attr-eomt 0.1.3

Standalone EoMT (Encoder-only Mask Transformer) for instance segmentation, with DINOv2 init, COCO training/validation and inference.

One query embedding, many independent labels.

📖 Read the annotated explainer →

---

Abstract

attr-eomt is a standalone EoMT (Encoder-only Mask Transformer) for instance segmentation and object detection, extended with one feature that sets it apart: independent per-instance attribute heads. Alongside the usual mask/box + class output, the model predicts one or several orthogonal attributes for every detected instance — colour, state, grade, laterality, anything — read straight off the same per-query embedding the detector already computes. Instead of folding every distinction into one combinatorial class list, attr-eomt factorizes the label space: a small, well-populated primary taxonomy plus thin attribute heads that add rather than multiply. Each head reuses the detector's own Hungarian match, so attributes train and infer for near-zero extra compute — no second model, no second pass, and the primary detection metric is left exactly as it was. The result is composable, generalizing per-instance labels that can even predict attribute × class combinations never seen during training.

It is a clean-room, Apache-2.0-compatible reimplementation: weights you train are yours to release.

from eomt import EoMT

model = EoMT("l")                            # fresh large model (DINOv2 backbone)
model.train(data="coco", epochs=50)          # COCO 2017 auto-downloads if missing

model = EoMT("runs/train/eomt-l")            # reload a run — size/classes/heads auto-detected
model.predict("images/", plot=True)          # render masks/boxes + per-instance attributes

---

Architecture

EoMT is a DINOv2-with-registers ViT whose last few transformer blocks are augmented with a fixed set of learnable queries (the Mask2Former idea) — each query is one "slot" that latches onto one object instance. After the encoder runs, every query emits a single vector, the per-query embedding of shape [B, Q, hidden]. The whole model is then just "turn that embedding into predictions": a class head for the primary label and a mask/box head for geometry. It is NMS-free, so two overlapping garments stay two distinct queries instead of being merged — the property that lets attributes stay attached to the right instance.

The attribute heads add nothing to this picture except themselves: they tap the exact same embedding (captured non-invasively with a forward hook), each a small classifier on top.

Two model families: segmentation & detection

Both families share the same DINOv2 encoder, query mechanism, NMS-free matching and auxiliary heads — they differ only in the head on top and what they output:

family	--task	output	metric driving best.pt
instance (default)	instance	per-instance masks + boxes + class	segm/mAP
detect	detect	per-instance boxes + class (DETR-style box head, no masks)	bbox/mAP

EoMT("l").train(data="coco", family="instance")   # masks (default)
EoMT("l").train(data="coco", family="detect")     # boxes only

The family is recorded in the checkpoint, so val / predict pick the right post-processing automatically. Everything below applies identically to both.

Models & sizes

size	backbone	hidden	layers	heads	queries
s	DINOv2-small	384	12	6	100
b	DINOv2-base	768	12	12	200
l	DINOv2-large	1024	24	16	200

Default input is a patch-14-aligned square (644 = 14 × 46) so DINOv2 weights load 1:1.

Compute & inference speed

Measured on a single NVIDIA GeForce RTX 5090, 644 × 644 input, batch size 1. GFLOPs are multiply-accumulates at that resolution (attention included); latency / throughput are the median over 50 runs after warm-up, under torch.amp.autocast (fp16) — the package's own inference path.

instance family (masks + boxes + class):

size	params	GFLOPs	latency (fp16)	throughput (fp16)	throughput (fp32)
s	24.0 M	128	8.4 ms	119 img/s	70 img/s
b	93.9 M	430	17.4 ms	58 img/s	32 img/s
l	317 M	1144	30.2 ms	33 img/s	15 img/s

detect family (boxes + class, no mask head):

size	params	GFLOPs	latency (fp16)	throughput (fp16)	throughput (fp32)
s	22.7 M	89	2.9 ms	348 img/s	120 img/s
b	88.6 M	276	5.3 ms	190 img/s	60 img/s
l	308 M	881	13.6 ms	74 img/s	21 img/s

Dropping the mask-upsampling head makes detect substantially lighter and ~1.3–3× faster. Figures are for the detector itself (backbone + queries + heads); the attribute heads add a thin linear/MLP per head and are negligible by design.

