attr-eomt 0.1.2__tar.gz → 1.0.0__tar.gz

{attr_eomt-0.1.2 → attr_eomt-1.0.0}/.gitignore

@@ -29,3 +29,4 @@ scripts/sample_images/
 .idea/
 .vscode/
 .DS_Store
+.env

{attr_eomt-0.1.2 → attr_eomt-1.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: attr-eomt
-Version: 0.1.2
+Version: 1.0.0
 Summary: Standalone EoMT (Encoder-only Mask Transformer) for instance segmentation, with DINOv2 init, COCO training/validation and inference.
 Project-URL: Homepage, https://github.com/imagra93/attr-eomt
 Project-URL: Documentation, https://imagra93.github.io/attr-eomt
@@ -23,6 +23,7 @@ Requires-Dist: pyyaml>=6.0
 Requires-Dist: requests>=2.25.0
 Requires-Dist: scipy>=1.7.0
 Requires-Dist: torch>=2.4.0
+Requires-Dist: torchao>=0.7.0
 Requires-Dist: torchvision>=0.19.0
 Requires-Dist: tqdm>=4.65.0
 Requires-Dist: transformers>=5.1.0
@@ -92,9 +93,9 @@ with a fixed set of **learnable queries** (the Mask2Former idea) — each query
 "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 cars
-stay two distinct queries instead of being merged — the property that lets attributes
-stay attached to the right instance.
+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
@@ -164,9 +165,9 @@ 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 car's
-make **and** its colour **and** whether a door is open — the only way to express that is
-the Cartesian product `type × colour × state × …`. That space explodes combinatorially,
+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.
 
@@ -181,11 +182,11 @@ per-instance query embedding the detector already computes — they **add, not m
 Because the heads are independent:
 
 - **Categories stay collapsed and general.** Keep a compact, well-populated primary
-  taxonomy (`car`, `person`, `fruit`) 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 `colour` head trained across many object types
-  predicts `red` on a class it *never co-occurred with in red*, because colour is
+  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.
@@ -196,25 +197,25 @@ Because the heads are independent:
   matched queries — they add only a thin linear/MLP head and a cross-entropy term, not
   a second model or a second pass.
 
-### Example: a bowl of fruit
+### Example: clothing with per-instance attributes
 
-One model segments each fruit (primary classes `apple` / `banana` / `orange` /
-`pear`) and, for **every** detection, reads off two **independent** attribute heads —
-`ripeness` (`unripe` / `turning` / `ripe`) and a quality `grade` (`A` / `B`). The
-renderer prints the primary class + score on the first row and each attribute + its
-confidence on the row beneath it.
+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.
 
-![A bowl of fruit; each box labelled with its fruit class plus a ripeness and a grade](https://raw.githubusercontent.com/imagra93/attr-eomt/main/docs/examples/example_1.png)
+![Two people in dresses; each instance labelled with its garment class plus scale, occlusion, zoom and viewpoint attributes](https://raw.githubusercontent.com/imagra93/attr-eomt/main/docs/examples/sample.jpg)
 
-`ripeness` and `grade` are *orthogonal* — 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**, **vehicles → type + colour + damage state**, **cells / leaves → type +
-health**, **apparel → garment + pattern**.
+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**.
 
-> The image above is a **simulation**: a stock photo with hand-placed detections fed
-> through the package's own renderer ([`eomt.visualize.draw_instances`](eomt/visualize.py))
-> to show the output format — not a trained model's predictions.
+> Trained on the public **[DeepFashion2](https://github.com/switchablenorms/DeepFashion2)**
+> dataset (13 garment classes + 4 attribute heads) and rendered with the package's own
+> renderer ([`eomt.visualize.draw_instances`](eomt/visualize.py)).
 
