden[:, 0]  # first token of each row is the [CLS] summary
print(hidden.shape, pooled.shape)

One training step

# A linear probe on pooled hidden states is the only trainable part: the checkpoint has no task head.
probe = torch.nn.Linear(model.config.hidden_size, 2).to(device=device, dtype=model.dtype)
model.train()
optimizer = torch.optim.SGD(list(model.parameters()) + list(probe.parameters()), lr=0.01)
labels = torch.tensor([0, 1], device=device)

outputs = model(**inputs)
logits = probe(outputs.last_hidden_state[:, 0])
loss = torch.nn.functional.cross_entropy(logits, labels)
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 = (
        "encoder.layer.0.attention.self.query.weight",
        "encoder.layer.0.attention.self.key_conv_attn_layer.depthwise.weight",
        "encoder.layer.0.intermediate.dense.weight",
    )
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "hidden_probe": hidden[0, :3, :2].float().tolist(),
        "pooled_probe": pooled[0, :4].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))