feat: Claude/yolo semantic segmentation mcwhh8

configs/AeroScapes_info.json

@@ -0,0 +1,122 @@
+[
+  {
+    "hasInstances": false,
+    "category": "void",
+    "catid": 0,
+    "name": "Background",
+    "ignoreInEval": false,
+    "id": 0,
+    "color": [0, 0, 0],
+    "trainId": 0
+  },
+  {
+    "hasInstances": false,
+    "category": "human",
+    "catid": 1,
+    "name": "Person",
+    "ignoreInEval": false,
+    "id": 1,
+    "color": [192, 128, 128],
+    "trainId": 1
+  },
+  {
+    "hasInstances": false,
+    "category": "vehicle",
+    "catid": 2,
+    "name": "Bike",
+    "ignoreInEval": false,
+    "id": 2,
+    "color": [0, 128, 0],
+    "trainId": 2
+  },
+  {
+    "hasInstances": false,
+    "category": "vehicle",
+    "catid": 2,
+    "name": "Car",
+    "ignoreInEval": false,
+    "id": 3,
+    "color": [128, 128, 128],
+    "trainId": 3
+  },
+  {
+    "hasInstances": false,
+    "category": "vehicle",
+    "catid": 2,
+    "name": "Drone",
+    "ignoreInEval": false,
+    "id": 4,
+    "color": [128, 0, 0],
+    "trainId": 4
+  },
+  {
+    "hasInstances": false,
+    "category": "vehicle",
+    "catid": 2,
+    "name": "Boat",
+    "ignoreInEval": false,
+    "id": 5,
+    "color": [0, 0, 128],
+    "trainId": 5
+  },
+  {
+    "hasInstances": false,
+    "category": "animal",
+    "catid": 3,
+    "name": "Animal",
+    "ignoreInEval": false,
+    "id": 6,
+    "color": [192, 0, 128],
+    "trainId": 6
+  },
+  {
+    "hasInstances": false,
+    "category": "void",
+    "catid": 0,
+    "name": "Obstacle",
+    "ignoreInEval": false,
+    "id": 7,
+    "color": [192, 0, 0],
+    "trainId": 7
+  },
+  {
+    "hasInstances": false,
+    "category": "construction",
+    "catid": 4,
+    "name": "Construction",
+    "ignoreInEval": false,
+    "id": 8,
+    "color": [192, 128, 0],
+    "trainId": 8
+  },
+  {
+    "hasInstances": false,
+    "category": "nature",
+    "catid": 5,
+    "name": "Vegetation",
+    "ignoreInEval": false,
+    "id": 9,
+    "color": [0, 64, 0],
+    "trainId": 9
+  },
+  {
+    "hasInstances": false,
+    "category": "flat",
+    "catid": 6,
+    "name": "Road",
+    "ignoreInEval": false,
+    "id": 10,
+    "color": [128, 128, 0],
+    "trainId": 10
+  },
+  {
+    "hasInstances": false,
+    "category": "nature",
+    "catid": 5,
+    "name": "Sky",
+    "ignoreInEval": false,
+    "id": 11,
+    "color": [0, 128, 128],
+    "trainId": 11
+  }
+]