 ---
 
@@ -259,18 +260,19 @@ names) are discovered from the JSON, the same as `nc`.
 ```jsonc
 "attributes": [
   {
-    "name": "ripeness",
+    "name": "scale",
     "categories": [
-      {"id": 0, "name": "unripe"},
-      {"id": 1, "name": "turning"},
-      {"id": 2, "name": "ripe"}
+      {"id": 1, "name": "small"},
+      {"id": 2, "name": "modest"},
+      {"id": 3, "name": "large"}
     ]
   },
   {
-    "name": "grade",
+    "name": "viewpoint",
     "categories": [
-      {"id": 10, "name": "A"},
-      {"id": 20, "name": "B"}
+      {"id": 0, "name": "frontal"},
+      {"id": 1, "name": "side"},
+      {"id": 2, "name": "back"}
     ]
   }
 ]
@@ -282,14 +284,14 @@ names) are discovered from the JSON, the same as `nc`.
 {
   "id": 1, "image_id": 42, "category_id": 1,
   "segmentation": [...], "bbox": [...], "area": 1234, "iscrowd": 0,
-  "attributes": {"ripeness": 1, "grade": 20}
+  "attributes": {"scale": 3, "viewpoint": 0}
 }
 ```
 
 Notes:
 
-- Raw ids are remapped to a contiguous `0..n-1` per head (so `grade`'s `10`/`20` become
-  `0`/`1`); `categories` may be omitted, in which case the id set is inferred.
+- 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.
 
@@ -326,6 +328,27 @@ 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](https://github.com/pytorch/ao))
+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.
+
+```python
+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(...)`:

{attr_eomt-0.1.2 → attr_eomt-1.0.0}/README.md

@@ -55,9 +55,9 @@ with a fixed set of **learnable queries** (the Mask2Former idea) — each query
 "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 cars
-stay two distinct queries instead of being merged — the property that lets attributes
-stay attached to the right instance.
+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
@@ -127,9 +127,9 @@ 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 car's
-make **and** its colour **and** whether a door is open — the only way to express that is
-the Cartesian product `type × colour × state × …`. That space explodes combinatorially,
+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.
 
@@ -144,11 +144,11 @@ per-instance query embedding the detector already computes — they **add, not m
 Because the heads are independent:
 
 - **Categories stay collapsed and general.** Keep a compact, well-populated primary
-  taxonomy (`car`, `person`, `fruit`) 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 `colour` head trained across many object types
-  predicts `red` on a class it *never co-occurred with in red*, because colour is
+  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.
@@ -159,25 +159,25 @@ Because the heads are independent:
   matched queries — they add only a thin linear/MLP head and a cross-entropy term, not
   a second model or a second pass.
 
-### Example: a bowl of fruit
+### Example: clothing with per-instance attributes
 
-One model segments each fruit (primary classes `apple` / `banana` / `orange` /
-`pear`) and, for **every** detection, reads off two **independent** attribute heads —
-`ripeness` (`unripe` / `turning` / `ripe`) and a quality `grade` (`A` / `B`). The
-renderer prints the primary class + score on the first row and each attribute + its
-confidence on the row beneath it.
+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.
 
-![A bowl of fruit; each box labelled with its fruit class plus a ripeness and a grade](https://raw.githubusercontent.com/imagra93/attr-eomt/main/docs/examples/example_1.png)
+![Two people in dresses; each instance labelled with its garment class plus scale, occlusion, zoom and viewpoint attributes](https://raw.githubusercontent.com/imagra93/attr-eomt/main/docs/examples/sample.jpg)
 
-`ripeness` and `grade` are *orthogonal* — 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**, **vehicles → type + colour + damage state**, **cells / leaves → type +
-health**, **apparel → garment + pattern**.
+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**.
 
-> The image above is a **simulation**: a stock photo with hand-placed detections fed
-> through the package's own renderer ([`eomt.visualize.draw_instances`](eomt/visualize.py))
-> to show the output format — not a trained model's predictions.
+> Trained on the public **[DeepFashion2](https://github.com/switchablenorms/DeepFashion2)**
+> dataset (13 garment classes + 4 attribute heads) and rendered with the package's own
+> renderer ([`eomt.visualize.draw_instances`](eomt/visualize.py)).
 
 ---
 
@@ -222,18 +222,19 @@ names) are discovered from the JSON, the same as `nc`.
 ```jsonc
 "attributes": [
   {
-    "name": "ripeness",
+    "name": "scale",
     "categories": [
-      {"id": 0, "name": "unripe"},
-      {"id": 1, "name": "turning"},
-      {"id": 2, "name": "ripe"}
+      {"id": 1, "name": "small"},
+      {"id": 2, "name": "modest"},
+      {"id": 3, "name": "large"}
     ]
   },
   {
-    "name": "grade",
+    "name": "viewpoint",
     "categories": [
-      {"id": 10, "name": "A"},
-      {"id": 20, "name": "B"}
+      {"id": 0, "name": "frontal"},
+      {"id": 1, "name": "side"},
+      {"id": 2, "name": "back"}
     ]
   }
 ]
@@ -245,14 +246,14 @@ names) are discovered from the JSON, the same as `nc`.
 {
   "id": 1, "image_id": 42, "category_id": 1,
   "segmentation": [...], "bbox": [...], "area": 1234, "iscrowd": 0,
-  "attributes": {"ripeness": 1, "grade": 20}
+  "attributes": {"scale": 3, "viewpoint": 0}
 }
 ```
 
