fotonet 0.1.0a0__py3-none-any.whl

fotonet/__init__.py

@@ -0,0 +1,32 @@
+"""FOTO-NET: fast Apache-friendly object detection with an Ultralytics-like API."""
+
+from fotonet._torch_notice import maybe_print_torch_notice
+
+
+__version__ = "0.1.0a0"
+__all__ = ["FOTONET", "Vector2", "AnchorPoint", "FocusRegion", "BoxTransform", "__version__"]
+
+_PUBLIC_API = {"FOTONET", "Vector2", "AnchorPoint", "FocusRegion", "BoxTransform"}
+
+maybe_print_torch_notice()
+
+
+def _load_public_api():
+    from fotonet.engine.model import FOTONET
+    from fotonet.engine.results import AnchorPoint, BoxTransform, FocusRegion, Vector2
+
+    public = {
+        "FOTONET": FOTONET,
+        "Vector2": Vector2,
+        "AnchorPoint": AnchorPoint,
+        "FocusRegion": FocusRegion,
+        "BoxTransform": BoxTransform,
+    }
+    globals().update(public)
+    return public
+
+
+def __getattr__(name):
+    if name in _PUBLIC_API:
+        return _load_public_api()[name]
+    raise AttributeError(f"module 'fotonet' has no attribute {name!r}")

fotonet/_torch_notice.py

@@ -0,0 +1,21 @@
+"""Helpers for guiding users to install PyTorch outside package dependencies."""
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+
+
+TORCH_INSTALL_MESSAGE = "please download torch for your system at https://pytorch.org"
+
+
+def is_torch_available():
+    return importlib.util.find_spec("torch") is not None
+
+
+def maybe_print_torch_notice(torch_available=None):
+    if torch_available is None:
+        torch_available = is_torch_available()
+    if not torch_available:
+        print(TORCH_INSTALL_MESSAGE, file=sys.stderr)
+