configs/dataset/aeroscapes.yaml

@@ -0,0 +1,48 @@
+# Dataset Configuration for CABiNet training on AeroScapes
+#
+# dataset_path must point at the PRE-CONVERTED dataset produced by
+# src/scripts/convert_aeroscapes_to_yolo.py:
+#
+#   <dataset_path>/
+#   ├── images/{train,val}/   ← RGB JPEGs (uniformly 1280x720)
+#   └── masks/{train,val}/    ← single-channel PNGs, pixel value = class ID
+#
+# This is NOT the raw AeroScapes distribution (JPEGImages/, SegmentationClass/,
+# ImageSets/) — convert it first:
+#   python src/scripts/convert_aeroscapes_to_yolo.py \
+#       --src /path/to/raw/aeroscapes --dst /path/to/converted
+#
+# All 12 classes are valid trainIds 0-11 — see configs/AeroScapes_info.json.
+# No colour palette or lookup table is needed at load time; the conversion
+# step's source masks were already single-channel class IDs.
+#
+# Unlike UAVid, AeroScapes source images are uniformly 1280x720, so there is
+# no batch_size=1 validation constraint here — standard batched validation
+# works.
+#
+# NOTE: the source dataset has no test split (only ImageSets/trn.txt and
+# val.txt) — the converter only ever produces train/val.
+
+name: aeroscapes
+num_classes: 12 # Background(0) … Sky(11); all 12 classes valid, none ignored
+cropsize: [720, 720] # native res is 1280x720 (uniform) — smaller than UAVid's 3840x2160
+dataset_path: ${oc.env:AEROSCAPES_YOLO_ROOT,''} # path to the CONVERTED dataset root
+ignore_idx: 255 # label for genuinely unrecognized pixel values (should not occur)
+seed: 0
+
+# Class weights are computed at training startup from a sample of the
+# training set via training_config.cls_pw (see configs/train.yaml) — same
+# mechanism as uavid.yaml, not a static per-dataset table.
+
+# Training augmentation — same knobs/defaults as configs/dataset/uavid.yaml;
+# AeroScapes is also UAV-captured imagery so the same recipe applies.
+augmentation:
+  degrees: 10.0 # rotation ±10° (matches RandomRotate)
+  translate: 0.05 # small translation, fraction of image size
+  scale: 0.3 # scale jitter ±30% (continuous)
+  flipud: 0.2 # vertical flip — AeroScapes includes top-down and oblique views
+  fliplr: 0.5 # horizontal flip
+  hsv_h: 0.01 # small hue shift (controlled imaging conditions)
+  hsv_s: 0.4 # saturation jitter
+  hsv_v: 0.3 # brightness/value jitter
+  mixup: 0.1 # probability of blending with a second random sample

configs/dataset/aeroscapes_yolo.yaml

@@ -0,0 +1,72 @@
+# AeroScapes dataset configuration for Ultralytics YOLO semantic segmentation
+#
+# Format: https://docs.ultralytics.com/datasets/semantic/
+# Masks:  single-channel PNG, pixel value = class index (0-11). 255 remains
+#         reserved for genuinely unrecognized values.
+#
+# Pre-processing required
+# -----------------------
+# Convert the raw AeroScapes distribution (JPEGImages/, SegmentationClass/,
+# ImageSets/) to the shared images/+masks/ layout first:
+#
+#   python src/scripts/convert_aeroscapes_to_yolo.py \
+#       --src /path/to/aeroscapes \
+#       --dst /path/to/aeroscapes_yolo
+#
+# Then set the ``path`` field below to /path/to/aeroscapes_yolo.
+#
+# Class mapping (all 12 classes are valid; none are ignored)
+# -----------------------------------------------------------
+#   0  Background     [  0,   0,   0]
+#   1  Person         [192, 128, 128]
+#   2  Bike           [  0, 128,   0]
+#   3  Car            [128, 128, 128]
+#   4  Drone          [128,   0,   0]
+#   5  Boat           [  0,   0, 128]
+#   6  Animal         [192,   0, 128]
+#   7  Obstacle       [192,   0,   0]
+#   8  Construction   [192, 128,   0]
+#   9  Vegetation     [  0,  64,   0]
+#  10  Road           [128, 128,   0]
+#  11  Sky            [  0, 128, 128]
+
+# Root path. Ultralytics parses this file with plain YAML (no Hydra/OmegaConf
+# env-var resolution), so this must be a literal path — edit it directly for
+# your machine, or point `data=` at a copy of this file with your own path.
+path: /home/neural_debugger/Downloads/datasets/AeroScapes/AeroScapes-YOLO
+
+train: images/train # relative to path
+val: images/val # relative to path
+# NOTE: no 'test:' key — the source AeroScapes distribution has no test
+# split (only ImageSets/trn.txt and val.txt); the converter never produces one.
+
+# Mask directory (must mirror images/ structure): images/train -> masks/train, etc.
+masks_dir: masks # relative to path
+
+# Number of semantic classes (all 12 AeroScapes classes)
+nc: 12
+
+# Class names — order must match the integer class IDs above
+names:
+  0: Background
+  1: Person
+  2: Bike
+  3: Car
+  4: Drone
+  5: Boat
+  6: Animal
+  7: Obstacle
+  8: Construction
+  9: Vegetation
+  10: Road
+  11: Sky
+
+# Ultralytics training command example
+# -------------------------------------
+# yolo semantic train \
+#     model=yolo26n-sem.pt \
+#     data=configs/dataset/aeroscapes_yolo.yaml \
+#     imgsz=896 \
+#     epochs=100 \
+#     batch=8 \
+#     device=0

