SAM2: promptable image segmentation with Transformers

This page is the Transformers contract of the SAM2 MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against facebook/sam2.1-hiera-tiny. 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.

Segment an object from a point prompt

import numpy as np
import torch
from PIL import Image
from transformers import Sam2Model, Sam2Processor, set_seed
import json  # doc-builder: hide
import os  # doc-builder: hide

model_id = "facebook/sam2.1-hiera-tiny"
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

processor = Sam2Processor.from_pretrained(model_id)
model = Sam2Model.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    del model
    model = Sam2Model.from_pretrained(model_id, **variant).to(device)

# Two synthetic 480x640 images: a bright rectangle on a dim noisy background.
# `boxes` holds each rectangle as (x0, y0, x1, y1) in pixels.
rng = np.random.default_rng(42)
boxes = [(160, 120, 400, 300), (240, 180, 560, 420)]  # both cover the image centre
images, object_masks = [], []
for x0, y0, x1, y1 in boxes:
    pixels = rng.integers(0, 80, (480, 640, 3), dtype=np.uint8)
    pixels[y0:y1, x0:x1] = (230, 180, 40)
    images.append(Image.fromarray(pixels))
    mask = np.zeros((480, 640), dtype=np.float32)
    mask[y0:y1, x0:x1] = 1.0
    object_masks.append(mask)

# The same positive point on both images, at the image centre: only the pixels
# decide which rectangle is segmented.
# Nesting: [image][object][point][x, y] and [image][object][label].
input_points = [[[[320, 240]]] for _ in boxes]
input_labels = [[[1]] for _ in boxes]
inputs = processor(images=images, input_points=input_points, input_labels=input_labels, return_tensors="pt")
# Moves every tensor to the device; only floating tensors (pixels, points) take the model's dtype
inputs = inputs.to(device, dtype=model.dtype)

model.eval()
with torch.no_grad():
    outputs = model(**inputs, multimask_output=False)

# pred_masks: low-resolution mask logits, (images, objects, masks, 256, 256)
# iou_scores: the model's estimate of each mask's quality, (images, objects, masks)
masks = processor.post_process_masks(outputs.pred_masks.float().cpu(), inputs["original_sizes"])
print(f"pred_masks: {list(outputs.pred_masks.shape)}, upscaled to {list(masks[0].shape)}")
print(f"iou_scores: {outputs.iou_scores.flatten().tolist()}")

# Compact probes: a 4x4 grid of mask logits per image, and the IoU estimates
mask_probe = outputs.pred_masks[:, 0, 0, ::64, ::64].float().flatten().tolist()
iou_probe = outputs.iou_scores.float().flatten().tolist()

One fine-tuning step

Sam2Model returns predictions but no loss, so a fine-tuning loop computes its own. The loss below is binary cross-entropy between the low-resolution mask logits and each object's ground-truth mask, resized to the logits' 256x256 grid. It is computed in float32 so that a bfloat16 model gets a full-precision loss.

import torch.nn.functional as F

# Ground truth at the mask logits' resolution, (images, objects, masks, 256, 256)
target = torch.from_numpy(np.stack(object_masks))[:, None]
target = F.interpolate(target, size=outputs.pred_masks.shape[-2:], mode="nearest")
target = target[:, None].to(device)

model.train()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
outputs = model(**inputs, multimask_output=False)
loss = F.binary_cross_entropy_with_logits(outputs.pred_masks.float(), target)
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 = (
        "vision_encoder.backbone.patch_embed.projection.weight",
        "vision_encoder.neck.convs.0.weight",
        "prompt_encoder.point_embed.weight",
        "mask_decoder.upscale_conv1.weight",
    )
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "pred_masks": mask_probe,
        "iou_scores": iou_probe,
        "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))