Upload 12 head(s) from 1a8130c92226

DMS/dinov2_clip_gated/README.md

@@ -0,0 +1,52 @@
+# DINOv2-CLIP + Gated Multimodal Unit on Apple Dense Material Segmentation
+
+A material-classification **fusion head** from SegObjectMaterial. It classifies the
+material of a masked object from two frozen embeddings:
+
+| | |
+|---|---|
+| vision | DINOv2 ViT-B/14 (CLS token, frozen) -> 768-d |
+| language | CLIP ViT-L/14 text encoder over Qwen2.5-VL-7B captions (frozen) -> 768-d |
+| head | Gated Multimodal Unit (`gated`, defaults), 1,722,937 parameters |
+| classes | 57 (Apple Dense Material Segmentation) |
+
+**This repo holds only the head.** Features must come from the same encoders,
+captioner and prompts as training (see the SegObjectMaterial repository); the
+encoders are downloaded from their own model pages under their own licenses.
+
+## Evaluation
+
+Held-out split recorded in `split.json` (source images never seen in training).
+Single training run -- differences under ~0.5 mIoU points between heads are within
+run-to-run noise.
+
+| mIoU | mAcc | eval samples |
+|---|---|---|
+| 0.4543 | 0.6373 | 186,231 |
+
+## Training
+
+```json
+{
+  "learning_rate": 5e-05,
+  "epochs": 50,
+  "batch_size": 32,
+  "weight_decay": 0.0001,
+  "lr_schedule": "constant",
+  "early_stopping_patience": 0,
+  "checkpoint": "best validation loss"
+}
+```
+
+## Use
+
+```python
+from utils.hub import load_head_from_hub      # SegObjectMaterial repository
+head, info = load_head_from_hub("s0um0r0y/SegObjectMaterial", "DMS/dinov2_clip_gated")
+logits = head(vision_features, language_features)   # [B, 768], [B, 768]
+```
+
+## Integrity
+
+- `model.safetensors` sha256 `0715e31f97766d6da1185b402bf47e44a44db6f0bcb24f7e9cb743b30232200c`
+- source config `config/config_dinov2_clip.toml`, code commit `1a8130c922267fb0786a33e3854a7e00763cc38f+dirty`

DMS/dinov2_clip_gated/config.toml

@@ -0,0 +1,62 @@
+# Apple DMS -- BASE pipeline: DINOv2 ViT-B/14 (768) + Qwen2.5-VL -> CLIP text (768)
+#
+# Pair this with:
+#   src/vision_stream.py + src/language_stream.py
+#   run_serialization.py  ->  train.py  (the raw/original scripts)
+#
+# config/config.toml is the OTHER pipeline (DINOv3 ViT-L/16 1024 + SigLIP2),
+# driven by run_serialization_latest_models.py -> train_latest_models.py.
+# Mixing them raises the dimension-mismatch ValueError in the fusion head.
+
+[data]
+dataset_type = "apple_dms"
+image_dir = "<local>/images"
+mask_dir = "<local>/labels"
+# Rebuilt automatically on first run; the index now stores one entry per
+# CONNECTED COMPONENT, so any pre-existing labels.json must be removed first.
+labels_file = "<local>/labels_instances.json"
+# Fresh directory: features are now object-grounded (bbox-cropped vision,
+# bbox-prompted captions), so the old tensor_cache_dir/ contents are stale.
+tensor_cache_dir = "<local>/tensor_cache_dir_grounded"
+
+# Connected-component filtering (measured on 400 real DMS masks):
+# keeping every component gives 7.5x more samples than the old per-(image,class)
+# scheme -- 2.45M samples / 47 GB / ~7.5x the VLM captioning time -- because 69%
+# of components are under 0.1% of the frame and crop to a few pixels.
+# A component is kept when area >= max(min_instance_area, min_instance_frac * frame).
+min_instance_area = 1000        # absolute pixel floor
+min_instance_frac = 0.001       # and at least 0.1% of the frame
+max_instances_per_class = 0     # 0 = keep all; set e.g. 3 to cap the long tail
+
+
+[model]
+# fusion head, read by train_latest_models.py. This config always trained "gated"
+# (the default when the key was absent) -- best_model_dinov2_clip.pth / best_model_fmd.pth
+# are gated heads. config_dinov2_clip_mlp.toml / fmd_config_dinov2_clip_mlp.toml hold the MLP.
+fusion = "gated"
+vision_dim = 768      # DINOv2 ViT-B/14 CLS token
+language_dim = 768    # CLIP ViT-L/14 text pooler_output
+num_classes = 57
+
+[training]
+batch_size = 32
+
+# Batch size for the OFFLINE SERIALIZATION pass, which is a completely different
+# memory regime from training: it pushes full-resolution images through the 7B
+# VLM, whereas [training].batch_size only moves cached 768-d vectors through a
+# small MLP. Measured on the 23.6 GiB 4090 -- static weights 16.2 GiB, so only
+# ~7.4 GiB is left: batch 8 peaks at 17.7 GiB, 16 at 19.2 GiB, and 32 OOMs.
+# 8 leaves headroom for a labmate's job to coexist; raise to 16 on a free card.
+serialization_batch_size = 8
+learning_rate = 5e-5
+epochs = 50
+train_ratio = 0.8  # 80% for training
+val_ratio = 0.2    # 20% for validation
+tensorboard_log_dir = "<local>/dms_dinov2_clip"
+
+[inference]
+checkpoint_path = "<local>/best_model_dinov2_clip.pth"
+
+[logging]
+wandb_project = "material-classification"
+wandb_run_name = "dms-dinov2-clip-grounded"

DMS/dinov2_clip_gated/head.json

@@ -0,0 +1,99 @@
+{
+  "fusion": "gated",
+  "head": "Gated Multimodal Unit",
+  "head_kwargs": {},
+  "vision_dim": 768,
+  "language_dim": 768,
+  "num_classes": 57,
+  "classes": [
+    "No label",
+    "Animal skin",
+    "Bone/teeth/horn",
+    "Brickwork",
+    "Cardboard",
+    "Carpet/rug",
+    "Ceiling tile",
+    "Ceramic",
+    "Chalkboard/blackboard",
+    "Clutter",
+    "Concrete",
+    "Cork/corkboard",
+    "Engineered stone",
+    "Fabric/cloth",
+    "Fiberglass wool",
+    "Fire",
+    "Foliage",
+    "Food",
+    "Fur",
+    "Gemstone/quartz",
+    "Glass",
+    "Hair",
+    "I cannot tell",
+    "Ice",
+    "Leather",
+    "Liquid, non-water",
+    "Metal",
+    "Mirror",
+    "Not on list",
+    "Paint/plaster/enamel",
+    "Paper",
+    "Pearl",
+    "Photograph/painting",
+    "Plastic, clear",
+    "Plastic, non-clear",
+    "Rubber/latex",
+    "Sand",
+    "Skin/lips",
+    "Sky",
+    "Snow",
+    "Soap",
+    "Soil/mud",
+    "Sponge",
+    "Stone, natural",
+    "Stone, polished",
+    "Styrofoam",
+    "Tile",
+    "Wallpaper",
+    "Water",
+    "Wax",
+    "Whiteboard",
+    "Wicker",
+    "Wood",
+    "Wood, tree",
+    "Bad polygon",
+    "Multiple materials",
+    "Asphalt"
+  ],
+  "dataset": "Apple Dense Material Segmentation",
+  "dataset_key": "DMS",
+  "encoders": {
+    "vision": "DINOv2 ViT-B/14 (CLS token, frozen)",
+    "language": "CLIP ViT-L/14 text encoder over Qwen2.5-VL-7B captions (frozen)",
+    "short": "DINOv2-CLIP",
+    "slug": "dinov2_clip"
+  },
+  "parameters": 1722937,
+  "training": {
+    "learning_rate": 5e-05,
+    "epochs": 50,
+    "batch_size": 32,
+    "weight_decay": 0.0001,
+    "lr_schedule": "constant",
+    "early_stopping_patience": 0,
+    "checkpoint": "best validation loss"
+  },
+  "sha256": {
+    "safetensors": "0715e31f97766d6da1185b402bf47e44a44db6f0bcb24f7e9cb743b30232200c",
+    "pth": "045aaa0553e384bca097835a3689c8aedeef15118c1c6f69c5b38d1d04dfb20f"
+  },
+  "source": {
+    "config": "config/config_dinov2_clip.toml",
+    "git_commit": "1a8130c922267fb0786a33e3854a7e00763cc38f+dirty",
+    "checkpoint_mtime": "2026-09-23T13:35:51"
+  },
+  "metrics": {
+    "mIoU": 0.45433338700912007,
+    "mAcc": 0.6372685795566289,
+    "eval_samples": 186231
+  }
+}

DMS/dinov2_clip_gated/metrics.json

@@ -0,0 +1,296 @@
+{
+  "dataset": "DMS",
+  "config": "config/config_dinov2_clip.toml",
+  "checkpoint": "<local>/best_model_dinov2_clip.pth",
+  "fusion_type": "gated",
+  "num_eval_samples": 186231,
+  "mIoU": 0.45433338700912007,
+  "mAcc": 0.6372685795566289,
+  "per_class": {
+    "No label": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Animal skin": {
+      "accuracy": 0.5964912280701754,
+      "iou": 0.4657534246575342,
+      "support": 114
+    },
+    "Bone/teeth/horn": {
+      "accuracy": 0.7954545454545454,
+      "iou": 0.603448275862069,
+      "support": 44
+    },
+    "Brickwork": {
+      "accuracy": 0.775692582663092,
+      "iou": 0.4849162011173184,
+      "support": 1119
+    },
+    "Cardboard": {
+      "accuracy": 0.7605083088954057,
+      "iou": 0.44431753283837805,
+      "support": 1023
+    },
+    "Carpet/rug": {
+      "accuracy": 0.7795339412360689,
+      "iou": 0.46450132818159867,
+      "support": 4935
+    },
+    "Ceiling tile": {
+      "accuracy": 0.7887323943661971,
+      "iou": 0.4783404514948139,
+      "support": 994
+    },
+    "Ceramic": {
+      "accuracy": 0.8392311052861512,
+      "iou": 0.5887220349371743,
+      "support": 2289
+    },
+    "Chalkboard/blackboard": {
+      "accuracy": 0.7023255813953488,
+      "iou": 0.5171232876712328,
+      "support": 215
+    },
+    "Clutter": {
+      "accuracy": 0.39622641509433965,
+      "iou": 0.2692307692307692,
+      "support": 53
+    },
+    "Concrete": {
+      "accuracy": 0.496372430471584,
+      "iou": 0.27467380394780866,
+      "support": 1654
+    },
+    "Cork/corkboard": {
+      "accuracy": 0.4117647058823529,
+      "iou": 0.2441860465116279,
+      "support": 51
+    },
+    "Engineered stone": {
+      "accuracy": 0.2247191011235955,
+      "iou": 0.13986013986013987,
+      "support": 89
+    },
+    "Fabric/cloth": {
+      "accuracy": 0.7702929239456503,
+      "iou": 0.702627660858718,
+      "support": 22668
+    },
+    "Fiberglass wool": {
+      "accuracy": 0.55,
+      "iou": 0.3055555555555556,
+      "support": 20
+    },
+    "Fire": {
+      "accuracy": 0.78,
+      "iou": 0.65,
+      "support": 50
+    },
+    "Foliage": {
+      "accuracy": 0.9052897705173084,
+      "iou": 0.7798626235550343,
+      "support": 5142
+    },
+    "Food": {
+      "accuracy": 0.8737430167597765,
+      "iou": 0.6085603112840466,
+      "support": 895
+    },
+    "Fur": {
+      "accuracy": 0.8283062645011601,
+      "iou": 0.68,
+      "support": 431
+    },
+    "Gemstone/quartz": {
+      "accuracy": 0.72,
+      "iou": 0.2926829268292683,
+      "support": 50
+    },
+    "Glass": {
+      "accuracy": 0.760976692563818,
+      "iou": 0.6663686195233838,
+      "support": 22525
+    },
+    "Hair": {
+      "accuracy": 0.9763929618768329,
+      "iou": 0.9157040704070407,
+      "support": 6820
+    },
+    "I cannot tell": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Ice": {
+      "accuracy": 0.6363636363636364,
+      "iou": 0.44871794871794873,
+      "support": 55
+    },
+    "Leather": {
+      "accuracy": 0.7150890346766635,
+      "iou": 0.42721164613661816,
+      "support": 2134
+    },
+    "Liquid, non-water": {
+      "accuracy": 0.4444444444444444,
+      "iou": 0.3404255319148936,
+      "support": 36
+    },
+    "Metal": {
+      "accuracy": 0.7975601714474118,
+      "iou": 0.5638256545722943,
+      "support": 9099
+    },
+    "Mirror": {
+      "accuracy": 0.6700125470514429,
+      "iou": 0.3729050279329609,
+      "support": 797
+    },
+    "Not on list": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Paint/plaster/enamel": {
+      "accuracy": 0.5913390528267496,
+      "iou": 0.5186604189636164,
+      "support": 31821
+    },
+    "Paper": {
+      "accuracy": 0.8254024411816735,
+      "iou": 0.6823632641123135,
+      "support": 11306
+    },
+    "Pearl": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 1
+    },
+    "Photograph/painting": {
+      "accuracy": 0.7118034751348112,
+      "iou": 0.3496174220129488,
+      "support": 1669
+    },
+    "Plastic, clear": {
+      "accuracy": 0.6816199376947041,
+      "iou": 0.43830128205128205,
+      "support": 1605
+    },
+    "Plastic, non-clear": {
+      "accuracy": 0.6442321302193914,
+      "iou": 0.5032896555537126,
+      "support": 14130
+    },
+    "Rubber/latex": {
+      "accuracy": 0.6570363466915191,
+      "iou": 0.34474327628361856,
+      "support": 1073
+    },
+    "Sand": {
+      "accuracy": 0.42045454545454547,
+      "iou": 0.2890625,
+      "support": 88
+    },
+    "Skin/lips": {
+      "accuracy": 0.9697943988493104,
+      "iou": 0.9154952076677316,
+      "support": 11819
+    },
+    "Sky": {
+      "accuracy": 0.9127205199628597,
+      "iou": 0.6920098556846181,
+      "support": 2154
+    },
+    "Snow": {
+      "accuracy": 0.5963302752293578,
+      "iou": 0.40372670807453415,
+      "support": 109
+    },
+    "Soap": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 7
+    },
+    "Soil/mud": {
+      "accuracy": 0.624048706240487,
+      "iou": 0.4257528556593977,
+      "support": 657
+    },
+    "Sponge": {
+      "accuracy": 0.16666666666666666,
+      "iou": 0.125,
+      "support": 18
+    },
+    "Stone, natural": {
+      "accuracy": 0.6322686322686323,
+      "iou": 0.4419004006868918,
+      "support": 1221
+    },
+    "Stone, polished": {
+      "accuracy": 0.39580602883355176,
+      "iou": 0.252297410192147,
+      "support": 763
+    },
+    "Styrofoam": {
+      "accuracy": 0.16129032258064516,
+      "iou": 0.13513513513513514,
+      "support": 31
+    },
+    "Tile": {
+      "accuracy": 0.645155679248565,
+      "iou": 0.42412807318467693,
+      "support": 5749
+    },
+    "Wallpaper": {
+      "accuracy": 0.5654761904761905,
+      "iou": 0.3445149592021759,
+      "support": 672
+    },
+    "Water": {
+      "accuracy": 0.6651532349602725,
+      "iou": 0.5195035460992907,
+      "support": 881
+    },
+    "Wax": {
+      "accuracy": 0.8245614035087719,
+      "iou": 0.6861313868613139,
+      "support": 114
+    },
+    "Whiteboard": {
+      "accuracy": 0.6486486486486487,
+      "iou": 0.46652267818574517,
+      "support": 333
+    },
+    "Wicker": {
+      "accuracy": 0.7702479338842976,
+      "iou": 0.548235294117647,
+      "support": 605
+    },
+    "Wood": {
+      "accuracy": 0.7891884833844748,
+      "iou": 0.6155726434791465,
+      "support": 15317
+    },
+    "Wood, tree": {
+      "accuracy": 0.7431972789115646,
+      "iou": 0.4765539803707743,
+      "support": 588
+    },
+    "Bad polygon": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Multiple materials": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Asphalt": {
+      "accuracy": 0.5,
+      "iou": 0.2972972972972973,
+      "support": 198
+    }
+  }
+}