---

⭐ Method — factorizing the label space

This is the contribution. Conventional detectors fold every distinction into a single flat label space. When an object has both a type and several attributes — a garment's class and its viewpoint and whether it's occluded — the only way to express that is the Cartesian product type × viewpoint × occlusion × …. That space explodes combinatorially, starves each leaf class of training examples, and breaks Hungarian matching by multiplying the query targets.

attr-eomt factorizes instead. The primary head stays small and general; orthogonal attributes are predicted by independent secondary heads, each reading the same per-instance query embedding the detector already computes — they add, not multiply.

Because the heads are independent:

• Categories stay collapsed and general. Keep a compact, well-populated primary taxonomy (short_sleeve_top, dress, trousers) and push fine-grained or orthogonal distinctions into attributes — every primary class keeps its full sample count instead of being shattered into rare leaves.
• Unseen combinations generalize. A viewpoint head trained across many garment types predicts side on a class it never co-occurred with sideways, because viewpoint is learned independently of type. The model composes attribute × class combinations that never appear in the training data — combinations a flat label space cannot even represent.
• Instances stay separate. EoMT is NMS-free, so two overlapping same-class objects remain two distinct queries; each carries its own attribute predictions rather than being merged.
• It rides along for free. Attributes reuse the backbone and the detector's own matched queries — they add only a thin linear/MLP head and a cross-entropy term, not a second model or a second pass.

Example: clothing with per-instance attributes

One model segments each garment (primary classes like vest_dress / short_sleeve_top / long_sleeve_dress / skirt / trousers …) and, for every detection, reads off four independent attribute heads — scale (small / modest / large), occlusion (no / slight / medium), zoom_in (no / medium / large) and viewpoint (frontal / side / back). The renderer prints the primary class + score on the first row and each attribute + its confidence on the rows beneath it.

The four attributes are orthogonal to the garment class — they vary independently — which is exactly the case that's awkward to fold into the primary class space. The same pattern fits any "class plus per-instance sub-labels" task: retail shelves → product + facing, documents → element + role, cells → type + health.

Trained on the public DeepFashion2 dataset (13 garment classes + 4 attribute heads) and rendered with the package's own renderer (eomt.visualize.draw_instances).

---

Training — it rides on the detector's own match

Attributes never run their own matcher. Detection already solves "which query is responsible for which ground-truth object" via the Hungarian matcher; attributes simply reuse that same query→GT assignment and read the answer off the matched queries.

• Embedding source. Each head reads the per-query embedding — the input to EoMT's class_predictor, captured with a forward hook ([B, Q, hidden]).
• Matching. Supervision reuses EoMT's own Hungarian matcher (model.eomt.criterion.matcher), so every attribute is trained on the same query→GT assignment the detection loss used; the attribute is read after matching.
• Gate. An optional IoU gate drops barely-overlapping matched pairs (common early in training) so attributes only learn from queries that actually localize the object.
• Loss. Cross-entropy per head over matched queries, summed across heads and scaled by aux_w (default 1.0), added to the detector loss. Empty-match batches contribute a graph-preserving zero, and missing labels use ignore_index and contribute nothing.
• Checkpoint selection is unchanged. The attribute "rides along": its per-head matched-query accuracy is shown live and written to metrics.csv, but never drives best.pt (still segm/mAP or bbox/mAP).
• Inference. Each result attaches aux = {head: {"ids", "probs"}} for the kept detections, and predict(plot=True) renders each attribute next to the class label using names stored in the checkpoint.

---

Data format (auto-discovered from the COCO JSON)

Attributes live inside the COCO annotations — each annotation is already a per-instance object, so alignment is automatic and pycocotools still parses it. Just two additions to a standard COCO file; no YAML changes — heads (count, classes, names) are discovered from the JSON, the same as nc.

1. A top-level attributes list — one entry per head, defining its vocabulary:

"attributes": [
  {
    "name": "scale",
    "categories": [
      {"id": 1, "name": "small"},
      {"id": 2, "name": "modest"},
      {"id": 3, "name": "large"}
    ]
  },
  {
    "name": "viewpoint",
    "categories": [
      {"id": 0, "name": "frontal"},
      {"id": 1, "name": "side"},
      {"id": 2, "name": "back"}
    ]
  }
]

2. A per-annotation attributes map — {head: raw_id} on each instance:

{
  "id": 1, "image_id": 42, "category_id": 1,
  "segmentation": [...], "bbox": [...], "area": 1234, "iscrowd": 0,
  "attributes": {"scale": 3, "viewpoint": 0}
}

Notes:

• Raw ids are remapped to a contiguous 0..n-1 per head (so scale's 1/2/3 become 0/1/2); categories may be omitted, in which case the id set is inferred.
• A missing per-annotation value defaults to 0; a JSON with no attributes ⇒ detection-only, exactly as before.

