OWLv2: open-vocabulary object detection with Transformers

This page is the Transformers contract of the OWLv2 MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against google/owlv2-base-patch16-ensemble. 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.

Detect objects from text queries

import torch
from transformers import Owlv2ForObjectDetection, Owlv2Processor, set_seed
import json  # doc-builder: hide
import os  # doc-builder: hide

model_id = "google/owlv2-base-patch16-ensemble"
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 and dropout draw random numbers

processor = Owlv2Processor.from_pretrained(model_id)
model = Owlv2ForObjectDetection.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    del model  # doc-builder: hide
    model = Owlv2ForObjectDetection.from_pretrained(model_id, **variant).to(device)  # doc-builder: hide
model.eval()

# Two small synthetic images and two text queries per image.
images = [torch.rand(3, 224, 224), torch.rand(3, 224, 224)]
texts = [["a photo of a cat", "a photo of a dog"], ["a photo of a cat", "a photo of a dog"]]
inputs = processor(text=texts, images=images, return_tensors="pt")
# BatchFeature.to casts only floating tensors (pixel_values); token ids and masks stay integer.
inputs = inputs.to(device, dtype=model.dtype)

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

target_sizes = torch.tensor([[224, 224], [224, 224]])
results = processor.post_process_grounded_object_detection(
    outputs=outputs, target_sizes=target_sizes, threshold=0.1, text_labels=texts
)
# Keep each image's five highest-scoring detections.
detections = []
for result in results:
    top = result["scores"].argsort(descending=True)[:5].tolist()
    detections.append(
        {
            "boxes": result["boxes"][top].tolist(),
            "scores": result["scores"][top].tolist(),
            "text_labels": [result["text_labels"][i] for i in top],
        }
    )
print(json.dumps({"detections": detections}))

One training step

# Detection labels: one dict per image with class labels and normalized cx, cy, w, h boxes.
labels = [
    {
        "class_labels": torch.tensor([0], dtype=torch.int64, device=device),
        "boxes": torch.tensor([[0.5, 0.5, 0.4, 0.4]], dtype=torch.float32, device=device),
    },
    {
        "class_labels": torch.tensor([1], dtype=torch.int64, device=device),
        "boxes": torch.tensor([[0.3, 0.3, 0.2, 0.2]], dtype=torch.float32, device=device),
    },
]

model.train()
optimizer = torch.optim.SGD([p for p in model.parameters() if p.requires_grad], lr=0.01)
# Owlv2ForObjectDetection has no training loss in Transformers: match each target
# box to its nearest predicted box, then score that query's class logits (binary
# cross-entropy per text query) and regress its box (L1).
train_outputs = model(**inputs)
loss = 0
for i, target in enumerate(labels):
    query = (train_outputs.pred_boxes[i].detach() - target["boxes"][0]).abs().sum(-1).argmin()
    logits = train_outputs.logits[i, query].float()
    wanted = torch.nn.functional.one_hot(target["class_labels"][0], logits.shape[-1]).float()
    loss = loss + torch.nn.functional.binary_cross_entropy_with_logits(logits, wanted)
    loss = loss + (train_outputs.pred_boxes[i, query].float() - target["boxes"][0]).abs().sum()
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 = (
        "owlv2.vision_model.encoder.layers.0.self_attn.q_proj.weight",
        "owlv2.text_model.encoder.layers.0.self_attn.q_proj.weight",
        "class_head.dense0.weight",
    )
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "detections": detections,
        "logits_probe": outputs.logits[0, :2, :2].float().tolist(),
        "boxes_probe": outputs.pred_boxes[0, :2, :2].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))