DMS/dinov2_clip_gated/model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:045aaa0553e384bca097835a3689c8aedeef15118c1c6f69c5b38d1d04dfb20f
+size 6910974

DMS/dinov2_clip_gated/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0715e31f97766d6da1185b402bf47e44a44db6f0bcb24f7e9cb743b30232200c
+size 6904172

DMS/dinov2_clip_mlp/README.md

@@ -0,0 +1,52 @@
+# DINOv2-CLIP + Concat + MLP on Apple Dense Material Segmentation
+
+A material-classification **fusion head** from SegObjectMaterial. It classifies the
+material of a masked object from two frozen embeddings:
+
+| | |
+|---|---|
+| vision | DINOv2 ViT-B/14 (CLS token, frozen) -> 768-d |
+| language | CLIP ViT-L/14 text encoder over Qwen2.5-VL-7B captions (frozen) -> 768-d |
+| head | Concat + MLP (`mlp`, defaults), 934,457 parameters |
+| classes | 57 (Apple Dense Material Segmentation) |
+
+**This repo holds only the head.** Features must come from the same encoders,
+captioner and prompts as training (see the SegObjectMaterial repository); the
+encoders are downloaded from their own model pages under their own licenses.
+
+## Evaluation
+
+Held-out split recorded in `split.json` (source images never seen in training).
+Single training run -- differences under ~0.5 mIoU points between heads are within
+run-to-run noise.
+
+| mIoU | mAcc | eval samples |
+|---|---|---|
+| 0.4744 | 0.6672 | 186,231 |
+
+## Training
+
+```json
+{
+  "learning_rate": 5e-05,
+  "epochs": 50,
+  "batch_size": 32,
+  "weight_decay": 0.0001,
+  "lr_schedule": "constant",
+  "early_stopping_patience": 0,
+  "checkpoint": "best validation loss"
+}
+```
+
+## Use
+
+```python
+from utils.hub import load_head_from_hub      # SegObjectMaterial repository
+head, info = load_head_from_hub("s0um0r0y/SegObjectMaterial", "DMS/dinov2_clip_mlp")
+logits = head(vision_features, language_features)   # [B, 768], [B, 768]
+```
+
+## Integrity
+
+- `model.safetensors` sha256 `b91f86016abb3734031738c48537d2ed499d7c0bbbaf90ea8a8fc4219f649c9e`
+- source config `config/config_dinov2_clip_mlp.toml`, code commit `1a8130c922267fb0786a33e3854a7e00763cc38f+dirty`

DMS/dinov2_clip_mlp/config.toml

@@ -0,0 +1,68 @@
+# DINOv2-CLIP + concat-MLP head: the ORIGINAL-PAPER fusion, retrained on the
+# current grounded cache. The old MLP checkpoints (best_model.pth / _v2) were
+# trained on a cache that no longer exists, with no held-out split recorded, so
+# they cannot be evaluated fairly. Same cache as config_dinov2_clip.toml;
+# own checkpoint, TensorBoard dir and W&B run, so nothing is overwritten.
+# Baseline recipe (wd 1e-4, constant LR, 50 epochs) to match the other baseline rows.
+
+# Apple DMS -- BASE pipeline: DINOv2 ViT-B/14 (768) + Qwen2.5-VL -> CLIP text (768)
+#
+# Pair this with:
+#   src/vision_stream.py + src/language_stream.py
+#   run_serialization.py  ->  train.py  (the raw/original scripts)
+#
+# config/config.toml is the OTHER pipeline (DINOv3 ViT-L/16 1024 + SigLIP2),
+# driven by run_serialization_latest_models.py -> train_latest_models.py.
+# Mixing them raises the dimension-mismatch ValueError in the fusion head.
+
+[data]
+dataset_type = "apple_dms"
+image_dir = "<local>/images"
+mask_dir = "<local>/labels"
+# Rebuilt automatically on first run; the index now stores one entry per
+# CONNECTED COMPONENT, so any pre-existing labels.json must be removed first.
+labels_file = "<local>/labels_instances.json"
+# Fresh directory: features are now object-grounded (bbox-cropped vision,
+# bbox-prompted captions), so the old tensor_cache_dir/ contents are stale.
+tensor_cache_dir = "<local>/tensor_cache_dir_grounded"
+
+# Connected-component filtering (measured on 400 real DMS masks):
+# keeping every component gives 7.5x more samples than the old per-(image,class)
+# scheme -- 2.45M samples / 47 GB / ~7.5x the VLM captioning time -- because 69%
+# of components are under 0.1% of the frame and crop to a few pixels.
+# A component is kept when area >= max(min_instance_area, min_instance_frac * frame).
+min_instance_area = 1000        # absolute pixel floor
+min_instance_frac = 0.001       # and at least 0.1% of the frame
+max_instances_per_class = 0     # 0 = keep all; set e.g. 3 to cap the long tail
+
+
+[model]
+# fusion head: read by train_latest_models.py. Without this key it defaults to
+# "gated" -- which is how best_model_dinov2_clip.pth / best_model_fmd.pth were trained.
+fusion = "mlp"
+vision_dim = 768      # DINOv2 ViT-B/14 CLS token
+language_dim = 768    # CLIP ViT-L/14 text pooler_output
+num_classes = 57
+
+[training]
+batch_size = 32
+
+# Batch size for the OFFLINE SERIALIZATION pass, which is a completely different
+# memory regime from training: it pushes full-resolution images through the 7B
+# VLM, whereas [training].batch_size only moves cached 768-d vectors through a
+# small MLP. Measured on the 23.6 GiB 4090 -- static weights 16.2 GiB, so only
+# ~7.4 GiB is left: batch 8 peaks at 17.7 GiB, 16 at 19.2 GiB, and 32 OOMs.
+# 8 leaves headroom for a labmate's job to coexist; raise to 16 on a free card.
+serialization_batch_size = 8
+learning_rate = 5e-5
+epochs = 50
+train_ratio = 0.8  # 80% for training
+val_ratio = 0.2    # 20% for validation
+tensorboard_log_dir = "<local>/dms_dinov2_clip_mlp"
+
+[inference]
+checkpoint_path = "<local>/best_model_dms_dinov2_clip_mlp.pth"
+
+[logging]
+wandb_project = "material-classification"
+wandb_run_name = "dms-dinov2-clip-mlp"

DMS/dinov2_clip_mlp/head.json

@@ -0,0 +1,99 @@
+{
+  "fusion": "mlp",
+  "head": "Concat + MLP",
+  "head_kwargs": {},
+  "vision_dim": 768,
+  "language_dim": 768,
+  "num_classes": 57,
+  "classes": [
+    "No label",
+    "Animal skin",
+    "Bone/teeth/horn",
+    "Brickwork",
+    "Cardboard",
+    "Carpet/rug",
+    "Ceiling tile",
+    "Ceramic",
+    "Chalkboard/blackboard",
+    "Clutter",
+    "Concrete",
+    "Cork/corkboard",
+    "Engineered stone",
+    "Fabric/cloth",
+    "Fiberglass wool",
+    "Fire",
+    "Foliage",
+    "Food",
+    "Fur",
+    "Gemstone/quartz",
+    "Glass",
+    "Hair",
+    "I cannot tell",
+    "Ice",
+    "Leather",
+    "Liquid, non-water",
+    "Metal",
+    "Mirror",
+    "Not on list",
+    "Paint/plaster/enamel",
+    "Paper",
+    "Pearl",
+    "Photograph/painting",
+    "Plastic, clear",
+    "Plastic, non-clear",
+    "Rubber/latex",
+    "Sand",
+    "Skin/lips",
+    "Sky",
+    "Snow",
+    "Soap",
+    "Soil/mud",
+    "Sponge",
+    "Stone, natural",
+    "Stone, polished",
+    "Styrofoam",
+    "Tile",
+    "Wallpaper",
+    "Water",
+    "Wax",
+    "Whiteboard",
+    "Wicker",
+    "Wood",
+    "Wood, tree",
+    "Bad polygon",
+    "Multiple materials",
+    "Asphalt"
+  ],
+  "dataset": "Apple Dense Material Segmentation",
+  "dataset_key": "DMS",
+  "encoders": {
+    "vision": "DINOv2 ViT-B/14 (CLS token, frozen)",
+    "language": "CLIP ViT-L/14 text encoder over Qwen2.5-VL-7B captions (frozen)",
+    "short": "DINOv2-CLIP",
+    "slug": "dinov2_clip"
+  },
+  "parameters": 934457,
+  "training": {
+    "learning_rate": 5e-05,
+    "epochs": 50,
+    "batch_size": 32,
+    "weight_decay": 0.0001,
+    "lr_schedule": "constant",
+    "early_stopping_patience": 0,
+    "checkpoint": "best validation loss"
+  },
+  "sha256": {
+    "safetensors": "b91f86016abb3734031738c48537d2ed499d7c0bbbaf90ea8a8fc4219f649c9e",
+    "pth": "0fe9af49a1f493b61a55aed60cef6a809458c6baeea9a5e8d2f15d3e2c4d0ad8"
+  },
+  "source": {
+    "config": "config/config_dinov2_clip_mlp.toml",
+    "git_commit": "1a8130c922267fb0786a33e3854a7e00763cc38f+dirty",
+    "checkpoint_mtime": "2026-10-09T00:26:13"
+  },
+  "metrics": {
+    "mIoU": 0.4743560437589232,
+    "mAcc": 0.6672396382567987,
+    "eval_samples": 186231
+  }
+}

DMS/dinov2_clip_mlp/metrics.json

@@ -0,0 +1,296 @@
+{
+  "dataset": "DMS",
+  "config": "config/config_dinov2_clip_mlp.toml",
+  "checkpoint": "<local>/best_model_dms_dinov2_clip_mlp.pth",
+  "fusion_type": "mlp",
+  "num_eval_samples": 186231,
+  "mIoU": 0.4743560437589232,
+  "mAcc": 0.6672396382567987,
+  "per_class": {
+    "No label": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Animal skin": {
+      "accuracy": 0.6228070175438597,
+      "iou": 0.47019867549668876,
+      "support": 114
+    },
+    "Bone/teeth/horn": {
+      "accuracy": 0.7954545454545454,
+      "iou": 0.625,
+      "support": 44
+    },
+    "Brickwork": {
+      "accuracy": 0.7587131367292225,
+      "iou": 0.5322884012539185,
+      "support": 1119
+    },
+    "Cardboard": {
+      "accuracy": 0.7223851417399805,
+      "iou": 0.4737179487179487,
+      "support": 1023
+    },
+    "Carpet/rug": {
+      "accuracy": 0.8028368794326242,
+      "iou": 0.4522314804246091,
+      "support": 4935
+    },
+    "Ceiling tile": {
+      "accuracy": 0.7937625754527163,
+      "iou": 0.499683343888537,
+      "support": 994
+    },
+    "Ceramic": {
+      "accuracy": 0.8217562254259502,
+      "iou": 0.6011505273250239,
+      "support": 2289
+    },
+    "Chalkboard/blackboard": {
+      "accuracy": 0.7209302325581395,
+      "iou": 0.49363057324840764,
+      "support": 215
+    },
+    "Clutter": {
+      "accuracy": 0.4528301886792453,
+      "iou": 0.32432432432432434,
+      "support": 53
+    },
+    "Concrete": {
+      "accuracy": 0.56590084643289,
+      "iou": 0.2826940501359106,
+      "support": 1654
+    },
+    "Cork/corkboard": {
+      "accuracy": 0.35294117647058826,
+      "iou": 0.21951219512195122,
+      "support": 51
+    },
+    "Engineered stone": {
+      "accuracy": 0.21348314606741572,
+      "iou": 0.1484375,
+      "support": 89
+    },
+    "Fabric/cloth": {
+      "accuracy": 0.765528498323628,
+      "iou": 0.700876448968052,
+      "support": 22668
+    },
+    "Fiberglass wool": {
+      "accuracy": 0.55,
+      "iou": 0.2558139534883721,
+      "support": 20
+    },
+    "Fire": {
+      "accuracy": 0.82,
+      "iou": 0.6307692307692307,
+      "support": 50
+    },
+    "Foliage": {
+      "accuracy": 0.9142357059509918,
+      "iou": 0.7736998025016458,
+      "support": 5142
+    },
+    "Food": {
+      "accuracy": 0.8737430167597765,
+      "iou": 0.6201427438540841,
+      "support": 895
+    },
+    "Fur": {
+      "accuracy": 0.8213457076566125,
+      "iou": 0.6941176470588235,
+      "support": 431
+    },
+    "Gemstone/quartz": {
+      "accuracy": 0.64,
+      "iou": 0.24806201550387597,
+      "support": 50
+    },
+    "Glass": {
+      "accuracy": 0.7659045504994451,
+      "iou": 0.667672897557955,
+      "support": 22525
+    },
+    "Hair": {
+      "accuracy": 0.9778592375366569,
+      "iou": 0.9158198297171107,
+      "support": 6820
+    },
+    "I cannot tell": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Ice": {
+      "accuracy": 0.6545454545454545,
+      "iou": 0.4931506849315068,
+      "support": 55
+    },
+    "Leather": {
+      "accuracy": 0.7146204311152765,
+      "iou": 0.43348493462194426,
+      "support": 2134
+    },
+    "Liquid, non-water": {
+      "accuracy": 0.5833333333333334,
+      "iou": 0.42,
+      "support": 36
+    },
+    "Metal": {
+      "accuracy": 0.8061325420375866,
+      "iou": 0.5545055941941337,
+      "support": 9099
+    },
+    "Mirror": {
+      "accuracy": 0.6750313676286073,
+      "iou": 0.38156028368794326,
+      "support": 797
+    },
+    "Not on list": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Paint/plaster/enamel": {
+      "accuracy": 0.5959586436629899,
+      "iou": 0.5222228341686402,
+      "support": 31821
+    },
+    "Paper": {
+      "accuracy": 0.8380505926056961,
+      "iou": 0.6929715497696189,
+      "support": 11306
+    },
+    "Pearl": {
+      "accuracy": 1.0,
+      "iou": 0.5,
+      "support": 1
+    },
+    "Photograph/painting": {
+      "accuracy": 0.6674655482324745,
+      "iou": 0.36369572314724125,
+      "support": 1669
+    },
+    "Plastic, clear": {
+      "accuracy": 0.6990654205607477,
+      "iou": 0.42275810097965333,
+      "support": 1605
+    },
+    "Plastic, non-clear": {
+      "accuracy": 0.6428167020523708,
+      "iou": 0.5052286127489154,
+      "support": 14130
+    },
+    "Rubber/latex": {
+      "accuracy": 0.6365330848089469,
+      "iou": 0.3490035769034236,
+      "support": 1073
+    },
+    "Sand": {
+      "accuracy": 0.4318181818181818,
+      "iou": 0.2733812949640288,
+      "support": 88
+    },
+    "Skin/lips": {
+      "accuracy": 0.9684406464167865,
+      "iou": 0.9211331079993562,
+      "support": 11819
+    },
+    "Sky": {
+      "accuracy": 0.9103992571959145,
+      "iou": 0.675973802137194,
+      "support": 2154
+    },
+    "Snow": {
+      "accuracy": 0.6055045871559633,
+      "iou": 0.4074074074074074,
+      "support": 109
+    },
+    "Soap": {
+      "accuracy": 0.2857142857142857,
+      "iou": 0.2857142857142857,
+      "support": 7
+    },
+    "Soil/mud": {
+      "accuracy": 0.6529680365296804,
+      "iou": 0.41690962099125367,
+      "support": 657
+    },
+    "Sponge": {
+      "accuracy": 0.2777777777777778,
+      "iou": 0.17857142857142858,
+      "support": 18
+    },
+    "Stone, natural": {
+      "accuracy": 0.5995085995085995,
+      "iou": 0.4566437928883344,
+      "support": 1221
+    },
+    "Stone, polished": {
+      "accuracy": 0.38794233289646135,
+      "iou": 0.26690712353471596,
+      "support": 763
+    },
+    "Styrofoam": {
+      "accuracy": 0.16129032258064516,
+      "iou": 0.1388888888888889,
+      "support": 31
+    },
+    "Tile": {
+      "accuracy": 0.6155853191859454,
+      "iou": 0.4333292518672707,
+      "support": 5749
+    },
+    "Wallpaper": {
+      "accuracy": 0.5818452380952381,
+      "iou": 0.3010007698229407,
+      "support": 672
+    },
+    "Water": {
+      "accuracy": 0.6867196367763905,
+      "iou": 0.4954954954954955,
+      "support": 881
+    },
+    "Wax": {
+      "accuracy": 0.8421052631578947,
+      "iou": 0.7007299270072993,
+      "support": 114
+    },
+    "Whiteboard": {
+      "accuracy": 0.6576576576576577,
+      "iou": 0.4524793388429752,
+      "support": 333
+    },
+    "Wicker": {
+      "accuracy": 0.7735537190082644,
+      "iou": 0.572827417380661,
+      "support": 605
+    },
+    "Wood": {
+      "accuracy": 0.7940197166546974,
+      "iou": 0.6218109310291937,
+      "support": 15317
+    },
+    "Wood, tree": {
+      "accuracy": 0.7329931972789115,
+      "iou": 0.5149342891278376,
+      "support": 588
+    },
+    "Bad polygon": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Multiple materials": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Asphalt": {
+      "accuracy": 0.46464646464646464,
+      "iou": 0.2839506172839506,
+      "support": 198
+    }
+  }
+}