 Notes:
 
-- Raw ids are remapped to a contiguous `0..n-1` per head (so `grade`'s `10`/`20` become
-  `0`/`1`); `categories` may be omitted, in which case the id set is inferred.
+- 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.
 
@@ -289,6 +290,27 @@ 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](https://github.com/pytorch/ao))
+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.
+
+```python
+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(...)`:

{attr_eomt-0.1.2 → attr_eomt-1.0.0}/docs/assets/01-architecture.svg

@@ -1 +1 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 940 470" width="940" height="470"><rect x="1" y="1" width="938" height="468" rx="16" fill="#0e1318" stroke="#2a3744" stroke-width="1.5"/><text x="32" y="46" font-family="DejaVu Sans, sans-serif" font-size="23" fill="#d7e0e8" text-anchor="start" font-weight="bold">attr-eomt</text><text x="33" y="66" font-family="DejaVu Sans Mono, monospace" font-size="12" fill="#8c9aa8" text-anchor="start" font-weight="normal">architecture · the heads tap one shared embedding</text><circle cx="525" cy="42" r="4.5" fill="#4ec9b0"/><text x="536" y="46" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#8c9aa8">shared embedding</text><circle cx="659.6" cy="42" r="4.5" fill="#5aa6f0"/><text x="670.6" y="46" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#8c9aa8">primary head</text><circle cx="767.8" cy="42" r="4.5" fill="#e2b341"/><text x="778.8" y="46" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#8c9aa8">attribute head</text><defs><marker id="a" markerWidth="9" markerHeight="9" refX="7" refY="4.5" orient="auto"><path d="M0,0 L9,4.5 L0,9 z" fill="#5e6b78"/></marker><marker id="aq" markerWidth="9" markerHeight="9" refX="7" refY="4.5" orient="auto"><path d="M0,0 L9,4.5 L0,9 z" fill="#4ec9b0"/></marker></defs><rect x="34" y="220" width="104" height="60" rx="9" fill="#1b232c" stroke="#2a3744" stroke-width="1.4" /><text x="86" y="248" font-family="DejaVu Sans Mono, monospace" font-size="13" fill="#d7e0e8" text-anchor="middle" font-weight="normal">image</text><text x="86" y="266" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">644²</text><path d="M138,250 H176" stroke="#5e6b78" stroke-width="1.6" fill="none" marker-end="url(#a)"/><rect x="180" y="188" width="178" height="124" rx="9" fill="#1b232c" stroke="#5aa6f0" stroke-width="1.4" /><text x="269" y="224" font-family="DejaVu Sans, sans-serif" font-size="14.5" fill="#5aa6f0" text-anchor="middle" font-weight="bold">DINOv2 ViT</text><text x="269" y="244" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#5e6b78" text-anchor="middle" font-weight="normal">last blocks +</text><text x="269" y="260" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#5e6b78" text-anchor="middle" font-weight="normal">learnable queries</text><text x="269" y="286" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#8c9aa8" text-anchor="middle" font-weight="normal">NMS-free</text><path d="M358,250 H426" stroke="#5e6b78" stroke-width="1.6" fill="none" marker-end="url(#a)"/><rect x="430" y="221" width="150" height="58" rx="9" fill="#10201c" stroke="#4ec9b0" stroke-width="1.8" /><text x="505" y="246" font-family="DejaVu Sans Mono, monospace" font-size="13.5" fill="#4ec9b0" text-anchor="middle" font-weight="bold">query_embed</text><text x="505" y="265" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">[B, Q, hidden]</text><text x="505" y="204" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#8c9aa8" text-anchor="middle" font-weight="normal">▲ forward hook on class_predictor</text><path d="M580,205 C638,205 638,94 704,94" stroke="#4ec9b0" stroke-width="2.1" fill="none" marker-end="url(#aq)" /><rect x="708" y="70" width="200" height="48" rx="9" fill="#1b232c" stroke="#5aa6f0" stroke-width="1.4" /><text x="808.0" y="91" font-family="DejaVu Sans Mono, monospace" font-size="12.5" fill="#5aa6f0" text-anchor="middle" font-weight="normal">class_predictor</text><text