A tiny, self-contained example (two heads, including a non-contiguous id set) lives in sample_data/.

---

Install

pip install attr-eomt                  # from PyPI
pip install "attr-eomt[logging]"       # + tensorboard/wandb
pip install -e ".[dev]"                # from source (editable; [dev] adds pytest/build/twine)

Usage

Everything goes through one class. Initialize from a size (fresh model, pretrained DINOv2 backbone) or from a checkpoint / run folder (family, size, classes, image size, normalization and any auxiliary heads are auto-detected from the .pt):

from eomt import EoMT

# Train on COCO 2017 (auto-downloaded on first run):
EoMT("l").train(data="coco", epochs=50, batch=4)

# ...or any COCO-format dataset (point at its data.yaml):
EoMT("s").train(data="sample_data/data.yaml", epochs=1, batch=1)

# Validate and predict from a trained run:
EoMT("runs/train/eomt-l").val(data="coco")
EoMT("runs/train/eomt-l").predict("images/", plot=True)   # writes annotated images

Compress to int8

compress() applies int8 weight-only quantization (via torchao) to the ViT transformer blocks — the bulk of the parameters — while keeping the prediction heads in full precision. It's data-free (no calibration) and shrinks the checkpoint ≈3.4× on l with no measurable mAP loss.

m = EoMT("runs/train/eomt-l")
m.compress("int8", data="coco", save="runs/train/eomt-int8/weights/best.pt")  # quantize, validate, save
# size/mAP/latency deltas are returned (and written to compression_metrics.json in the
# run folder); omit `data` to skip the before/after validation

# Reload like any run — the recipe is recorded in the checkpoint, so the int8 layout
# is rebuilt automatically:
EoMT("runs/train/eomt-int8").predict("images/")

The compressed model is GPU-only (torchao's int8 kernels have no CPU path), and best.pt exported from an EMA run quantizes exactly those EMA weights.

Useful train() options

Passed as keyword arguments to model.train(...):

arg	default	effect
family	"instance"	"instance" (masks) or "detect" (boxes only)
nominal_batch / accum	16 / 0	gradient accumulation to an effective batch of nominal_batch; accum=N sets the step count explicitly
ema / ema_decay / ema_tau	True / 0.9999 / 2000	validate & export best.pt from an EMA of the weights
llrd	0.85	layer-wise LR decay on the DINOv2 backbone (1.0 = flat backbone_lr_mult)
min_scale / max_scale	0.1 / 2.0	Large-Scale Jitter range
letterbox	True	aspect-preserving letterbox eval (vs legacy square stretch); recorded in the checkpoint
flip_prob	0.5	horizontal-flip probability. Set 0 for datasets with a left/right attribute — hflip mirrors pixels without swapping the laterality label
mask_anneal	True	anneal masked attention 1→0 over training (EoMT recipe)
aux_w	1.0	weight on the summed secondary-head loss

Training recipe (defaults)

Defaults follow the EoMT/Mask2Former fine-tuning recipe; each piece is a keyword argument, so legacy behaviour is one override away.

• Effective batch via gradient accumulation — LR/WD/clip are tuned for an effective batch of 16; EoMT is a ViT (LayerNorm, no BatchNorm) so accumulation ≈ a true large batch at a fraction of the memory.
• EMA weights validated and saved as best.pt; last.pt holds live weights + optimizer + EMA state for exact resume.
• Large-Scale Jitter for training, letterbox for eval/inference (mode stored per-checkpoint so val/predict match training automatically).
• AdamW with no weight decay on norms/biases/embeddings and layer-wise LR decay on the backbone.
• Masked-attention annealing 1 → 0 so the final stretch trains mask-free and matches efficient (mask-less) inference.

---

Roadmap / future work

• Model export. ONNX / TensorRT (and friends) for deployment — currently out of scope; the inference path is being kept export-friendly.
• Keypoints. A keypoint/pose head family alongside instance and detect (the code already carries a family parameter so new heads slot in without API churn).
• Pretrained COCO checkpoints. None are published yet. COCO-trained s/b/l weights will be released on the Hugging Face Hub (the from_pretrained / hf:// loading plumbing is already in place and waiting for them).
• Multi-image re-ID via contrastive learning. Train the auxiliary head with a contrastive objective so each instance's query embedding becomes a re-identification vector — matching the same object across images, frames and cameras for tracking and retrieval. The aux head already produces a per-instance embedding from the detector's own matched queries; re-ID reuses that signal instead of bolting on a separate model.