DMS/dinov2_clip_mlp/model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fe9af49a1f493b61a55aed60cef6a809458c6baeea9a5e8d2f15d3e2c4d0ad8
+size 3750066

DMS/dinov2_clip_mlp/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b91f86016abb3734031738c48537d2ed499d7c0bbbaf90ea8a8fc4219f649c9e
+size 3745260

DMS/dinov3_siglip2_gated/README.md

@@ -0,0 +1,52 @@
+# DINOv3-SigLIP2 + Gated Multimodal Unit on Apple Dense Material Segmentation
+
+A material-classification **fusion head** from SegObjectMaterial. It classifies the
+material of a masked object from two frozen embeddings:
+
+| | |
+|---|---|
+| vision | DINOv3 ViT-L/16 (CLS token, frozen) -> 1024-d |
+| language | SigLIP2 base text encoder over Qwen3-VL-8B captions (frozen) -> 768-d |
+| head | Gated Multimodal Unit (`gated`, defaults), 1,985,081 parameters |
+| classes | 57 (Apple Dense Material Segmentation) |
+
+**This repo holds only the head.** Features must come from the same encoders,
+captioner and prompts as training (see the SegObjectMaterial repository); the
+encoders are downloaded from their own model pages under their own licenses.
+
+## Evaluation
+
+Held-out split recorded in `split.json` (source images never seen in training).
+Single training run -- differences under ~0.5 mIoU points between heads are within
+run-to-run noise.
+
+| mIoU | mAcc | eval samples |
+|---|---|---|
+| 0.4630 | 0.6386 | 186,231 |
+
+## Training
+
+```json
+{
+  "learning_rate": 5e-05,
+  "epochs": 50,
+  "batch_size": 32,
+  "weight_decay": 0.0001,
+  "lr_schedule": "constant",
+  "early_stopping_patience": 0,
+  "checkpoint": "best validation loss"
+}
+```
+
+## Use
+
+```python
+from utils.hub import load_head_from_hub      # SegObjectMaterial repository
+head, info = load_head_from_hub("s0um0r0y/SegObjectMaterial", "DMS/dinov3_siglip2_gated")
+logits = head(vision_features, language_features)   # [B, 1024], [B, 768]
+```
+
+## Integrity
+
+- `model.safetensors` sha256 `8c71607a87521793b66f1fc13ba984fe2b6954774b0caad559ae94d72de33b12`
+- source config `config/config.toml`, code commit `1a8130c922267fb0786a33e3854a7e00763cc38f+dirty`

DMS/dinov3_siglip2_gated/config.toml

@@ -0,0 +1,59 @@
+# Apple DMS -- LATEST-MODELS pipeline: DINOv3 ViT-L/16 (1024) + Qwen3-VL -> SigLIP2 text (768)
+#
+# Pair this with:
+#   src/vision_stream_latest_models.py + src/language_stream_latest_models.py
+#   run_serialization_latest_models.py  ->  train_latest_models.py
+#
+# For the base DINOv2/CLIP pipeline (run_serialization.py -> train.py) use
+# config/config_dinov2_clip.toml instead -- its vision_dim is 768, not 1024.
+
+[data]
+dataset_type = "apple_dms"
+image_dir = "<local>/images"
+mask_dir = "<local>/labels"
+# Rebuilt automatically on first run; the index now stores one entry per
+# CONNECTED COMPONENT, so any pre-existing labels.json must be removed first.
+labels_file = "<local>/labels_instances.json"
+# Separate cache: DINOv3 features are 1024-d and must not mix with the
+# 768-d DINOv2 cache written by run_serialization.py.
+tensor_cache_dir = "<local>/tensor_cache_dir_latest_models"
+
+# Connected-component filtering (measured on 400 real DMS masks):
+# keeping every component gives 7.5x more samples than the old per-(image,class)
+# scheme -- 2.45M samples / 47 GB / ~7.5x the VLM captioning time -- because 69%
+# of components are under 0.1% of the frame and crop to a few pixels.
+# A component is kept when area >= max(min_instance_area, min_instance_frac * frame).
+min_instance_area = 1000        # absolute pixel floor
+min_instance_frac = 0.001       # and at least 0.1% of the frame
+max_instances_per_class = 0     # 0 = keep all; set e.g. 3 to cap the long tail
+
+
+[model]
+# fusion head: "gated" (GMU) or "mlp" (concat+MLP). Read by train_latest_models.py; a contradicting --fusion flag is an error, so checkpoint_path can't hold the wrong head.
+fusion = "gated"
+vision_dim = 1024     # DINOv3 ViT-L/16 CLS token
+language_dim = 768    # SigLIP2 base text embedding
+num_classes = 57
+
+[training]
+batch_size = 32
+
+# Batch size for the OFFLINE SERIALIZATION pass, which is a completely different
+# memory regime from training: it pushes full-resolution images through the 7B
+# VLM, whereas [training].batch_size only moves cached 768-d vectors through a
+# small MLP. Measured on the 23.6 GiB 4090 -- static weights 16.2 GiB, so only
+# ~7.4 GiB is left: batch 8 peaks at 17.7 GiB, 16 at 19.2 GiB, and 32 OOMs.
+# 8 leaves headroom for a labmate's job to coexist; raise to 16 on a free card.
+serialization_batch_size = 8
+learning_rate = 5e-5
+epochs = 50
+train_ratio = 0.8  # 80% for training
+val_ratio = 0.2    # 20% for validation
+tensorboard_log_dir = "<local>/dms_dinov3_siglip2"
+
+[inference]
+checkpoint_path = "<local>/best_model_SigLIP_DINOv3.pth"
+
+[logging]
+wandb_project = "material-classification"
+wandb_run_name = "dinov3-siglip2"

DMS/dinov3_siglip2_gated/head.json

@@ -0,0 +1,99 @@
+{
+  "fusion": "gated",
+  "head": "Gated Multimodal Unit",
+  "head_kwargs": {},
+  "vision_dim": 1024,
+  "language_dim": 768,
+  "num_classes": 57,
+  "classes": [
+    "No label",
+    "Animal skin",
+    "Bone/teeth/horn",
+    "Brickwork",
+    "Cardboard",
+    "Carpet/rug",
+    "Ceiling tile",
+    "Ceramic",
+    "Chalkboard/blackboard",
+    "Clutter",
+    "Concrete",
+    "Cork/corkboard",
+    "Engineered stone",
+    "Fabric/cloth",
+    "Fiberglass wool",
+    "Fire",
+    "Foliage",
+    "Food",
+    "Fur",
+    "Gemstone/quartz",
+    "Glass",
+    "Hair",
+    "I cannot tell",
+    "Ice",
+    "Leather",
+    "Liquid, non-water",
+    "Metal",
+    "Mirror",
+    "Not on list",
+    "Paint/plaster/enamel",
+    "Paper",
+    "Pearl",
+    "Photograph/painting",
+    "Plastic, clear",
+    "Plastic, non-clear",
+    "Rubber/latex",
+    "Sand",
+    "Skin/lips",
+    "Sky",
+    "Snow",
+    "Soap",
+    "Soil/mud",
+    "Sponge",
+    "Stone, natural",
+    "Stone, polished",
+    "Styrofoam",
+    "Tile",
+    "Wallpaper",
+    "Water",
+    "Wax",
+    "Whiteboard",
+    "Wicker",
+    "Wood",
+    "Wood, tree",
+    "Bad polygon",
+    "Multiple materials",
+    "Asphalt"
+  ],
+  "dataset": "Apple Dense Material Segmentation",
+  "dataset_key": "DMS",
+  "encoders": {
+    "vision": "DINOv3 ViT-L/16 (CLS token, frozen)",
+    "language": "SigLIP2 base text encoder over Qwen3-VL-8B captions (frozen)",
+    "short": "DINOv3-SigLIP2",
+    "slug": "dinov3_siglip2"
+  },
+  "parameters": 1985081,
+  "training": {
+    "learning_rate": 5e-05,
+    "epochs": 50,
+    "batch_size": 32,
+    "weight_decay": 0.0001,
+    "lr_schedule": "constant",
+    "early_stopping_patience": 0,
+    "checkpoint": "best validation loss"
+  },
+  "sha256": {
+    "safetensors": "8c71607a87521793b66f1fc13ba984fe2b6954774b0caad559ae94d72de33b12",
+    "pth": "5751d9a8428036b671c11a9fc440931ad0888d497feb7b2a0f93271b4f18b293"
+  },
+  "source": {
+    "config": "config/config.toml",
+    "git_commit": "1a8130c922267fb0786a33e3854a7e00763cc38f+dirty",
+    "checkpoint_mtime": "2026-09-27T16:03:35"
+  },
+  "metrics": {
+    "mIoU": 0.4630038334240617,
+    "mAcc": 0.638554054441271,
+    "eval_samples": 186231
+  }
+}