x="808.0" y="107" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">primary class · nc+1</text><path d="M580,218 C638,218 638,164 704,164" stroke="#4ec9b0" stroke-width="2.1" fill="none" marker-end="url(#aq)" /><rect x="708" y="140" width="200" height="48" rx="9" fill="#1b232c" stroke="#5aa6f0" stroke-width="1.4" /><text x="808.0" y="161" font-family="DejaVu Sans Mono, monospace" font-size="12.5" fill="#5aa6f0" text-anchor="middle" font-weight="normal">mask / box head</text><text x="808.0" y="177" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">geometry</text><path d="M580,228 C638,228 638,254 704,254" stroke="#4ec9b0" stroke-width="2.1" fill="none" marker-end="url(#aq)" /><rect x="708" y="230" width="200" height="48" rx="9" fill="#1b232c" stroke="#e2b341" stroke-width="1.4" /><text x="808.0" y="251" font-family="DejaVu Sans Mono, monospace" font-size="12.5" fill="#e2b341" text-anchor="middle" font-weight="normal">aux · "ripeness"</text><text x="808.0" y="267" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">3 classes</text><path d="M580,236 C638,236 638,324 704,324" stroke="#4ec9b0" stroke-width="2.1" fill="none" marker-end="url(#aq)" /><rect x="708" y="300" width="200" height="48" rx="9" fill="#1b232c" stroke="#e2b341" stroke-width="1.4" /><text x="808.0" y="321" font-family="DejaVu Sans Mono, monospace" font-size="12.5" fill="#e2b341" text-anchor="middle" font-weight="normal">aux · "grade"</text><text x="808.0" y="337" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">2 classes</text><path d="M580,243 C638,243 638,396 704,396" stroke="#4ec9b0" stroke-width="2.1" fill="none" marker-end="url(#aq)" stroke-dasharray="3 4"/><rect x="708" y="372" width="200" height="48" rx="9" fill="#1b232c" stroke="#5e6b78" stroke-width="1.4" stroke-dasharray="4 4"/><text x="808.0" y="393" font-family="DejaVu Sans Mono, monospace" font-size="12.5" fill="#8c9aa8" text-anchor="middle" font-weight="normal">… as many as data defines</text><text x="470.0" y="444" font-family="DejaVu Sans Mono, monospace" font-size="11.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">teal = the single per-query embedding · every head is a small classifier on top of it</text></svg>
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 940 470" width="940" height="470"><rect x="1" y="1" width="938" height="468" rx="16" fill="#0e1318" stroke="#2a3744" stroke-width="1.5"/><text x="32" y="46" font-family="DejaVu Sans, sans-serif" font-size="23" fill="#d7e0e8" text-anchor="start" font-weight="bold">attr-eomt</text><text x="33" y="66" font-family="DejaVu Sans Mono, monospace" font-size="12" fill="#8c9aa8" text-anchor="start" font-weight="normal">architecture · the heads tap one shared embedding</text><circle cx="525" cy="42" r="4.5" fill="#4ec9b0"/><text x="536" y="46" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#8c9aa8">shared embedding</text><circle cx="659.6" cy="42" r="4.5" fill="#5aa6f0"/><text x="670.6" y="46" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#8c9aa8">primary head</text><circle cx="767.8" cy="42" r="4.5" fill="#e2b341"/><text x="778.8" y="46" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#8c9aa8">attribute head</text><defs><marker id="a" markerWidth="9" markerHeight="9" refX="7" refY="4.5" orient="auto"><path d="M0,0 L9,4.5 L0,9 z" fill="#5e6b78"/></marker><marker id="aq" markerWidth="9" markerHeight="9" refX="7" refY="4.5" orient="auto"><path d="M0,0 L9,4.5 L0,9 z" fill="#4ec9b0"/></marker></defs><rect x="34" y="220" width="104" height="60" rx="9" fill="#1b232c" stroke="#2a3744" stroke-width="1.4" /><text x="86" y="248" font-family="DejaVu Sans Mono, monospace" font-size="13" fill="#d7e0e8" text-anchor="middle" font-weight="normal">image</text><text x="86" y="266" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">644²</text><path d="M138,250 H176" stroke="#5e6b78" stroke-width="1.6" fill="none" marker-end="url(#a)"/><rect