configs/dataset/vdd.yaml

@@ -0,0 +1,47 @@
+# Dataset Configuration for CABiNet training on VDD (Varied Drone Dataset)
+#
+# dataset_path must point at the PRE-CONVERTED dataset produced by
+# src/scripts/convert_vdd_to_yolo.py:
+#
+#   <dataset_path>/
+#   ├── images/{train,val,test}/   ← RGB JPEGs (uniformly 4000x3000)
+#   └── masks/{train,val,test}/    ← single-channel PNGs, pixel value = class ID
+#
+# This is NOT the raw VDD distribution (train/val/test with src/, gt/
+# sub-dirs) — convert it first:
+#   python src/scripts/convert_vdd_to_yolo.py \
+#       --src /path/to/raw/VDD --dst /path/to/converted
+#
+# All 7 classes are valid trainIds 0-6 — see https://github.com/RussRobin/VDD.
+# No colour palette or lookup table is needed at load time; the source masks
+# were already single-channel class IDs.
+#
+# Like AeroScapes and unlike UAVid, VDD source images are uniformly
+# 4000x3000, so there is no batch_size=1 validation constraint here —
+# standard batched validation works. Unlike AeroScapes, VDD ships with a
+# real train/val/test split already defined by the source distribution.
+
+name: vdd
+num_classes: 7 # Other(0), Wall(1), Road(2), Vegetation(3), Vehicle(4), Roof(5), Water(6)
+cropsize: [1024, 1024] # native res is 4000x3000 (uniform) — much larger than UAVid/AeroScapes
+dataset_path: ${oc.env:VDD_YOLO_ROOT,''} # path to the CONVERTED dataset root
+ignore_idx: 255 # label for genuinely unrecognized pixel values (should not occur)
+seed: 0
+
+# Class weights are computed at training startup from a sample of the
+# training set via training_config.cls_pw (see configs/train.yaml) — same
+# mechanism as uavid.yaml/aeroscapes.yaml, not a static per-dataset table.
+
+# Training augmentation — same knobs/defaults as configs/dataset/uavid.yaml
+# and aeroscapes.yaml; VDD is also UAV-captured imagery so the same recipe
+# applies.
+augmentation:
+  degrees: 10.0 # rotation ±10° (matches RandomRotate)
+  translate: 0.05 # small translation, fraction of image size
+  scale: 0.3 # scale jitter ±30% (continuous)
+  flipud: 0.2 # vertical flip — VDD includes top-down and oblique views
+  fliplr: 0.5 # horizontal flip
+  hsv_h: 0.01 # small hue shift (controlled imaging conditions)
+  hsv_s: 0.4 # saturation jitter
+  hsv_v: 0.3 # brightness/value jitter
+  mixup: 0.1 # probability of blending with a second random sample