DMS/dinov3_siglip2_gated/metrics.json

@@ -0,0 +1,296 @@
+{
+  "dataset": "DMS",
+  "checkpoint": "<local>/best_model_SigLIP_DINOv3.pth",
+  "fusion_type": "gated",
+  "num_eval_samples": 186231,
+  "mIoU": 0.4630038334240617,
+  "mAcc": 0.638554054441271,
+  "per_class": {
+    "No label": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Animal skin": {
+      "accuracy": 0.6052631578947368,
+      "iou": 0.4107142857142857,
+      "support": 114
+    },
+    "Bone/teeth/horn": {
+      "accuracy": 0.7272727272727273,
+      "iou": 0.64,
+      "support": 44
+    },
+    "Brickwork": {
+      "accuracy": 0.7667560321715817,
+      "iou": 0.516245487364621,
+      "support": 1119
+    },
+    "Cardboard": {
+      "accuracy": 0.7526881720430108,
+      "iou": 0.48951048951048953,
+      "support": 1023
+    },
+    "Carpet/rug": {
+      "accuracy": 0.7797365754812563,
+      "iou": 0.49056603773584906,
+      "support": 4935
+    },
+    "Ceiling tile": {
+      "accuracy": 0.772635814889336,
+      "iou": 0.5140562248995983,
+      "support": 994
+    },
+    "Ceramic": {
+      "accuracy": 0.8396679772826562,
+      "iou": 0.5924784217016029,
+      "support": 2289
+    },
+    "Chalkboard/blackboard": {
+      "accuracy": 0.7534883720930232,
+      "iou": 0.5472972972972973,
+      "support": 215
+    },
+    "Clutter": {
+      "accuracy": 0.5094339622641509,
+      "iou": 0.40298507462686567,
+      "support": 53
+    },
+    "Concrete": {
+      "accuracy": 0.5126964933494559,
+      "iou": 0.2870683818551117,
+      "support": 1654
+    },
+    "Cork/corkboard": {
+      "accuracy": 0.3333333333333333,
+      "iou": 0.2125,
+      "support": 51
+    },
+    "Engineered stone": {
+      "accuracy": 0.23595505617977527,
+      "iou": 0.13725490196078433,
+      "support": 89
+    },
+    "Fabric/cloth": {
+      "accuracy": 0.7962766896064938,
+      "iou": 0.7181221404416153,
+      "support": 22668
+    },
+    "Fiberglass wool": {
+      "accuracy": 0.25,
+      "iou": 0.1724137931034483,
+      "support": 20
+    },
+    "Fire": {
+      "accuracy": 0.86,
+      "iou": 0.6142857142857143,
+      "support": 50
+    },
+    "Foliage": {
+      "accuracy": 0.9225982108129133,
+      "iou": 0.7788540469545231,
+      "support": 5142
+    },
+    "Food": {
+      "accuracy": 0.870391061452514,
+      "iou": 0.6411522633744856,
+      "support": 895
+    },
+    "Fur": {
+      "accuracy": 0.8190255220417634,
+      "iou": 0.6948818897637795,
+      "support": 431
+    },
+    "Gemstone/quartz": {
+      "accuracy": 0.66,
+      "iou": 0.2578125,
+      "support": 50
+    },
+    "Glass": {
+      "accuracy": 0.7757602663706992,
+      "iou": 0.6716119609501114,
+      "support": 22525
+    },
+    "Hair": {
+      "accuracy": 0.9737536656891496,
+      "iou": 0.9276435256320715,
+      "support": 6820
+    },
+    "I cannot tell": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Ice": {
+      "accuracy": 0.45454545454545453,
+      "iou": 0.3787878787878788,
+      "support": 55
+    },
+    "Leather": {
+      "accuracy": 0.7155576382380506,
+      "iou": 0.44094715564539416,
+      "support": 2134
+    },
+    "Liquid, non-water": {
+      "accuracy": 0.5,
+      "iou": 0.36,
+      "support": 36
+    },
+    "Metal": {
+      "accuracy": 0.8040444004835696,
+      "iou": 0.5635061233921281,
+      "support": 9099
+    },
+    "Mirror": {
+      "accuracy": 0.685069008782936,
+      "iou": 0.3976693372177713,
+      "support": 797
+    },
+    "Not on list": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Paint/plaster/enamel": {
+      "accuracy": 0.6256245875365325,
+      "iou": 0.5428516892536743,
+      "support": 31821
+    },
+    "Paper": {
+      "accuracy": 0.8281443481337343,
+      "iou": 0.6981582283200358,
+      "support": 11306
+    },
+    "Pearl": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 1
+    },
+    "Photograph/painting": {
+      "accuracy": 0.7112043139604554,
+      "iou": 0.3851395197923426,
+      "support": 1669
+    },
+    "Plastic, clear": {
+      "accuracy": 0.6785046728971963,
+      "iou": 0.4403558431055398,
+      "support": 1605
+    },
+    "Plastic, non-clear": {
+      "accuracy": 0.6440905874026893,
+      "iou": 0.5220558710491596,
+      "support": 14130
+    },
+    "Rubber/latex": {
+      "accuracy": 0.5712954333643989,
+      "iou": 0.3383002207505519,
+      "support": 1073
+    },
+    "Sand": {
+      "accuracy": 0.32954545454545453,
+      "iou": 0.25,
+      "support": 88
+    },
+    "Skin/lips": {
+      "accuracy": 0.975886284795668,
+      "iou": 0.9366574630501868,
+      "support": 11819
+    },
+    "Sky": {
+      "accuracy": 0.9308263695450325,
+      "iou": 0.6852358168147642,
+      "support": 2154
+    },
+    "Snow": {
+      "accuracy": 0.7064220183486238,
+      "iou": 0.5,
+      "support": 109
+    },
+    "Soap": {
+      "accuracy": 0.14285714285714285,
+      "iou": 0.14285714285714285,
+      "support": 7
+    },
+    "Soil/mud": {
+      "accuracy": 0.6636225266362252,
+      "iou": 0.41366223908918404,
+      "support": 657
+    },
+    "Sponge": {
+      "accuracy": 0.2777777777777778,
+      "iou": 0.19230769230769232,
+      "support": 18
+    },
+    "Stone, natural": {
+      "accuracy": 0.6101556101556102,
+      "iou": 0.4482551143200963,
+      "support": 1221
+    },
+    "Stone, polished": {
+      "accuracy": 0.41153342070773263,
+      "iou": 0.26364399664147775,
+      "support": 763
+    },
+    "Styrofoam": {
+      "accuracy": 0.16129032258064516,
+      "iou": 0.11627906976744186,
+      "support": 31
+    },
+    "Tile": {
+      "accuracy": 0.6489824317272569,
+      "iou": 0.4471476510067114,
+      "support": 5749
+    },
+    "Wallpaper": {
+      "accuracy": 0.59375,
+      "iou": 0.328125,
+      "support": 672
+    },
+    "Water": {
+      "accuracy": 0.6821793416572077,
+      "iou": 0.5207972270363952,
+      "support": 881
+    },
+    "Wax": {
+      "accuracy": 0.868421052631579,
+      "iou": 0.5789473684210527,
+      "support": 114
+    },
+    "Whiteboard": {
+      "accuracy": 0.6606606606606606,
+      "iou": 0.46218487394957986,
+      "support": 333
+    },
+    "Wicker": {
+      "accuracy": 0.7851239669421488,
+      "iou": 0.5771567436208992,
+      "support": 605
+    },
+    "Wood": {
+      "accuracy": 0.7963700463537247,
+      "iou": 0.6331689592525305,
+      "support": 15317
+    },
+    "Wood, tree": {
+      "accuracy": 0.7346938775510204,
+      "iou": 0.48758465011286684,
+      "support": 588
+    },
+    "Bad polygon": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Multiple materials": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Asphalt": {
+      "accuracy": 0.4898989898989899,
+      "iou": 0.3069620253164557,
+      "support": 198
+    }
+  },
+  "config": "config/config.toml"
+}

DMS/dinov3_siglip2_gated/model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5751d9a8428036b671c11a9fc440931ad0888d497feb7b2a0f93271b4f18b293
+size 7959608

DMS/dinov3_siglip2_gated/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c71607a87521793b66f1fc13ba984fe2b6954774b0caad559ae94d72de33b12
+size 7952748

DMS/dinov3_siglip2_gated_reg/README.md

@@ -0,0 +1,52 @@
+# DINOv3-SigLIP2 + Gated Multimodal Unit on Apple Dense Material Segmentation
+
+A material-classification **fusion head** from SegObjectMaterial. It classifies the
+material of a masked object from two frozen embeddings:
+
+| | |
+|---|---|
+| vision | DINOv3 ViT-L/16 (CLS token, frozen) -> 1024-d |
+| language | SigLIP2 base text encoder over Qwen3-VL-8B captions (frozen) -> 768-d |
+| head | Gated Multimodal Unit (`gated`, defaults), 1,985,081 parameters |
+| classes | 57 (Apple Dense Material Segmentation) |
+
+**This repo holds only the head.** Features must come from the same encoders,
+captioner and prompts as training (see the SegObjectMaterial repository); the
+encoders are downloaded from their own model pages under their own licenses.
+
+## Evaluation
+
+Held-out split recorded in `split.json` (source images never seen in training).
+Single training run -- differences under ~0.5 mIoU points between heads are within
+run-to-run noise.
+
+| mIoU | mAcc | eval samples |
+|---|---|---|
+| 0.4606 | 0.6403 | 186,231 |
+
+## Training
+
+```json
+{
+  "learning_rate": 5e-05,
+  "epochs": 50,
+  "batch_size": 32,
+  "weight_decay": 0.01,
+  "lr_schedule": "cosine",
+  "early_stopping_patience": 8,
+  "checkpoint": "best validation loss"
+}
+```
+
+## Use
+
+```python
+from utils.hub import load_head_from_hub      # SegObjectMaterial repository
+head, info = load_head_from_hub("s0um0r0y/SegObjectMaterial", "DMS/dinov3_siglip2_gated_reg")
+logits = head(vision_features, language_features)   # [B, 1024], [B, 768]
+```
+
+## Integrity
+
+- `model.safetensors` sha256 `651edac500aacfb32bec53e14e7b2da4c660eb67acdaa350a8a700f5cf060453`
+- source config `config/config_latest_models_gated_reg.toml`, code commit `1a8130c922267fb0786a33e3854a7e00763cc38f+dirty`

DMS/dinov3_siglip2_gated_reg/config.toml

@@ -0,0 +1,72 @@
+# REGULARISED GATED run: config.toml (gated head) trained with the SAME recipe
+# as config_latest_models_lmf_reg.toml -- weight decay 1e-2, cosine LR, early
+# stopping. Without this control, an LMF gain could be the recipe, not the head.
+# (No modality dropout: only the LMF head implements it.) Own checkpoint.
+
+# Apple DMS -- LATEST-MODELS pipeline: DINOv3 ViT-L/16 (1024) + Qwen3-VL -> SigLIP2 text (768)
+#
+# Pair this with:
+#   src/vision_stream_latest_models.py + src/language_stream_latest_models.py
+#   run_serialization_latest_models.py  ->  train_latest_models.py
+#
+# For the base DINOv2/CLIP pipeline (run_serialization.py -> train.py) use
+# config/config_dinov2_clip.toml instead -- its vision_dim is 768, not 1024.
+
+[data]
+dataset_type = "apple_dms"
+image_dir = "<local>/images"
+mask_dir = "<local>/labels"
+# Rebuilt automatically on first run; the index now stores one entry per
+# CONNECTED COMPONENT, so any pre-existing labels.json must be removed first.
+labels_file = "<local>/labels_instances.json"
+# Separate cache: DINOv3 features are 1024-d and must not mix with the
+# 768-d DINOv2 cache written by run_serialization.py.
+tensor_cache_dir = "<local>/tensor_cache_dir_latest_models"
+
+# Connected-component filtering (measured on 400 real DMS masks):
+# keeping every component gives 7.5x more samples than the old per-(image,class)
+# scheme -- 2.45M samples / 47 GB / ~7.5x the VLM captioning time -- because 69%
+# of components are under 0.1% of the frame and crop to a few pixels.
+# A component is kept when area >= max(min_instance_area, min_instance_frac * frame).
+min_instance_area = 1000        # absolute pixel floor
+min_instance_frac = 0.001       # and at least 0.1% of the frame
+max_instances_per_class = 0     # 0 = keep all; set e.g. 3 to cap the long tail
+
+
+[model]
+# fusion head: "gated" (GMU) or "mlp" (concat+MLP). Read by train_latest_models.py; a contradicting --fusion flag is an error, so checkpoint_path can't hold the wrong head.
+fusion = "gated"
+vision_dim = 1024     # DINOv3 ViT-L/16 CLS token
+language_dim = 768    # SigLIP2 base text embedding
+num_classes = 57
+
+[training]
+batch_size = 32
+
+# Batch size for the OFFLINE SERIALIZATION pass, which is a completely different
+# memory regime from training: it pushes full-resolution images through the 7B
+# VLM, whereas [training].batch_size only moves cached 768-d vectors through a
+# small MLP. Measured on the 23.6 GiB 4090 -- static weights 16.2 GiB, so only
+# ~7.4 GiB is left: batch 8 peaks at 17.7 GiB, 16 at 19.2 GiB, and 32 OOMs.
+# 8 leaves headroom for a labmate's job to coexist; raise to 16 on a free card.
+serialization_batch_size = 8
+learning_rate = 5e-5
+epochs = 50
+train_ratio = 0.8  # 80% for training
+val_ratio = 0.2    # 20% for validation
+tensorboard_log_dir = "<local>/dms_dinov3_siglip2_gated_reg"
+
+# ---- regularisation / schedule (train_latest_models.py; all optional) ----
+# The baseline runs (wd 1e-4, constant LR, 50 epochs) overfit on DMS: val loss
+# bottomed at epoch 13-15 and then rose while train accuracy kept climbing.
+weight_decay = 1e-2            # was 1e-4 (the default when this key is absent)
+lr_schedule = "cosine"         # "constant" (default) | "cosine", stepped per epoch over `epochs`
+min_lr = 1e-6                  # cosine floor
+early_stopping_patience = 8    # stop after 8 epochs without val-loss improvement; 0 = off
+
+[inference]
+checkpoint_path = "<local>/best_model_dms_latest_gated_reg.pth"
+
+[logging]
+wandb_project = "material-classification"
+wandb_run_name = "dms-dinov3-siglip2-gated-reg"

DMS/dinov3_siglip2_gated_reg/head.json

@@ -0,0 +1,99 @@
+{
+  "fusion": "gated",
+  "head": "Gated Multimodal Unit",
+  "head_kwargs": {},
+  "vision_dim": 1024,
+  "language_dim": 768,
+  "num_classes": 57,
+  "classes": [
+    "No label",
+    "Animal skin",
+    "Bone/teeth/horn",
+    "Brickwork",
+    "Cardboard",
+    "Carpet/rug",
+    "Ceiling tile",
+    "Ceramic",
+    "Chalkboard/blackboard",
+    "Clutter",
+    "Concrete",
+    "Cork/corkboard",
+    "Engineered stone",
+    "Fabric/cloth",
+    "Fiberglass wool",
+    "Fire",
+    "Foliage",
+    "Food",
+    "Fur",
+    "Gemstone/quartz",
+    "Glass",
+    "Hair",
+    "I cannot tell",
+    "Ice",
+    "Leather",
+    "Liquid, non-water",
+    "Metal",
+    "Mirror",
+    "Not on list",
+    "Paint/plaster/enamel",
+    "Paper",
+    "Pearl",
+    "Photograph/painting",
+    "Plastic, clear",
+    "Plastic, non-clear",
+    "Rubber/latex",
+    "Sand",
+    "Skin/lips",
+    "Sky",
+    "Snow",
+    "Soap",
+    "Soil/mud",
+    "Sponge",
+    "Stone, natural",
+    "Stone, polished",
+    "Styrofoam",
+    "Tile",
+    "Wallpaper",
+    "Water",
+    "Wax",
+    "Whiteboard",
+    "Wicker",
+    "Wood",
+    "Wood, tree",
+    "Bad polygon",
+    "Multiple materials",
+    "Asphalt"
+  ],
+  "dataset": "Apple Dense Material Segmentation",
+  "dataset_key": "DMS",
+  "encoders": {
+    "vision": "DINOv3 ViT-L/16 (CLS token, frozen)",
+    "language": "SigLIP2 base text encoder over Qwen3-VL-8B captions (frozen)",
+    "short": "DINOv3-SigLIP2",
+    "slug": "dinov3_siglip2"
+  },
+  "parameters": 1985081,
+  "training": {
+    "learning_rate": 5e-05,
+    "epochs": 50,
+    "batch_size": 32,
+    "weight_decay": 0.01,
+    "lr_schedule": "cosine",
+    "early_stopping_patience": 8,
+    "checkpoint": "best validation loss"
+  },
+  "sha256": {
+    "safetensors": "651edac500aacfb32bec53e14e7b2da4c660eb67acdaa350a8a700f5cf060453",
+    "pth": "b3dd5b15372df360a54de3889039b15cced829cbe485bfd4ab07e0e92bc9d340"
+  },
+  "source": {
+    "config": "config/config_latest_models_gated_reg.toml",
+    "git_commit": "1a8130c922267fb0786a33e3854a7e00763cc38f+dirty",
+    "checkpoint_mtime": "2026-10-07T18:00:41"
+  },
+  "metrics": {
+    "mIoU": 0.4606188697826707,
+    "mAcc": 0.6402625240517726,
+    "eval_samples": 186231
+  }
+}