x="180" y="188" width="178" height="124" rx="9" fill="#1b232c" stroke="#5aa6f0" stroke-width="1.4" /><text x="269" y="224" font-family="DejaVu Sans, sans-serif" font-size="14.5" fill="#5aa6f0" text-anchor="middle" font-weight="bold">DINOv2 ViT</text><text x="269" y="244" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#5e6b78" text-anchor="middle" font-weight="normal">last blocks +</text><text x="269" y="260" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#5e6b78" text-anchor="middle" font-weight="normal">learnable queries</text><text x="269" y="286" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#8c9aa8" text-anchor="middle" font-weight="normal">NMS-free</text><path d="M358,250 H426" stroke="#5e6b78" stroke-width="1.6" fill="none" marker-end="url(#a)"/><rect x="430" y="221" width="150" height="58" rx="9" fill="#10201c" stroke="#4ec9b0" stroke-width="1.8" /><text x="505" y="246" font-family="DejaVu Sans Mono, monospace" font-size="13.5" fill="#4ec9b0" text-anchor="middle" font-weight="bold">query_embed</text><text x="505" y="265" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">[B, Q, hidden]</text><text x="505" y="204" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#8c9aa8" text-anchor="middle" font-weight="normal">▲ forward hook on class_predictor</text><path d="M580,205 C638,205 638,94 704,94" stroke="#4ec9b0" stroke-width="2.1" fill="none" marker-end="url(#aq)" /><rect x="708" y="70" width="200" height="48" rx="9" fill="#1b232c" stroke="#5aa6f0" stroke-width="1.4" /><text x="808.0" y="91" font-family="DejaVu Sans Mono, monospace" font-size="12.5" fill="#5aa6f0" text-anchor="middle" font-weight="normal">class_predictor</text><text x="808.0" y="107" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">primary class · nc+1</text><path d="M580,218 C638,218 638,164 704,164" stroke="#4ec9b0" stroke-width="2.1" fill="none" marker-end="url(#aq)" /><rect x="708" y="140" width="200" height="48" rx="9" fill="#1b232c" stroke="#5aa6f0" stroke-width="1.4" /><text x="808.0" y="161" font-family="DejaVu Sans Mono, monospace" font-size="12.5" fill="#5aa6f0" text-anchor="middle" font-weight="normal">mask / box head</text><text x="808.0" y="177" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">geometry</text><path d="M580,228 C638,228 638,254 704,254" stroke="#4ec9b0" stroke-width="2.1" fill="none" marker-end="url(#aq)" /><rect x="708" y="230" width="200" height="48" rx="9" fill="#1b232c" stroke="#e2b341" stroke-width="1.4" /><text x="808.0" y="251" font-family="DejaVu Sans Mono, monospace" font-size="12.5" fill="#e2b341" text-anchor="middle" font-weight="normal">aux · "scale"</text><text x="808.0" y="267" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">3 classes</text><path d="M580,236 C638,236 638,324 704,324" stroke="#4ec9b0" stroke-width="2.1" fill="none" marker-end="url(#aq)" /><rect x="708" y="300" width="200" height="48" rx="9" fill="#1b232c" stroke="#e2b341" stroke-width="1.4" /><text x="808.0" y="321" font-family="DejaVu Sans Mono, monospace" font-size="12.5" fill="#e2b341" text-anchor="middle" font-weight="normal">aux · "viewpoint"</text><text x="808.0" y="337" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">3 classes</text><path d="M580,243 C638,243 638,396 704,396" stroke="#4ec9b0" stroke-width="2.1" fill="none" marker-end="url(#aq)" stroke-dasharray="3 4"/><rect x="708" y="372" width="200" height="48" rx="9" fill="#1b232c" stroke="#5e6b78" stroke-width="1.4" stroke-dasharray="4 4"/><text x="808.0" y="393" font-family="DejaVu Sans Mono, monospace" font-size="12.5" fill="#8c9aa8" text-anchor="middle" font-weight="normal">… as many as data defines</text><text x="470.0" y="444" font-family="DejaVu Sans Mono, monospace" font-size="11.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">teal = the single per-query embedding · every head is a small classifier on top of it</text></svg>