configs/dataset/vdd_yolo.yaml

@@ -0,0 +1,61 @@
+# VDD (Varied Drone Dataset) configuration for Ultralytics YOLO semantic segmentation
+#
+# Format: https://docs.ultralytics.com/datasets/semantic/
+# Masks:  single-channel PNG, pixel value = class index (0-6). 255 remains
+#         reserved for genuinely unrecognized values.
+#
+# Pre-processing required
+# -----------------------
+# Convert the raw VDD distribution (train/val/test, each with src/ and gt/)
+# to the shared images/+masks/ layout first:
+#
+#   python src/scripts/convert_vdd_to_yolo.py \
+#       --src /path/to/VDD \
+#       --dst /path/to/vdd_yolo
+#
+# Then set the ``path`` field below to /path/to/vdd_yolo.
+#
+# Class mapping (all 7 classes are valid; none are ignored)
+# -----------------------------------------------------------
+#   0  Other        (background clutter)
+#   1  Wall
+#   2  Road
+#   3  Vegetation
+#   4  Vehicle
+#   5  Roof
+#   6  Water
+
+# Root path. Ultralytics parses this file with plain YAML (no Hydra/OmegaConf
+# env-var resolution), so this must be a literal path — edit it directly for
+# your machine, or point `data=` at a copy of this file with your own path.
+path: /home/neural_debugger/Downloads/datasets/VDD/VDD-YOLO
+
+train: images/train # relative to path
+val: images/val # relative to path
+test: images/test # relative to path
+
+# Mask directory (must mirror images/ structure): images/train -> masks/train, etc.
+masks_dir: masks # relative to path
+
+# Number of semantic classes
+nc: 7
+
+# Class names — order must match the integer class IDs above
+names:
+  0: Other
+  1: Wall
+  2: Road
+  3: Vegetation
+  4: Vehicle
+  5: Roof
+  6: Water
+
+# Ultralytics training command example
+# -------------------------------------
+# yolo semantic train \
+#     model=yolo26n-sem.pt \
+#     data=configs/dataset/vdd_yolo.yaml \
+#     imgsz=1024 \
+#     epochs=120 \
+#     batch=4 \
+#     device=0

configs/train.yaml

@@ -18,6 +18,17 @@ training_config:
   optimizer_power: 0.9
   warmup_steps: 4000
   warmup_start_lr: 1e-5
+  # Warm-start model weights from a full CABiNet checkpoint trained on a
+  # (possibly different) dataset — e.g. finetune AeroScapes/VDD from a
+  # UAVid-trained checkpoint, which converges much faster than
+  # backbone-only (ImageNet) init. Only parameters whose name AND shape
+  # match are loaded — the two classifier heads (sized by num_classes) are
+  # skipped when they differ, everything else (backbone/CAB/spatial/fusion)
+  # transfers. Independent of `resume` below: this never restores
+  # optimizer/EMA/epoch state, it's a fresh training run with warm-started
+  # weights. Accepts either a full training checkpoint (checkpoint_last.pth)
+  # or a raw model state_dict (*_best.pth).
+  pretrained_ckpt_path: null
   # max_iterations is in OPTIMIZER STEPS (not batches).
   # train.py computes this as ceil(epochs * batches_per_epoch / accum_steps).
   # Setting it here is optional — if absent, train.py derives it automatically.
@@ -34,7 +45,7 @@ training_config:
 
   max_grad_norm: 1.0 # Gradient clipping threshold; set to null/0 to disable
   log_iter: 20 # Renamed from 'msg_iterations' — clearer intent
-  experiments_path: experiments/cabinet_experiments_large_1024x1024
+  experiments_path: experiments/${dataset.name}/cabinet_experiments_large_${dataset.cropsize[0]}x${dataset.cropsize[1]} # Output path for checkpoints, logs, etc.
   model_save_name: cabinet.pth
   resume: false # Set to true to resume from checkpoint_last.pth in experiments_path