DMS/dinov3_siglip2_gated_reg/metrics.json

@@ -0,0 +1,296 @@
+{
+  "dataset": "DMS",
+  "config": "config/config_latest_models_gated_reg.toml",
+  "checkpoint": "<local>/best_model_dms_latest_gated_reg.pth",
+  "fusion_type": "gated",
+  "num_eval_samples": 186231,
+  "mIoU": 0.4606188697826707,
+  "mAcc": 0.6402625240517726,
+  "per_class": {
+    "No label": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Animal skin": {
+      "accuracy": 0.631578947368421,
+      "iou": 0.4,
+      "support": 114
+    },
+    "Bone/teeth/horn": {
+      "accuracy": 0.75,
+      "iou": 0.6470588235294118,
+      "support": 44
+    },
+    "Brickwork": {
+      "accuracy": 0.7631814119749777,
+      "iou": 0.5381222432262129,
+      "support": 1119
+    },
+    "Cardboard": {
+      "accuracy": 0.7526881720430108,
+      "iou": 0.48826886493341787,
+      "support": 1023
+    },
+    "Carpet/rug": {
+      "accuracy": 0.8024316109422492,
+      "iou": 0.4708120318630365,
+      "support": 4935
+    },
+    "Ceiling tile": {
+      "accuracy": 0.7756539235412475,
+      "iou": 0.5102581072137657,
+      "support": 994
+    },
+    "Ceramic": {
+      "accuracy": 0.8300567933595456,
+      "iou": 0.6176853055916776,
+      "support": 2289
+    },
+    "Chalkboard/blackboard": {
+      "accuracy": 0.7302325581395349,
+      "iou": 0.548951048951049,
+      "support": 215
+    },
+    "Clutter": {
+      "accuracy": 0.4716981132075472,
+      "iou": 0.3424657534246575,
+      "support": 53
+    },
+    "Concrete": {
+      "accuracy": 0.5145102781136639,
+      "iou": 0.28263035536366654,
+      "support": 1654
+    },
+    "Cork/corkboard": {
+      "accuracy": 0.37254901960784315,
+      "iou": 0.21348314606741572,
+      "support": 51
+    },
+    "Engineered stone": {
+      "accuracy": 0.23595505617977527,
+      "iou": 0.1640625,
+      "support": 89
+    },
+    "Fabric/cloth": {
+      "accuracy": 0.7821157578965943,
+      "iou": 0.7193167525459488,
+      "support": 22668
+    },
+    "Fiberglass wool": {
+      "accuracy": 0.3,
+      "iou": 0.18181818181818182,
+      "support": 20
+    },
+    "Fire": {
+      "accuracy": 0.8,
+      "iou": 0.49382716049382713,
+      "support": 50
+    },
+    "Foliage": {
+      "accuracy": 0.9101516919486581,
+      "iou": 0.7878787878787878,
+      "support": 5142
+    },
+    "Food": {
+      "accuracy": 0.8636871508379889,
+      "iou": 0.6578723404255319,
+      "support": 895
+    },
+    "Fur": {
+      "accuracy": 0.8259860788863109,
+      "iou": 0.688588007736944,
+      "support": 431
+    },
+    "Gemstone/quartz": {
+      "accuracy": 0.68,
+      "iou": 0.2677165354330709,
+      "support": 50
+    },
+    "Glass": {
+      "accuracy": 0.7577358490566037,
+      "iou": 0.6721008072455208,
+      "support": 22525
+    },
+    "Hair": {
+      "accuracy": 0.9782991202346041,
+      "iou": 0.9238438105787871,
+      "support": 6820
+    },
+    "I cannot tell": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Ice": {
+      "accuracy": 0.5454545454545454,
+      "iou": 0.46875,
+      "support": 55
+    },
+    "Leather": {
+      "accuracy": 0.7197750702905342,
+      "iou": 0.43986254295532645,
+      "support": 2134
+    },
+    "Liquid, non-water": {
+      "accuracy": 0.4722222222222222,
+      "iou": 0.2698412698412698,
+      "support": 36
+    },
+    "Metal": {
+      "accuracy": 0.8048137157929443,
+      "iou": 0.5654389622422978,
+      "support": 9099
+    },
+    "Mirror": {
+      "accuracy": 0.6888331242158093,
+      "iou": 0.43364928909952605,
+      "support": 797
+    },
+    "Not on list": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Paint/plaster/enamel": {
+      "accuracy": 0.619308004148204,
+      "iou": 0.5437013739447112,
+      "support": 31821
+    },
+    "Paper": {
+      "accuracy": 0.8385812842738369,
+      "iou": 0.693562545720556,
+      "support": 11306
+    },
+    "Pearl": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 1
+    },
+    "Photograph/painting": {
+      "accuracy": 0.7142001198322349,
+      "iou": 0.38402061855670105,
+      "support": 1669
+    },
+    "Plastic, clear": {
+      "accuracy": 0.7065420560747664,
+      "iou": 0.4186046511627907,
+      "support": 1605
+    },
+    "Plastic, non-clear": {
+      "accuracy": 0.6639065817409766,
+      "iou": 0.5230263157894737,
+      "support": 14130
+    },
+    "Rubber/latex": {
+      "accuracy": 0.6011183597390494,
+      "iou": 0.3425385023898035,
+      "support": 1073
+    },
+    "Sand": {
+      "accuracy": 0.3409090909090909,
+      "iou": 0.24793388429752067,
+      "support": 88
+    },
+    "Skin/lips": {
+      "accuracy": 0.973601827565784,
+      "iou": 0.9409600130836536,
+      "support": 11819
+    },
+    "Sky": {
+      "accuracy": 0.9201485608170845,
+      "iou": 0.7038352272727273,
+      "support": 2154
+    },
+    "Snow": {
+      "accuracy": 0.7431192660550459,
+      "iou": 0.5436241610738255,
+      "support": 109
+    },
+    "Soap": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 7
+    },
+    "Soil/mud": {
+      "accuracy": 0.6773211567732116,
+      "iou": 0.3938053097345133,
+      "support": 657
+    },
+    "Sponge": {
+      "accuracy": 0.2777777777777778,
+      "iou": 0.2,
+      "support": 18
+    },
+    "Stone, natural": {
+      "accuracy": 0.6347256347256347,
+      "iou": 0.4487550665894615,
+      "support": 1221
+    },
+    "Stone, polished": {
+      "accuracy": 0.399737876802097,
+      "iou": 0.2654482158398607,
+      "support": 763
+    },
+    "Styrofoam": {
+      "accuracy": 0.12903225806451613,
+      "iou": 0.08695652173913043,
+      "support": 31
+    },
+    "Tile": {
+      "accuracy": 0.649852148199687,
+      "iou": 0.44124247076886736,
+      "support": 5749
+    },
+    "Wallpaper": {
+      "accuracy": 0.6175595238095238,
+      "iou": 0.3312051077414206,
+      "support": 672
+    },
+    "Water": {
+      "accuracy": 0.6696935300794552,
+      "iou": 0.508182601205857,
+      "support": 881
+    },
+    "Wax": {
+      "accuracy": 0.8771929824561403,
+      "iou": 0.6711409395973155,
+      "support": 114
+    },
+    "Whiteboard": {
+      "accuracy": 0.6636636636636637,
+      "iou": 0.4258188824662813,
+      "support": 333
+    },
+    "Wicker": {
+      "accuracy": 0.7884297520661157,
+      "iou": 0.5740072202166066,
+      "support": 605
+    },
+    "Wood": {
+      "accuracy": 0.8096232943787948,
+      "iou": 0.631544102668568,
+      "support": 15317
+    },
+    "Wood, tree": {
+      "accuracy": 0.7482993197278912,
+      "iou": 0.5034324942791762,
+      "support": 588
+    },
+    "Bad polygon": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Multiple materials": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Asphalt": {
+      "accuracy": 0.4696969696969697,
+      "iou": 0.3345323741007194,
+      "support": 198
+    }
+  }
+}

DMS/dinov3_siglip2_gated_reg/model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3dd5b15372df360a54de3889039b15cced829cbe485bfd4ab07e0e92bc9d340
+size 7959811

DMS/dinov3_siglip2_gated_reg/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:651edac500aacfb32bec53e14e7b2da4c660eb67acdaa350a8a700f5cf060453
+size 7952748

DMS/dinov3_siglip2_lmf/README.md

@@ -0,0 +1,52 @@
+# DINOv3-SigLIP2 + Low-rank Multimodal Fusion on Apple Dense Material Segmentation
+
+A material-classification **fusion head** from SegObjectMaterial. It classifies the
+material of a masked object from two frozen embeddings:
+
+| | |
+|---|---|
+| vision | DINOv3 ViT-L/16 (CLS token, frozen) -> 1024-d |
+| language | SigLIP2 base text encoder over Qwen3-VL-8B captions (frozen) -> 768-d |
+| head | Low-rank Multimodal Fusion (`lmf`, rank=4, hidden_dim=256, modality_dropout=0.0), 1,922,365 parameters |
+| classes | 57 (Apple Dense Material Segmentation) |
+
+**This repo holds only the head.** Features must come from the same encoders,
+captioner and prompts as training (see the SegObjectMaterial repository); the
+encoders are downloaded from their own model pages under their own licenses.
+
+## Evaluation
+
+Held-out split recorded in `split.json` (source images never seen in training).
+Single training run -- differences under ~0.5 mIoU points between heads are within
+run-to-run noise.
+
+| mIoU | mAcc | eval samples |
+|---|---|---|
+| 0.4407 | 0.6181 | 186,231 |
+
+## Training
+
+```json
+{
+  "learning_rate": 5e-05,
+  "epochs": 50,
+  "batch_size": 32,
+  "weight_decay": 0.0001,
+  "lr_schedule": "constant",
+  "early_stopping_patience": 0,
+  "checkpoint": "best validation loss"
+}
+```
+
+## Use
+
+```python
+from utils.hub import load_head_from_hub      # SegObjectMaterial repository
+head, info = load_head_from_hub("s0um0r0y/SegObjectMaterial", "DMS/dinov3_siglip2_lmf")
+logits = head(vision_features, language_features)   # [B, 1024], [B, 768]
+```
+
+## Integrity
+
+- `model.safetensors` sha256 `ee69c3ea4a8f7ecee29b767287987ec1c19082370b2dad48ae4d4b76f45b386a`
+- source config `config/config_latest_models_lmf.toml`, code commit `1a8130c922267fb0786a33e3854a7e00763cc38f+dirty`

DMS/dinov3_siglip2_lmf/config.toml

@@ -0,0 +1,77 @@
+# LMF-head twin of config/config.toml.
+# SAME embedding cache -- the head type does not change the cached features, so
+# nothing needs re-serializing. Only the checkpoint, TensorBoard dir and W&B run
+# differ, so this run can never overwrite the gated or mlp checkpoints.
+
+# Apple DMS -- LATEST-MODELS pipeline: DINOv3 ViT-L/16 (1024) + Qwen3-VL -> SigLIP2 text (768)
+#
+# Pair this with:
+#   src/vision_stream_latest_models.py + src/language_stream_latest_models.py
+#   run_serialization_latest_models.py  ->  train_latest_models.py
+#
+# For the base DINOv2/CLIP pipeline (run_serialization.py -> train.py) use
+# config/config_dinov2_clip.toml instead -- its vision_dim is 768, not 1024.
+
+[data]
+dataset_type = "apple_dms"
+image_dir = "<local>/images"
+mask_dir = "<local>/labels"
+# Rebuilt automatically on first run; the index now stores one entry per
+# CONNECTED COMPONENT, so any pre-existing labels.json must be removed first.
+labels_file = "<local>/labels_instances.json"
+# Separate cache: DINOv3 features are 1024-d and must not mix with the
+# 768-d DINOv2 cache written by run_serialization.py.
+tensor_cache_dir = "<local>/tensor_cache_dir_latest_models"
+
+# Connected-component filtering (measured on 400 real DMS masks):
+# keeping every component gives 7.5x more samples than the old per-(image,class)
+# scheme -- 2.45M samples / 47 GB / ~7.5x the VLM captioning time -- because 69%
+# of components are under 0.1% of the frame and crop to a few pixels.
+# A component is kept when area >= max(min_instance_area, min_instance_frac * frame).
+min_instance_area = 1000        # absolute pixel floor
+min_instance_frac = 0.001       # and at least 0.1% of the frame
+max_instances_per_class = 0     # 0 = keep all; set e.g. 3 to cap the long tail
+
+
+[model]
+# fusion head: "lmf" = low-rank multimodal fusion (src/fusion_classifier.py,
+# LowRankFusionHead). Read by train_latest_models.py; a contradicting --fusion
+# flag is an error, so checkpoint_path can't hold the wrong head.
+fusion = "lmf"
+vision_dim = 1024     # DINOv3 ViT-L/16 CLS token
+language_dim = 768    # SigLIP2 base text embedding
+num_classes = 57
+
+# LMF hyperparameters. rank 4 x hidden 256 = ~1.84M fusion parameters, close to
+# the gated head (~1.97M), so gated-vs-lmf compares the fusion operator rather
+# than model size. Both are read back from the checkpoint at load time, so
+# evaluate/inference scripts need no extra flags.
+lmf_rank = 4
+lmf_hidden = 256
+# Fraction of training samples with ONE stream zeroed (half vision, half
+# language). 0.0 = plain LMF, as in the paper. Try 0.2 as a second run: it
+# forces the head to work from either stream alone.
+modality_dropout = 0.0
+
+[training]
+batch_size = 32
+
+# Batch size for the OFFLINE SERIALIZATION pass, which is a completely different
+# memory regime from training: it pushes full-resolution images through the 7B
+# VLM, whereas [training].batch_size only moves cached 768-d vectors through a
+# small MLP. Measured on the 23.6 GiB 4090 -- static weights 16.2 GiB, so only
+# ~7.4 GiB is left: batch 8 peaks at 17.7 GiB, 16 at 19.2 GiB, and 32 OOMs.
+# 8 leaves headroom for a labmate's job to coexist; raise to 16 on a free card.
+serialization_batch_size = 8
+learning_rate = 5e-5
+epochs = 50
+train_ratio = 0.8  # 80% for training
+val_ratio = 0.2    # 20% for validation
+tensorboard_log_dir = "<local>/dms_dinov3_siglip2_lmf"
+
+[inference]
+checkpoint_path = "<local>/best_model_dms_latest_lmf.pth"
+
+[logging]
+wandb_project = "material-classification"
+wandb_run_name = "dms-dinov3-siglip2-lmf"

DMS/dinov3_siglip2_lmf/head.json

@@ -0,0 +1,103 @@
+{
+  "fusion": "lmf",
+  "head": "Low-rank Multimodal Fusion",
+  "head_kwargs": {
+    "rank": 4,
+    "hidden_dim": 256,
+    "modality_dropout": 0.0
+  },
+  "vision_dim": 1024,
+  "language_dim": 768,
+  "num_classes": 57,
+  "classes": [
+    "No label",
+    "Animal skin",
+    "Bone/teeth/horn",
+    "Brickwork",
+    "Cardboard",
+    "Carpet/rug",
+    "Ceiling tile",
+    "Ceramic",
+    "Chalkboard/blackboard",
+    "Clutter",
+    "Concrete",
+    "Cork/corkboard",
+    "Engineered stone",
+    "Fabric/cloth",
+    "Fiberglass wool",
+    "Fire",
+    "Foliage",
+    "Food",
+    "Fur",
+    "Gemstone/quartz",
+    "Glass",
+    "Hair",
+    "I cannot tell",
+    "Ice",
+    "Leather",
+    "Liquid, non-water",
+    "Metal",
+    "Mirror",
+    "Not on list",
+    "Paint/plaster/enamel",
+    "Paper",
+    "Pearl",
+    "Photograph/painting",
+    "Plastic, clear",
+    "Plastic, non-clear",
+    "Rubber/latex",
+    "Sand",
+    "Skin/lips",
+    "Sky",
+    "Snow",
+    "Soap",
+    "Soil/mud",
+    "Sponge",
+    "Stone, natural",
+    "Stone, polished",
+    "Styrofoam",
+    "Tile",
+    "Wallpaper",
+    "Water",
+    "Wax",
+    "Whiteboard",
+    "Wicker",
+    "Wood",
+    "Wood, tree",
+    "Bad polygon",
+    "Multiple materials",
+    "Asphalt"
+  ],
+  "dataset": "Apple Dense Material Segmentation",
+  "dataset_key": "DMS",
+  "encoders": {
+    "vision": "DINOv3 ViT-L/16 (CLS token, frozen)",
+    "language": "SigLIP2 base text encoder over Qwen3-VL-8B captions (frozen)",
+    "short": "DINOv3-SigLIP2",
+    "slug": "dinov3_siglip2"
+  },
+  "parameters": 1922365,
+  "training": {
+    "learning_rate": 5e-05,
+    "epochs": 50,
+    "batch_size": 32,
+    "weight_decay": 0.0001,
+    "lr_schedule": "constant",
+    "early_stopping_patience": 0,
+    "checkpoint": "best validation loss"
+  },
+  "sha256": {
+    "safetensors": "ee69c3ea4a8f7ecee29b767287987ec1c19082370b2dad48ae4d4b76f45b386a",
+    "pth": "a5b00ebaeb97a3bb93314ed849bef99d49f222cb4f9f82c5be4d659ed87dca30"
+  },
+  "source": {
+    "config": "config/config_latest_models_lmf.toml",
+    "git_commit": "1a8130c922267fb0786a33e3854a7e00763cc38f+dirty",
+    "checkpoint_mtime": "2026-10-07T14:53:13"
+  },
+  "metrics": {
+    "mIoU": 0.440722402484214,
+    "mAcc": 0.6180633687278619,
+    "eval_samples": 186231
+  }
+}