attr_eomt-1.0.0/docs/assets/02-factorized-labels.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 940 500" width="940" height="500"><rect x="1" y="1" width="938" height="498" rx="16" fill="#0e1318" stroke="#2a3744" stroke-width="1.5"/><text x="32" y="48" font-family="DejaVu Sans, sans-serif" font-size="23" fill="#d7e0e8" text-anchor="start" font-weight="bold">Factorize the label space</text><text x="33" y="69" font-family="DejaVu Sans Mono, monospace" font-size="12" fill="#8c9aa8" text-anchor="start" font-weight="normal">a garment's type, viewpoint and occlusion are orthogonal — don't multiply them</text><line x1="470.0" y1="110" x2="470.0" y2="430" stroke="#2a3744" stroke-width="1.2" stroke-dasharray="3 5"/><text x="470.0" y="254.0" font-family="DejaVu Sans, sans-serif" font-size="13" fill="#5e6b78" text-anchor="middle" font-weight="bold">vs</text><circle cx="470.0" cy="248.0" r="18" fill="#0e1318" stroke="#2a3744"/><text x="470.0" y="254.0" font-family="DejaVu Sans, sans-serif" font-size="13" fill="#8c9aa8" text-anchor="middle" font-weight="bold">vs</text><text x="245" y="128" font-family="DejaVu Sans Mono, monospace" font-size="12" fill="#e07a5f" text-anchor="middle" font-weight="bold">FLAT · ONE HEAD</text><text x="108" y="171.0" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#8c9aa8" text-anchor="end" font-weight="normal">top</text><rect x="120" y="150" width="72" height="34" rx="5" fill="#1b232c" stroke="rgba(224,122,95,.45)" stroke-width="1"/><text x="156.0" y="171.0" font-family="DejaVu Sans Mono, monospace" font-size="9.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">top·frnt</text><rect x="200" y="150" width="72" height="34" rx="5" fill="#1b232c" stroke="rgba(224,122,95,.45)" stroke-width="1"/><text x="236.0" y="171.0" font-family="DejaVu Sans Mono, monospace" font-size="9.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">top·side</text><rect x="280" y="150" width="72" height="34" rx="5" fill="#1b232c" stroke="rgba(224,122,95,.45)" stroke-width="1"/><text x="316.0" y="171.0" font-family="DejaVu Sans Mono, monospace" font-size="9.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">top·back</text><text x="108" y="213.0" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#8c9aa8" text-anchor="end" font-weight="normal">dress</text><rect x="120" y="192" width="72" height="34" rx="5" fill="#1b232c" stroke="rgba(224,122,95,.45)" stroke-width="1"/><text x="156.0" y="213.0" font-family="DejaVu Sans Mono, monospace" font-size="9.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">drs·frnt</text><rect x="200" y="192" width="72" height="34" rx="5" fill="#1b232c" stroke="rgba(224,122,95,.45)" stroke-width="1"/><text x="236.0" y="213.0" font-family="DejaVu Sans Mono, monospace" font-size="9.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">drs·side</text><rect x="280" y="192" width="72" height="34" rx="5" fill="#1b232c" stroke="rgba(224,122,95,.45)" stroke-width="1"/><text x="316.0" y="213.0" font-family="DejaVu Sans Mono, monospace" font-size="9.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">drs·back</text><text x="108" y="255.0" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#8c9aa8" text-anchor="end" font-weight="normal">skirt</text><rect x="120" y="234" width="72" height="34" rx="5" fill="#1b232c" stroke="rgba(224,122,95,.45)" stroke-width="1"/><text x="156.0" y="255.0" font-family="DejaVu Sans Mono, monospace" font-size="9.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">skt·frnt</text><rect x="200" y="234" width="72" height="34" rx="5" fill="#1b232c" stroke="rgba(224,122,95,.45)" stroke-width="1"/><text x="236.0" y="255.0" font-family="DejaVu Sans Mono, monospace" font-size="9.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">skt·side</text><rect x="280" y="234" width="72" height="34" rx="5" fill="#1b232c" stroke="rgba(224,122,95,.45)" stroke-width="1"/><text x="316.0" y="255.0" font-family="DejaVu Sans Mono, monospace" font-size="9.