DMS/dinov3_siglip2_lmf/metrics.json

@@ -0,0 +1,296 @@
+{
+  "dataset": "DMS",
+  "checkpoint": "<local>/best_model_dms_latest_lmf.pth",
+  "fusion_type": "lmf",
+  "num_eval_samples": 186231,
+  "mIoU": 0.440722402484214,
+  "mAcc": 0.6180633687278619,
+  "per_class": {
+    "No label": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Animal skin": {
+      "accuracy": 0.6052631578947368,
+      "iou": 0.41566265060240964,
+      "support": 114
+    },
+    "Bone/teeth/horn": {
+      "accuracy": 0.6590909090909091,
+      "iou": 0.5918367346938775,
+      "support": 44
+    },
+    "Brickwork": {
+      "accuracy": 0.7363717605004468,
+      "iou": 0.5061425061425061,
+      "support": 1119
+    },
+    "Cardboard": {
+      "accuracy": 0.7448680351906158,
+      "iou": 0.47270471464019853,
+      "support": 1023
+    },
+    "Carpet/rug": {
+      "accuracy": 0.7785207700101318,
+      "iou": 0.483757239989927,
+      "support": 4935
+    },
+    "Ceiling tile": {
+      "accuracy": 0.7716297786720322,
+      "iou": 0.4845230574857865,
+      "support": 994
+    },
+    "Ceramic": {
+      "accuracy": 0.8283093053735255,
+      "iou": 0.5846438482886216,
+      "support": 2289
+    },
+    "Chalkboard/blackboard": {
+      "accuracy": 0.7348837209302326,
+      "iou": 0.531986531986532,
+      "support": 215
+    },
+    "Clutter": {
+      "accuracy": 0.41509433962264153,
+      "iou": 0.3235294117647059,
+      "support": 53
+    },
+    "Concrete": {
+      "accuracy": 0.5126964933494559,
+      "iou": 0.2663316582914573,
+      "support": 1654
+    },
+    "Cork/corkboard": {
+      "accuracy": 0.29411764705882354,
+      "iou": 0.18292682926829268,
+      "support": 51
+    },
+    "Engineered stone": {
+      "accuracy": 0.19101123595505617,
+      "iou": 0.14049586776859505,
+      "support": 89
+    },
+    "Fabric/cloth": {
+      "accuracy": 0.7754543850361744,
+      "iou": 0.708876073718595,
+      "support": 22668
+    },
+    "Fiberglass wool": {
+      "accuracy": 0.25,
+      "iou": 0.15625,
+      "support": 20
+    },
+    "Fire": {
+      "accuracy": 0.78,
+      "iou": 0.47560975609756095,
+      "support": 50
+    },
+    "Foliage": {
+      "accuracy": 0.9074290159471023,
+      "iou": 0.7656711519527404,
+      "support": 5142
+    },
+    "Food": {
+      "accuracy": 0.8670391061452514,
+      "iou": 0.6308943089430894,
+      "support": 895
+    },
+    "Fur": {
+      "accuracy": 0.8167053364269141,
+      "iou": 0.6579439252336449,
+      "support": 431
+    },
+    "Gemstone/quartz": {
+      "accuracy": 0.6,
+      "iou": 0.21739130434782608,
+      "support": 50
+    },
+    "Glass": {
+      "accuracy": 0.753607103218646,
+      "iou": 0.6518566875312009,
+      "support": 22525
+    },
+    "Hair": {
+      "accuracy": 0.9782991202346041,
+      "iou": 0.9080021774632553,
+      "support": 6820
+    },
+    "I cannot tell": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Ice": {
+      "accuracy": 0.4909090909090909,
+      "iou": 0.39705882352941174,
+      "support": 55
+    },
+    "Leather": {
+      "accuracy": 0.7235238987816307,
+      "iou": 0.40450615666754,
+      "support": 2134
+    },
+    "Liquid, non-water": {
+      "accuracy": 0.4722222222222222,
+      "iou": 0.3148148148148148,
+      "support": 36
+    },
+    "Metal": {
+      "accuracy": 0.8027255742389273,
+      "iou": 0.5505389311826336,
+      "support": 9099
+    },
+    "Mirror": {
+      "accuracy": 0.671267252195734,
+      "iou": 0.4080854309687262,
+      "support": 797
+    },
+    "Not on list": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Paint/plaster/enamel": {
+      "accuracy": 0.6058577668835046,
+      "iou": 0.5275846970609162,
+      "support": 31821
+    },
+    "Paper": {
+      "accuracy": 0.823810366177251,
+      "iou": 0.6853064527996469,
+      "support": 11306
+    },
+    "Pearl": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 1
+    },
+    "Photograph/painting": {
+      "accuracy": 0.6854403834631516,
+      "iou": 0.38209752839011357,
+      "support": 1669
+    },
+    "Plastic, clear": {
+      "accuracy": 0.6834890965732088,
+      "iou": 0.40735239509840326,
+      "support": 1605
+    },
+    "Plastic, non-clear": {
+      "accuracy": 0.6435244161358811,
+      "iou": 0.5068844417191594,
+      "support": 14130
+    },
+    "Rubber/latex": {
+      "accuracy": 0.5526561043802423,
+      "iou": 0.32944444444444443,
+      "support": 1073
+    },
+    "Sand": {
+      "accuracy": 0.32954545454545453,
+      "iou": 0.25,
+      "support": 88
+    },
+    "Skin/lips": {
+      "accuracy": 0.9675099416194264,
+      "iou": 0.9362207303094809,
+      "support": 11819
+    },
+    "Sky": {
+      "accuracy": 0.9317548746518106,
+      "iou": 0.6459607338268426,
+      "support": 2154
+    },
+    "Snow": {
+      "accuracy": 0.6788990825688074,
+      "iou": 0.4567901234567901,
+      "support": 109
+    },
+    "Soap": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 7
+    },
+    "Soil/mud": {
+      "accuracy": 0.6438356164383562,
+      "iou": 0.3960674157303371,
+      "support": 657
+    },
+    "Sponge": {
+      "accuracy": 0.16666666666666666,
+      "iou": 0.13636363636363635,
+      "support": 18
+    },
+    "Stone, natural": {
+      "accuracy": 0.6224406224406225,
+      "iou": 0.45292014302741357,
+      "support": 1221
+    },
+    "Stone, polished": {
+      "accuracy": 0.36828309305373524,
+      "iou": 0.2526978417266187,
+      "support": 763
+    },
+    "Styrofoam": {
+      "accuracy": 0.03225806451612903,
+      "iou": 0.03125,
+      "support": 31
+    },
+    "Tile": {
+      "accuracy": 0.6536789006783789,
+      "iou": 0.43789326497319975,
+      "support": 5749
+    },
+    "Wallpaper": {
+      "accuracy": 0.5744047619047619,
+      "iou": 0.3534798534798535,
+      "support": 672
+    },
+    "Water": {
+      "accuracy": 0.6515323496027242,
+      "iou": 0.49653979238754326,
+      "support": 881
+    },
+    "Wax": {
+      "accuracy": 0.8771929824561403,
+      "iou": 0.5586592178770949,
+      "support": 114
+    },
+    "Whiteboard": {
+      "accuracy": 0.6606606606606606,
+      "iou": 0.4365079365079365,
+      "support": 333
+    },
+    "Wicker": {
+      "accuracy": 0.7867768595041322,
+      "iou": 0.5854858548585485,
+      "support": 605
+    },
+    "Wood": {
+      "accuracy": 0.7917999608278383,
+      "iou": 0.619122977181071,
+      "support": 15317
+    },
+    "Wood, tree": {
+      "accuracy": 0.7108843537414966,
+      "iou": 0.46756152125279643,
+      "support": 588
+    },
+    "Bad polygon": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Multiple materials": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Asphalt": {
+      "accuracy": 0.5353535353535354,
+      "iou": 0.3333333333333333,
+      "support": 198
+    }
+  },
+  "config": "config/config_latest_models_lmf.toml"
+}

DMS/dinov3_siglip2_lmf/model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5b00ebaeb97a3bb93314ed849bef99d49f222cb4f9f82c5be4d659ed87dca30
+size 7701480

DMS/dinov3_siglip2_lmf/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee69c3ea4a8f7ecee29b767287987ec1c19082370b2dad48ae4d4b76f45b386a
+size 7695412

DMS/dinov3_siglip2_lmf_reg/README.md

@@ -0,0 +1,52 @@
+# DINOv3-SigLIP2 + Low-rank Multimodal Fusion on Apple Dense Material Segmentation
+
+A material-classification **fusion head** from SegObjectMaterial. It classifies the
+material of a masked object from two frozen embeddings:
+
+| | |
+|---|---|
+| vision | DINOv3 ViT-L/16 (CLS token, frozen) -> 1024-d |
+| language | SigLIP2 base text encoder over Qwen3-VL-8B captions (frozen) -> 768-d |
+| head | Low-rank Multimodal Fusion (`lmf`, rank=4, hidden_dim=256, modality_dropout=0.2), 1,922,365 parameters |
+| classes | 57 (Apple Dense Material Segmentation) |
+
+**This repo holds only the head.** Features must come from the same encoders,
+captioner and prompts as training (see the SegObjectMaterial repository); the
+encoders are downloaded from their own model pages under their own licenses.
+
+## Evaluation
+
+Held-out split recorded in `split.json` (source images never seen in training).
+Single training run -- differences under ~0.5 mIoU points between heads are within
+run-to-run noise.
+
+| mIoU | mAcc | eval samples |
+|---|---|---|
+| 0.4453 | 0.6135 | 186,231 |
+
+## Training
+
+```json
+{
+  "learning_rate": 5e-05,
+  "epochs": 50,
+  "batch_size": 32,
+  "weight_decay": 0.01,
+  "lr_schedule": "cosine",
+  "early_stopping_patience": 10,
+  "checkpoint": "best validation loss"
+}
+```
+
+## Use
+
+```python
+from utils.hub import load_head_from_hub      # SegObjectMaterial repository
+head, info = load_head_from_hub("s0um0r0y/SegObjectMaterial", "DMS/dinov3_siglip2_lmf_reg")
+logits = head(vision_features, language_features)   # [B, 1024], [B, 768]
+```
+
+## Integrity
+
+- `model.safetensors` sha256 `ed2fc3b7ee4dd92b575b199ea2770e3a0e0d702355dfe5f97216e50917c68cab`
+- source config `config/config_latest_models_lmf_reg.toml`, code commit `1a8130c922267fb0786a33e3854a7e00763cc38f+dirty`

DMS/dinov3_siglip2_lmf_reg/config.toml

@@ -0,0 +1,89 @@
+# REGULARISED LMF run: config_latest_models_lmf.toml + modality dropout 0.2,
+# weight decay 1e-2, cosine LR and early stopping. Same cache; own checkpoint.
+# Compare against config_latest_models_gated_reg.toml (same recipe, gated head).
+
+# LMF-head twin of config/config.toml.
+# SAME embedding cache -- the head type does not change the cached features, so
+# nothing needs re-serializing. Only the checkpoint, TensorBoard dir and W&B run
+# differ, so this run can never overwrite the gated or mlp checkpoints.
+
+# Apple DMS -- LATEST-MODELS pipeline: DINOv3 ViT-L/16 (1024) + Qwen3-VL -> SigLIP2 text (768)
+#
+# Pair this with:
+#   src/vision_stream_latest_models.py + src/language_stream_latest_models.py
+#   run_serialization_latest_models.py  ->  train_latest_models.py
+#
+# For the base DINOv2/CLIP pipeline (run_serialization.py -> train.py) use
+# config/config_dinov2_clip.toml instead -- its vision_dim is 768, not 1024.
+
+[data]
+dataset_type = "apple_dms"
+image_dir = "<local>/images"
+mask_dir = "<local>/labels"
+# Rebuilt automatically on first run; the index now stores one entry per
+# CONNECTED COMPONENT, so any pre-existing labels.json must be removed first.
+labels_file = "<local>/labels_instances.json"
+# Separate cache: DINOv3 features are 1024-d and must not mix with the
+# 768-d DINOv2 cache written by run_serialization.py.
+tensor_cache_dir = "<local>/tensor_cache_dir_latest_models"
+
+# Connected-component filtering (measured on 400 real DMS masks):
+# keeping every component gives 7.5x more samples than the old per-(image,class)
+# scheme -- 2.45M samples / 47 GB / ~7.5x the VLM captioning time -- because 69%
+# of components are under 0.1% of the frame and crop to a few pixels.
+# A component is kept when area >= max(min_instance_area, min_instance_frac * frame).
+min_instance_area = 1000        # absolute pixel floor
+min_instance_frac = 0.001       # and at least 0.1% of the frame
+max_instances_per_class = 0     # 0 = keep all; set e.g. 3 to cap the long tail
+
+
+[model]
+# fusion head: "lmf" = low-rank multimodal fusion (src/fusion_classifier.py,
+# LowRankFusionHead). Read by train_latest_models.py; a contradicting --fusion
+# flag is an error, so checkpoint_path can't hold the wrong head.
+fusion = "lmf"
+vision_dim = 1024     # DINOv3 ViT-L/16 CLS token
+language_dim = 768    # SigLIP2 base text embedding
+num_classes = 57
+
+# LMF hyperparameters. rank 4 x hidden 256 = ~1.84M fusion parameters, close to
+# the gated head (~1.97M), so gated-vs-lmf compares the fusion operator rather
+# than model size. Both are read back from the checkpoint at load time, so
+# evaluate/inference scripts need no extra flags.
+lmf_rank = 4
+lmf_hidden = 256
+# Fraction of training samples with ONE stream zeroed (half vision, half
+# language). 0.0 = plain LMF, as in the paper. Try 0.2 as a second run: it
+# forces the head to work from either stream alone.
+modality_dropout = 0.2
+
+[training]
+batch_size = 32
+
+# Batch size for the OFFLINE SERIALIZATION pass, which is a completely different
+# memory regime from training: it pushes full-resolution images through the 7B
+# VLM, whereas [training].batch_size only moves cached 768-d vectors through a
+# small MLP. Measured on the 23.6 GiB 4090 -- static weights 16.2 GiB, so only
+# ~7.4 GiB is left: batch 8 peaks at 17.7 GiB, 16 at 19.2 GiB, and 32 OOMs.
+# 8 leaves headroom for a labmate's job to coexist; raise to 16 on a free card.
+serialization_batch_size = 8
+learning_rate = 5e-5
+epochs = 50
+train_ratio = 0.8  # 80% for training
+val_ratio = 0.2    # 20% for validation
+tensorboard_log_dir = "<local>/dms_dinov3_siglip2_lmf_reg"
+
+# ---- regularisation / schedule (train_latest_models.py; all optional) ----
+# The baseline runs (wd 1e-4, constant LR, 50 epochs) overfit on DMS: val loss
+# bottomed at epoch 13-15 and then rose while train accuracy kept climbing.
+weight_decay = 1e-2            # was 1e-4 (the default when this key is absent)
+lr_schedule = "cosine"         # "constant" (default) | "cosine", stepped per epoch over `epochs`
+min_lr = 1e-6                  # cosine floor
+early_stopping_patience = 10    # stop after 8 epochs without val-loss improvement; 0 = off
+
+[inference]
+checkpoint_path = "<local>/best_model_dms_latest_lmf_reg.pth"
+
+[logging]
+wandb_project = "material-classification"
+wandb_run_name = "dms-dinov3-siglip2-lmf-reg"

DMS/dinov3_siglip2_lmf_reg/head.json

@@ -0,0 +1,103 @@
+{
+  "fusion": "lmf",
+  "head": "Low-rank Multimodal Fusion",
+  "head_kwargs": {
+    "rank": 4,
+    "hidden_dim": 256,
+    "modality_dropout": 0.2
+  },
+  "vision_dim": 1024,
+  "language_dim": 768,
+  "num_classes": 57,
+  "classes": [
+    "No label",
+    "Animal skin",
+    "Bone/teeth/horn",
+    "Brickwork",
+    "Cardboard",
+    "Carpet/rug",
+    "Ceiling tile",
+    "Ceramic",
+    "Chalkboard/blackboard",
+    "Clutter",
+    "Concrete",
+    "Cork/corkboard",
+    "Engineered stone",
+    "Fabric/cloth",
+    "Fiberglass wool",
+    "Fire",
+    "Foliage",
+    "Food",
+    "Fur",
+    "Gemstone/quartz",
+    "Glass",
+    "Hair",
+    "I cannot tell",
+    "Ice",
+    "Leather",
+    "Liquid, non-water",
+    "Metal",
+    "Mirror",
+    "Not on list",
+    "Paint/plaster/enamel",
+    "Paper",
+    "Pearl",
+    "Photograph/painting",
+    "Plastic, clear",
+    "Plastic, non-clear",
+    "Rubber/latex",
+    "Sand",
+    "Skin/lips",
+    "Sky",
+    "Snow",
+    "Soap",
+    "Soil/mud",
+    "Sponge",
+    "Stone, natural",
+    "Stone, polished",
+    "Styrofoam",
+    "Tile",
+    "Wallpaper",
+    "Water",
+    "Wax",
+    "Whiteboard",
+    "Wicker",
+    "Wood",
+    "Wood, tree",
+    "Bad polygon",
+    "Multiple materials",
+    "Asphalt"
+  ],
+  "dataset": "Apple Dense Material Segmentation",
+  "dataset_key": "DMS",
+  "encoders": {
+    "vision": "DINOv3 ViT-L/16 (CLS token, frozen)",
+    "language": "SigLIP2 base text encoder over Qwen3-VL-8B captions (frozen)",
+    "short": "DINOv3-SigLIP2",
+    "slug": "dinov3_siglip2"
+  },
+  "parameters": 1922365,
+  "training": {
+    "learning_rate": 5e-05,
+    "epochs": 50,
+    "batch_size": 32,
+    "weight_decay": 0.01,
+    "lr_schedule": "cosine",
+    "early_stopping_patience": 10,
+    "checkpoint": "best validation loss"
+  },
+  "sha256": {
+    "safetensors": "ed2fc3b7ee4dd92b575b199ea2770e3a0e0d702355dfe5f97216e50917c68cab",
+    "pth": "fcbfcb811c95b16bba82f2765cc37aeb0861ea6067e559a36c605bfdc24f1ba0"
+  },
+  "source": {
+    "config": "config/config_latest_models_lmf_reg.toml",
+    "git_commit": "1a8130c922267fb0786a33e3854a7e00763cc38f+dirty",
+    "checkpoint_mtime": "2026-10-07T17:11:52"
+  },
+  "metrics": {
+    "mIoU": 0.4453273662621192,
+    "mAcc": 0.6135327722998756,
+    "eval_samples": 186231
+  }
+}

DMS/dinov3_siglip2_lmf_reg/metrics.json

@@ -0,0 +1,296 @@
+{
+  "dataset": "DMS",
+  "config": "config/config_latest_models_lmf_reg.toml",
+  "checkpoint": "<local>/best_model_dms_latest_lmf_reg.pth",
+  "fusion_type": "lmf",
+  "num_eval_samples": 186231,
+  "mIoU": 0.4453273662621192,
+  "mAcc": 0.6135327722998756,
+  "per_class": {
+    "No label": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Animal skin": {
+      "accuracy": 0.5526315789473685,
+      "iou": 0.37058823529411766,
+      "support": 114
+    },
+    "Bone/teeth/horn": {
+      "accuracy": 0.6818181818181818,
+      "iou": 0.6122448979591837,
+      "support": 44
+    },
+    "Brickwork": {
+      "accuracy": 0.7462019660411081,
+      "iou": 0.4979129397734049,
+      "support": 1119
+    },
+    "Cardboard": {
+      "accuracy": 0.7360703812316716,
+      "iou": 0.4944189100459619,
+      "support": 1023
+    },
+    "Carpet/rug": {
+      "accuracy": 0.7937183383991895,
+      "iou": 0.46300236406619383,
+      "support": 4935
+    },
+    "Ceiling tile": {
+      "accuracy": 0.7847082494969819,
+      "iou": 0.4701627486437613,
+      "support": 994
+    },
+    "Ceramic": {
+      "accuracy": 0.8235037134119703,
+      "iou": 0.5771586037966931,
+      "support": 2289
+    },
+    "Chalkboard/blackboard": {
+      "accuracy": 0.7534883720930232,
+      "iou": 0.552901023890785,
+      "support": 215
+    },
+    "Clutter": {
+      "accuracy": 0.4339622641509434,
+      "iou": 0.359375,
+      "support": 53
+    },
+    "Concrete": {
+      "accuracy": 0.5151148730350665,
+      "iou": 0.2623152709359606,
+      "support": 1654
+    },
+    "Cork/corkboard": {
+      "accuracy": 0.29411764705882354,
+      "iou": 0.18518518518518517,
+      "support": 51
+    },
+    "Engineered stone": {
+      "accuracy": 0.20224719101123595,
+      "iou": 0.15384615384615385,
+      "support": 89
+    },
+    "Fabric/cloth": {
+      "accuracy": 0.7728515969648844,
+      "iou": 0.7065537406735228,
+      "support": 22668
+    },
+    "Fiberglass wool": {
+      "accuracy": 0.3,
+      "iou": 0.24,
+      "support": 20
+    },
+    "Fire": {
+      "accuracy": 0.76,
+      "iou": 0.5757575757575758,
+      "support": 50
+    },
+    "Foliage": {
+      "accuracy": 0.8998444185141968,
+      "iou": 0.7700116491928773,
+      "support": 5142
+    },
+    "Food": {
+      "accuracy": 0.8614525139664805,
+      "iou": 0.622778675282714,
+      "support": 895
+    },
+    "Fur": {
+      "accuracy": 0.8259860788863109,
+      "iou": 0.6819923371647509,
+      "support": 431
+    },
+    "Gemstone/quartz": {
+      "accuracy": 0.54,
+      "iou": 0.1956521739130435,
+      "support": 50
+    },
+    "Glass": {
+      "accuracy": 0.7519200887902331,
+      "iou": 0.6554820232981152,
+      "support": 22525
+    },
+    "Hair": {
+      "accuracy": 0.975366568914956,
+      "iou": 0.9131091283459163,
+      "support": 6820
+    },
+    "I cannot tell": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Ice": {
+      "accuracy": 0.41818181818181815,
+      "iou": 0.35384615384615387,
+      "support": 55
+    },
+    "Leather": {
+      "accuracy": 0.697282099343955,
+      "iou": 0.4206955046649703,
+      "support": 2134
+    },
+    "Liquid, non-water": {
+      "accuracy": 0.4166666666666667,
+      "iou": 0.26785714285714285,
+      "support": 36
+    },
+    "Metal": {
+      "accuracy": 0.8197604132322234,
+      "iou": 0.5327476608813656,
+      "support": 9099
+    },
+    "Mirror": {
+      "accuracy": 0.6637390213299874,
+      "iou": 0.40350877192982454,
+      "support": 797
+    },
+    "Not on list": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Paint/plaster/enamel": {
+      "accuracy": 0.5977813393670847,
+      "iou": 0.5230278533916247,
+      "support": 31821
+    },
+    "Paper": {
+      "accuracy": 0.8264638245179551,
+      "iou": 0.6788724208079047,
+      "support": 11306
+    },
+    "Pearl": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 1
+    },
+    "Photograph/painting": {
+      "accuracy": 0.6998202516476932,
+      "iou": 0.36671899529042384,
+      "support": 1669
+    },
+    "Plastic, clear": {
+      "accuracy": 0.6641744548286604,
+      "iou": 0.42554890219560876,
+      "support": 1605
+    },
+    "Plastic, non-clear": {
+      "accuracy": 0.6261146496815286,
+      "iou": 0.5067590789322947,
+      "support": 14130
+    },
+    "Rubber/latex": {
+      "accuracy": 0.581547064305685,
+      "iou": 0.31060228969636633,
+      "support": 1073
+    },
+    "Sand": {
+      "accuracy": 0.29545454545454547,
+      "iou": 0.23853211009174313,
+      "support": 88
+    },
+    "Skin/lips": {
+      "accuracy": 0.9690329131060157,
+      "iou": 0.9340238134072745,
+      "support": 11819
+    },
+    "Sky": {
+      "accuracy": 0.9326833797585887,
+      "iou": 0.6621621621621622,
+      "support": 2154
+    },
+    "Snow": {
+      "accuracy": 0.6788990825688074,
+      "iou": 0.5211267605633803,
+      "support": 109
+    },
+    "Soap": {
+      "accuracy": 0.14285714285714285,
+      "iou": 0.14285714285714285,
+      "support": 7
+    },
+    "Soil/mud": {
+      "accuracy": 0.634703196347032,
+      "iou": 0.3926553672316384,
+      "support": 657
+    },
+    "Sponge": {
+      "accuracy": 0.1111111111111111,
+      "iou": 0.09090909090909091,
+      "support": 18
+    },
+    "Stone, natural": {
+      "accuracy": 0.6101556101556102,
+      "iou": 0.4328878558977339,
+      "support": 1221
+    },
+    "Stone, polished": {
+      "accuracy": 0.41022280471821754,
+      "iou": 0.25385239253852393,
+      "support": 763
+    },
+    "Styrofoam": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 31
+    },
+    "Tile": {
+      "accuracy": 0.6326317620455731,
+      "iou": 0.4346838771363691,
+      "support": 5749
+    },
+    "Wallpaper": {
+      "accuracy": 0.5758928571428571,
+      "iou": 0.33740191804707936,
+      "support": 672
+    },
+    "Water": {
+      "accuracy": 0.6458569807037458,
+      "iou": 0.5158658204895739,
+      "support": 881
+    },
+    "Wax": {
+      "accuracy": 0.8333333333333334,
+      "iou": 0.6129032258064516,
+      "support": 114
+    },
+    "Whiteboard": {
+      "accuracy": 0.6366366366366366,
+      "iou": 0.4491525423728814,
+      "support": 333
+    },
+    "Wicker": {
+      "accuracy": 0.7818181818181819,
+      "iou": 0.5789473684210527,
+      "support": 605
+    },
+    "Wood": {
+      "accuracy": 0.7983939413723314,
+      "iou": 0.6174079870752764,
+      "support": 15317
+    },
+    "Wood, tree": {
+      "accuracy": 0.717687074829932,
+      "iou": 0.49414519906323184,
+      "support": 588
+    },
+    "Bad polygon": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Multiple materials": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Asphalt": {
+      "accuracy": 0.4797979797979798,
+      "iou": 0.296875,
+      "support": 198
+    }
+  }
+}

DMS/dinov3_siglip2_lmf_reg/model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fcbfcb811c95b16bba82f2765cc37aeb0861ea6067e559a36c605bfdc24f1ba0
+size 7701584

DMS/dinov3_siglip2_lmf_reg/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed2fc3b7ee4dd92b575b199ea2770e3a0e0d702355dfe5f97216e50917c68cab
+size 7695412

DMS/dinov3_siglip2_mlp/README.md

@@ -0,0 +1,52 @@
+# DINOv3-SigLIP2 + Concat + MLP on Apple Dense Material Segmentation
+
+A material-classification **fusion head** from SegObjectMaterial. It classifies the
+material of a masked object from two frozen embeddings:
+
+| | |
+|---|---|
+| vision | DINOv3 ViT-L/16 (CLS token, frozen) -> 1024-d |
+| language | SigLIP2 base text encoder over Qwen3-VL-8B captions (frozen) -> 768-d |
+| head | Concat + MLP (`mlp`, defaults), 1,065,529 parameters |
+| classes | 57 (Apple Dense Material Segmentation) |
+
+**This repo holds only the head.** Features must come from the same encoders,
+captioner and prompts as training (see the SegObjectMaterial repository); the
+encoders are downloaded from their own model pages under their own licenses.
+
+## Evaluation
+
+Held-out split recorded in `split.json` (source images never seen in training).
+Single training run -- differences under ~0.5 mIoU points between heads are within
+run-to-run noise.
+
+| mIoU | mAcc | eval samples |
+|---|---|---|
+| 0.4592 | 0.6453 | 186,231 |
+
+## Training
+
+```json
+{
+  "learning_rate": 5e-05,
+  "epochs": 50,
+  "batch_size": 32,
+  "weight_decay": 0.0001,
+  "lr_schedule": "constant",
+  "early_stopping_patience": 0,
+  "checkpoint": "best validation loss"
+}
+```
+
+## Use
+
+```python
+from utils.hub import load_head_from_hub      # SegObjectMaterial repository
+head, info = load_head_from_hub("s0um0r0y/SegObjectMaterial", "DMS/dinov3_siglip2_mlp")
+logits = head(vision_features, language_features)   # [B, 1024], [B, 768]
+```
+
+## Integrity
+
+- `model.safetensors` sha256 `24b04b492f455d44e9ee3dc13c98bade94181ec60d0ce7363c6e3ff0e9996341`
+- source config `config/config_latest_models_mlp.toml`, code commit `1a8130c922267fb0786a33e3854a7e00763cc38f+dirty`