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">skt·back</text><text x="108" y="297.0" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#8c9aa8" text-anchor="end" font-weight="normal">trousers</text><rect x="120" y="276" width="72" height="34" rx="5" fill="#1b232c" stroke="rgba(224,122,95,.45)" stroke-width="1"/><text x="156.0" y="297.0" font-family="DejaVu Sans Mono, monospace" font-size="9.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">trs·frnt</text><rect x="200" y="276" width="72" height="34" rx="5" fill="#1b232c" stroke="rgba(224,122,95,.45)" stroke-width="1"/><text x="236.0" y="297.0" font-family="DejaVu Sans Mono, monospace" font-size="9.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">trs·side</text><rect x="280" y="276" width="72" height="34" rx="5" fill="#1b232c" stroke="rgba(224,122,95,.45)" stroke-width="1"/><text x="316.0" y="297.0" font-family="DejaVu Sans Mono, monospace" font-size="9.5" fill="#5e6b78" text-anchor="middle" font-weight="normal">trs·back</text><text x="245" y="338" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#8c9aa8" text-anchor="middle" font-weight="normal">× {visible, occluded} states</text><text x="245" y="376" font-family="DejaVu Sans Mono, monospace" font-size="15" fill="#e07a5f" text-anchor="middle" font-weight="bold">4 × 3 × 2</text><text x="245" y="402" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#5e6b78" text-anchor="middle" font-weight="normal">= 24 leaf classes</text><text x="700" y="128" font-family="DejaVu Sans Mono, monospace" font-size="12" fill="#e2b341" text-anchor="middle" font-weight="bold">FACTORIZED · INDEPENDENT HEADS</text><rect x="560" y="165" width="320" height="58" rx="9" fill="#1b232c" stroke="#e2b341" stroke-width="1.3"/><text x="576" y="190" font-family="DejaVu Sans Mono, monospace" font-size="13" fill="#e2b341" text-anchor="start" font-weight="bold">type head</text><text x="576" y="208" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="start" font-weight="normal">top · dress · skirt · trousers</text><circle cx="852" cy="194" r="15" fill="#10201c" stroke="#e2b341"/><text x="852" y="199" font-family="DejaVu Sans Mono, monospace" font-size="14" fill="#4ec9b0" text-anchor="middle" font-weight="bold">4</text><rect x="560" y="237" width="320" height="58" rx="9" fill="#1b232c" stroke="#e2b341" stroke-width="1.3"/><text x="576" y="262" font-family="DejaVu Sans Mono, monospace" font-size="13" fill="#e2b341" text-anchor="start" font-weight="bold">viewpoint head</text><text x="576" y="280" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="start" font-weight="normal">frontal · side · back</text><circle cx="852" cy="266" r="15" fill="#10201c" stroke="#e2b341"/><text x="852" y="271" font-family="DejaVu Sans Mono, monospace" font-size="14" fill="#4ec9b0" text-anchor="middle" font-weight="bold">3</text><rect x="560" y="309" width="320" height="58" rx="9" fill="#1b232c" stroke="#e2b341" stroke-width="1.3"/><text x="576" y="334" font-family="DejaVu Sans Mono, monospace" font-size="13" fill="#e2b341" text-anchor="start" font-weight="bold">occlusion head</text><text x="576" y="352" font-family="DejaVu Sans Mono, monospace" font-size="10.5" fill="#5e6b78" text-anchor="start" font-weight="normal">visible · occluded</text><circle cx="852" cy="338" r="15" fill="#10201c" stroke="#e2b341"/><text x="852" y="343" font-family="DejaVu Sans Mono, monospace" font-size="14" fill="#4ec9b0" text-anchor="middle" font-weight="bold">2</text><text x="700" y="399" font-family="DejaVu Sans Mono, monospace" font-size="15" fill="#4ec9b0" text-anchor="middle" font-weight="bold">4 + 3 + 2</text><text x="700" y="425" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#5e6b78" text-anchor="middle" font-weight="normal">= 9 outputs · each head sees every instance</text><text x="470.0" y="472" font-family="DejaVu Sans Mono, monospace" font-size="11" fill="#5e6b78" text-anchor="middle" font-weight="normal">compact primary taxonomy keeps its sample counts · unseen attribute × class combinations still generalize</text></svg>