DMS/dinov3_siglip2_mlp/config.toml

@@ -0,0 +1,64 @@
+# MLP-head twin of config/config.toml.
+# SAME embedding cache -- the head type does not change the cached features, so
+# nothing needs re-serializing. Only the checkpoint, TensorBoard dir and W&B run
+# differ, so this run can never overwrite the gated checkpoint.
+
+# Apple DMS -- LATEST-MODELS pipeline: DINOv3 ViT-L/16 (1024) + Qwen3-VL -> SigLIP2 text (768)
+#
+# Pair this with:
+#   src/vision_stream_latest_models.py + src/language_stream_latest_models.py
+#   run_serialization_latest_models.py  ->  train_latest_models.py
+#
+# For the base DINOv2/CLIP pipeline (run_serialization.py -> train.py) use
+# config/config_dinov2_clip.toml instead -- its vision_dim is 768, not 1024.
+
+[data]
+dataset_type = "apple_dms"
+image_dir = "<local>/images"
+mask_dir = "<local>/labels"
+# Rebuilt automatically on first run; the index now stores one entry per
+# CONNECTED COMPONENT, so any pre-existing labels.json must be removed first.
+labels_file = "<local>/labels_instances.json"
+# Separate cache: DINOv3 features are 1024-d and must not mix with the
+# 768-d DINOv2 cache written by run_serialization.py.
+tensor_cache_dir = "<local>/tensor_cache_dir_latest_models"
+
+# Connected-component filtering (measured on 400 real DMS masks):
+# keeping every component gives 7.5x more samples than the old per-(image,class)
+# scheme -- 2.45M samples / 47 GB / ~7.5x the VLM captioning time -- because 69%
+# of components are under 0.1% of the frame and crop to a few pixels.
+# A component is kept when area >= max(min_instance_area, min_instance_frac * frame).
+min_instance_area = 1000        # absolute pixel floor
+min_instance_frac = 0.001       # and at least 0.1% of the frame
+max_instances_per_class = 0     # 0 = keep all; set e.g. 3 to cap the long tail
+
+
+[model]
+# fusion head: "gated" (GMU) or "mlp" (concat+MLP). Read by train_latest_models.py; a contradicting --fusion flag is an error, so checkpoint_path can't hold the wrong head.
+fusion = "mlp"
+vision_dim = 1024     # DINOv3 ViT-L/16 CLS token
+language_dim = 768    # SigLIP2 base text embedding
+num_classes = 57
+
+[training]
+batch_size = 32
+
+# Batch size for the OFFLINE SERIALIZATION pass, which is a completely different
+# memory regime from training: it pushes full-resolution images through the 7B
+# VLM, whereas [training].batch_size only moves cached 768-d vectors through a
+# small MLP. Measured on the 23.6 GiB 4090 -- static weights 16.2 GiB, so only
+# ~7.4 GiB is left: batch 8 peaks at 17.7 GiB, 16 at 19.2 GiB, and 32 OOMs.
+# 8 leaves headroom for a labmate's job to coexist; raise to 16 on a free card.
+serialization_batch_size = 8
+learning_rate = 5e-5
+epochs = 50
+train_ratio = 0.8  # 80% for training
+val_ratio = 0.2    # 20% for validation
+tensorboard_log_dir = "<local>/dms_dinov3_siglip2_mlp"
+
+[inference]
+checkpoint_path = "<local>/best_model_dms_latest_mlp.pth"
+
+[logging]
+wandb_project = "material-classification"
+wandb_run_name = "dms-dinov3-siglip2-mlp"

DMS/dinov3_siglip2_mlp/head.json

@@ -0,0 +1,99 @@
+{
+  "fusion": "mlp",
+  "head": "Concat + MLP",
+  "head_kwargs": {},
+  "vision_dim": 1024,
+  "language_dim": 768,
+  "num_classes": 57,
+  "classes": [
+    "No label",
+    "Animal skin",
+    "Bone/teeth/horn",
+    "Brickwork",
+    "Cardboard",
+    "Carpet/rug",
+    "Ceiling tile",
+    "Ceramic",
+    "Chalkboard/blackboard",
+    "Clutter",
+    "Concrete",
+    "Cork/corkboard",
+    "Engineered stone",
+    "Fabric/cloth",
+    "Fiberglass wool",
+    "Fire",
+    "Foliage",
+    "Food",
+    "Fur",
+    "Gemstone/quartz",
+    "Glass",
+    "Hair",
+    "I cannot tell",
+    "Ice",
+    "Leather",
+    "Liquid, non-water",
+    "Metal",
+    "Mirror",
+    "Not on list",
+    "Paint/plaster/enamel",
+    "Paper",
+    "Pearl",
+    "Photograph/painting",
+    "Plastic, clear",
+    "Plastic, non-clear",
+    "Rubber/latex",
+    "Sand",
+    "Skin/lips",
+    "Sky",
+    "Snow",
+    "Soap",
+    "Soil/mud",
+    "Sponge",
+    "Stone, natural",
+    "Stone, polished",
+    "Styrofoam",
+    "Tile",
+    "Wallpaper",
+    "Water",
+    "Wax",
+    "Whiteboard",
+    "Wicker",
+    "Wood",
+    "Wood, tree",
+    "Bad polygon",
+    "Multiple materials",
+    "Asphalt"
+  ],
+  "dataset": "Apple Dense Material Segmentation",
+  "dataset_key": "DMS",
+  "encoders": {
+    "vision": "DINOv3 ViT-L/16 (CLS token, frozen)",
+    "language": "SigLIP2 base text encoder over Qwen3-VL-8B captions (frozen)",
+    "short": "DINOv3-SigLIP2",
+    "slug": "dinov3_siglip2"
+  },
+  "parameters": 1065529,
+  "training": {
+    "learning_rate": 5e-05,
+    "epochs": 50,
+    "batch_size": 32,
+    "weight_decay": 0.0001,
+    "lr_schedule": "constant",
+    "early_stopping_patience": 0,
+    "checkpoint": "best validation loss"
+  },
+  "sha256": {
+    "safetensors": "24b04b492f455d44e9ee3dc13c98bade94181ec60d0ce7363c6e3ff0e9996341",
+    "pth": "025001457a82b8db2365a08e3691fc9dd6594978b8c8467deb6bbefe77bfdda5"
+  },
+  "source": {
+    "config": "config/config_latest_models_mlp.toml",
+    "git_commit": "1a8130c922267fb0786a33e3854a7e00763cc38f+dirty",
+    "checkpoint_mtime": "2026-10-05T17:16:25"
+  },
+  "metrics": {
+    "mIoU": 0.45922884101465977,
+    "mAcc": 0.6453263039159061,
+    "eval_samples": 186231
+  }
+}

DMS/dinov3_siglip2_mlp/metrics.json

@@ -0,0 +1,296 @@
+{
+  "dataset": "DMS",
+  "checkpoint": "<local>/best_model_dms_latest_mlp.pth",
+  "fusion_type": "mlp",
+  "num_eval_samples": 186231,
+  "mIoU": 0.45922884101465977,
+  "mAcc": 0.6453263039159061,
+  "per_class": {
+    "No label": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Animal skin": {
+      "accuracy": 0.6228070175438597,
+      "iou": 0.4303030303030303,
+      "support": 114
+    },
+    "Bone/teeth/horn": {
+      "accuracy": 0.7272727272727273,
+      "iou": 0.6153846153846154,
+      "support": 44
+    },
+    "Brickwork": {
+      "accuracy": 0.743521000893655,
+      "iou": 0.5187032418952618,
+      "support": 1119
+    },
+    "Cardboard": {
+      "accuracy": 0.7536656891495601,
+      "iou": 0.47155963302752296,
+      "support": 1023
+    },
+    "Carpet/rug": {
+      "accuracy": 0.8107396149949342,
+      "iou": 0.46921543332942417,
+      "support": 4935
+    },
+    "Ceiling tile": {
+      "accuracy": 0.8038229376257545,
+      "iou": 0.4615829000577701,
+      "support": 994
+    },
+    "Ceramic": {
+      "accuracy": 0.8287461773700305,
+      "iou": 0.613915857605178,
+      "support": 2289
+    },
+    "Chalkboard/blackboard": {
+      "accuracy": 0.7348837209302326,
+      "iou": 0.5683453237410072,
+      "support": 215
+    },
+    "Clutter": {
+      "accuracy": 0.49056603773584906,
+      "iou": 0.3291139240506329,
+      "support": 53
+    },
+    "Concrete": {
+      "accuracy": 0.5108827085852479,
+      "iou": 0.2779605263157895,
+      "support": 1654
+    },
+    "Cork/corkboard": {
+      "accuracy": 0.3333333333333333,
+      "iou": 0.2236842105263158,
+      "support": 51
+    },
+    "Engineered stone": {
+      "accuracy": 0.20224719101123595,
+      "iou": 0.11612903225806452,
+      "support": 89
+    },
+    "Fabric/cloth": {
+      "accuracy": 0.7791159343568025,
+      "iou": 0.7166741062370653,
+      "support": 22668
+    },
+    "Fiberglass wool": {
+      "accuracy": 0.3,
+      "iou": 0.17647058823529413,
+      "support": 20
+    },
+    "Fire": {
+      "accuracy": 0.86,
+      "iou": 0.5657894736842105,
+      "support": 50
+    },
+    "Foliage": {
+      "accuracy": 0.91170750680669,
+      "iou": 0.7855227882037533,
+      "support": 5142
+    },
+    "Food": {
+      "accuracy": 0.8782122905027933,
+      "iou": 0.6374695863746959,
+      "support": 895
+    },
+    "Fur": {
+      "accuracy": 0.8283062645011601,
+      "iou": 0.697265625,
+      "support": 431
+    },
+    "Gemstone/quartz": {
+      "accuracy": 0.66,
+      "iou": 0.23741007194244604,
+      "support": 50
+    },
+    "Glass": {
+      "accuracy": 0.7552497225305217,
+      "iou": 0.6632617256033374,
+      "support": 22525
+    },
+    "Hair": {
+      "accuracy": 0.976099706744868,
+      "iou": 0.9226611226611227,
+      "support": 6820
+    },
+    "I cannot tell": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Ice": {
+      "accuracy": 0.5818181818181818,
+      "iou": 0.47058823529411764,
+      "support": 55
+    },
+    "Leather": {
+      "accuracy": 0.7282099343955014,
+      "iou": 0.4223973905952704,
+      "support": 2134
+    },
+    "Liquid, non-water": {
+      "accuracy": 0.4444444444444444,
+      "iou": 0.2857142857142857,
+      "support": 36
+    },
+    "Metal": {
+      "accuracy": 0.8286624903835587,
+      "iou": 0.5417445035206208,
+      "support": 9099
+    },
+    "Mirror": {
+      "accuracy": 0.6976160602258469,
+      "iou": 0.40942562592047127,
+      "support": 797
+    },
+    "Not on list": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Paint/plaster/enamel": {
+      "accuracy": 0.5956443857829735,
+      "iou": 0.5289685197588747,
+      "support": 31821
+    },
+    "Paper": {
+      "accuracy": 0.8389350787192641,
+      "iou": 0.6866719756750886,
+      "support": 11306
+    },
+    "Pearl": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 1
+    },
+    "Photograph/painting": {
+      "accuracy": 0.7231875374475734,
+      "iou": 0.37542768273716953,
+      "support": 1669
+    },
+    "Plastic, clear": {
+      "accuracy": 0.6890965732087228,
+      "iou": 0.41925701288855194,
+      "support": 1605
+    },
+    "Plastic, non-clear": {
+      "accuracy": 0.6375796178343949,
+      "iou": 0.5135088919288646,
+      "support": 14130
+    },
+    "Rubber/latex": {
+      "accuracy": 0.625349487418453,
+      "iou": 0.330704780680138,
+      "support": 1073
+    },
+    "Sand": {
+      "accuracy": 0.3181818181818182,
+      "iou": 0.24561403508771928,
+      "support": 88
+    },
+    "Skin/lips": {
+      "accuracy": 0.978847618241814,
+      "iou": 0.9362304766529093,
+      "support": 11819
+    },
+    "Sky": {
+      "accuracy": 0.9220055710306406,
+      "iou": 0.6768916155419223,
+      "support": 2154
+    },
+    "Snow": {
+      "accuracy": 0.6972477064220184,
+      "iou": 0.4342857142857143,
+      "support": 109
+    },
+    "Soap": {
+      "accuracy": 0.14285714285714285,
+      "iou": 0.14285714285714285,
+      "support": 7
+    },
+    "Soil/mud": {
+      "accuracy": 0.6636225266362252,
+      "iou": 0.40671641791044777,
+      "support": 657
+    },
+    "Sponge": {
+      "accuracy": 0.3888888888888889,
+      "iou": 0.2916666666666667,
+      "support": 18
+    },
+    "Stone, natural": {
+      "accuracy": 0.6445536445536445,
+      "iou": 0.44040290990486847,
+      "support": 1221
+    },
+    "Stone, polished": {
+      "accuracy": 0.4167758846657929,
+      "iou": 0.2550120288692863,
+      "support": 763
+    },
+    "Styrofoam": {
+      "accuracy": 0.16129032258064516,
+      "iou": 0.1111111111111111,
+      "support": 31
+    },
+    "Tile": {
+      "accuracy": 0.6460253957209949,
+      "iou": 0.43668430335097,
+      "support": 5749
+    },
+    "Wallpaper": {
+      "accuracy": 0.5967261904761905,
+      "iou": 0.348695652173913,
+      "support": 672
+    },
+    "Water": {
+      "accuracy": 0.6730987514188422,
+      "iou": 0.5188101487314086,
+      "support": 881
+    },
+    "Wax": {
+      "accuracy": 0.8771929824561403,
+      "iou": 0.6535947712418301,
+      "support": 114
+    },
+    "Whiteboard": {
+      "accuracy": 0.6906906906906907,
+      "iou": 0.4536489151873767,
+      "support": 333
+    },
+    "Wicker": {
+      "accuracy": 0.8132231404958677,
+      "iou": 0.5934861278648975,
+      "support": 605
+    },
+    "Wood": {
+      "accuracy": 0.7876215969184566,
+      "iou": 0.6349807884625507,
+      "support": 15317
+    },
+    "Wood, tree": {
+      "accuracy": 0.7363945578231292,
+      "iou": 0.4959908361970218,
+      "support": 588
+    },
+    "Bad polygon": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Multiple materials": {
+      "accuracy": 0.0,
+      "iou": 0.0,
+      "support": 0
+    },
+    "Asphalt": {
+      "accuracy": 0.5,
+      "iou": 0.32038834951456313,
+      "support": 198
+    }
+  },
+  "config": "config/config_latest_models_mlp.toml"
+}

DMS/dinov3_siglip2_mlp/model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:025001457a82b8db2365a08e3691fc9dd6594978b8c8467deb6bbefe77bfdda5
+size 4274254

DMS/dinov3_siglip2_mlp/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24b04b492f455d44e9ee3dc13c98bade94181ec60d0ce7363c6e3ff0e9996341
+size 4269548

FMD/dinov2_clip_gated/README.md

@@ -0,0 +1,52 @@
+# DINOv2-CLIP + Gated Multimodal Unit on Flickr Material Database
+
+A material-classification **fusion head** from SegObjectMaterial. It classifies the
+material of a masked object from two frozen embeddings:
+
+| | |
+|---|---|
+| vision | DINOv2 ViT-B/14 (CLS token, frozen) -> 768-d |
+| language | CLIP ViT-L/14 text encoder over Qwen2.5-VL-7B captions (frozen) -> 768-d |
+| head | Gated Multimodal Unit (`gated`, defaults), 1,710,858 parameters |
+| classes | 10 (Flickr Material Database) |
+
+**This repo holds only the head.** Features must come from the same encoders,
+captioner and prompts as training (see the SegObjectMaterial repository); the
+encoders are downloaded from their own model pages under their own licenses.
+
+## Evaluation
+
+Held-out split recorded in `split.json` (source images never seen in training).
+Single training run -- differences under ~0.5 mIoU points between heads are within
+run-to-run noise.
+
+| mIoU | mAcc | eval samples |
+|---|---|---|
+| 0.9234 | 0.9599 | 200 |
+
+## Training
+
+```json
+{
+  "learning_rate": 5e-05,
+  "epochs": 50,
+  "batch_size": 8,
+  "weight_decay": 0.0001,
+  "lr_schedule": "constant",
+  "early_stopping_patience": 0,
+  "checkpoint": "best validation loss"
+}
+```
+
+## Use
+
+```python
+from utils.hub import load_head_from_hub      # SegObjectMaterial repository
+head, info = load_head_from_hub("s0um0r0y/SegObjectMaterial", "FMD/dinov2_clip_gated")
+logits = head(vision_features, language_features)   # [B, 768], [B, 768]
+```
+
+## Integrity
+
+- `model.safetensors` sha256 `de0b96447b176c5791b4cda3b6121e852bf488943c1a7d6fe5c93aef923fe129`
+- source config `config/fmd_config.toml`, code commit `1a8130c922267fb0786a33e3854a7e00763cc38f+dirty`