fotonet/cli/main.py

@@ -0,0 +1,145 @@
+"""FOTO-NET command line interface.
+
+Examples:
+  fotonet train model=fotonetn data=datasets/coco/coco.yaml epochs=250 batch=16 imgsz=640
+  fotonet predict model=fotonet-n source=image.jpg conf=0.25 save=true
+  fotonet val model=fotonet-n data=datasets/coco/coco.yaml imgsz=640
+  fotonet export model=fotonet-n format=onnx path=dev-tools/runs/fotonet.onnx half=true
+
+YOLO-style mode form is also accepted:
+  fotonet mode=train model=fotonetn data=coco.yaml epochs=100
+"""
+
+import ast
+import os
+import sys
+
+
+TASK_ALIASES = {
+    "detect": "predict",
+    "infer": "predict",
+    "predict": "predict",
+    "track": "track",
+    "train": "train",
+    "val": "val",
+    "validate": "val",
+    "export": "export",
+}
+
+
+def _coerce(value):
+    raw = str(value).strip()
+    low = raw.lower()
+    if low in {"true", "false"}:
+        return low == "true"
+    if low in {"none", "null"}:
+        return None
+    try:
+        return ast.literal_eval(raw)
+    except (ValueError, SyntaxError):
+        return value
+
+
+def parse_args(argv):
+    args = {}
+    task = None
+    remaining = list(argv)
+
+    if remaining and "=" not in remaining[0]:
+        task = remaining.pop(0).lower()
+
+    for token in remaining:
+        if "=" not in token:
+            raise SystemExit(f"Invalid argument '{token}'. Use key=value.")
+        key, value = token.split("=", 1)
+        args[key.replace("-", "_")] = _coerce(value)
+
+    task = args.pop("mode", args.pop("task", task or "predict"))
+    task = TASK_ALIASES.get(str(task).lower())
+    if task is None:
+        valid = ", ".join(sorted(set(TASK_ALIASES.values())))
+        raise SystemExit(f"Invalid task. Use one of: {valid}.")
+    return task, args
+
+
+def _pop_bool(args, key, default=False):
+    value = args.pop(key, default)
+    return bool(value)
+
+
+def _print_help():
+    print(__doc__.strip())
+
+
+def main(argv=None):
+    argv = sys.argv[1:] if argv is None else argv
+    if not argv or argv[0] in {"-h", "--help", "help"}:
+        _print_help()
+        return
+
+    from fotonet import FOTONET
+
+    task, args = parse_args(argv)
+    model_name = args.pop("model", "fotonetn")
+    imgsz = int(args.pop("imgsz", 640))
+
+    model = FOTONET(model_name)
+
+    if task == "train":
+        data = args.pop("data", None)
+        if not data:
+            raise SystemExit("Training requires data=path/to/data.yaml.")
+        epochs = int(args.pop("epochs", 100))
+        save_path = args.pop("save", "fotonet_trained.pt")
+        print(f"FOTO-NET train | model={model_name} data={data} epochs={epochs} imgsz={imgsz}")
+        model.train(data=data, epochs=epochs, imgsz=imgsz, **args)
+        if save_path:
+            model.save(str(save_path), inference_only=True, half=True)
+            print(f"saved={save_path}")
+        return
+
+    if task in {"predict", "track"}:
+        source = args.pop("source", None)
+        if not source:
+            raise SystemExit(f"{task} requires source=image.jpg|folder.")
+        conf = float(args.pop("conf", 0.25))
+        save = _pop_bool(args, "save", True)
+        show = _pop_bool(args, "show", False)
+        use_nms = _pop_bool(args, "use_nms", False)
+        results = model.predict(source, imgsz=imgsz, conf=conf, use_nms=use_nms, **args)
+        result_list = results if isinstance(results, list) else [results]
+        print(
+            f"FOTO-NET {task} | model={model_name} source={source} "
+            f"images={len(result_list)} nms_free={not use_nms}"
+        )
+        if save:
+            save_dir = "runs"
+            os.makedirs(save_dir, exist_ok=True)
+            for i, result in enumerate(result_list):
+                result.save(os.path.join(save_dir, f"predict_{i}.jpg"))
+            print(f"saved_dir={save_dir}")
+        if show and not isinstance(results, list):
+            results.show()
+        return
+
+    if task == "val":
+        data = args.pop("data", None)
+        if not data:
+            raise SystemExit("Validation requires data=path/to/data.yaml.")
+        metrics = model.val(data=data, imgsz=imgsz, **args)
+        print("FOTO-NET val")
+        for key, value in metrics.items():
+            if key.endswith("per_class"):
+                continue
+            print(f"{key}={value}")
+        return
+
+    if task == "export":
+        fmt = str(args.pop("format", "onnx")).lower()
+        path = args.pop("path", f"fotonet_export.{fmt}")
+        out = model.export(path=path, format=fmt, imgsz=imgsz, **args)
+        print(f"exported={out}")
+
+
+if __name__ == "__main__":
+    main()

fotonet/config/models/fotonetl-p2.yaml

@@ -0,0 +1,29 @@
+# FOTO-NET Large P2 - high-capacity O2O target with P2 small-object head
+nc: 80
+depth_multiple: 0.50
+width_multiple: 0.55
+p2_head: true
+reg_max: 1
+p5_gate_blocks: 1
+neck_fusion: weighted
+p2_context_blocks: 1
+p3_context_blocks: 1
+quality_head: true
+
+backbone:
+  - [Conv, [3, 64, 3, 2]]
+  - [Conv, [64, 128, 3, 2]]
+  - [C2f, [128, 128, 3, True]]
+  - [Conv, [128, 256, 3, 2]]
+  - [C2f, [256, 256, 6, True]]
+  - [Conv, [256, 512, 3, 2]]
+  - [C2f, [512, 512, 9, True]]
+  - [Conv, [512, 1024, 3, 2]]
+  - [C2f, [1024, 1024, 3, True]]
+
+neck:
+  width_multiple: 1.0
+  depth_multiple: 1.0
+
+head:
+  width_multiple: 1.0

fotonet/config/models/fotonetl.yaml

@@ -0,0 +1,32 @@
+# FOTO-NET Large - P2-less high-capacity O2O target
+nc: 80
+depth_multiple: 0.50
+width_multiple: 0.55
+p2_head: false
+reg_max: 1
+p3_extra_blocks: 3
+p4_extra_blocks: 2
+p5_extra_blocks: 1
+p5_gate_blocks: 1
+neck_fusion: weighted
+p2_context_blocks: 0
+p3_context_blocks: 1
+quality_head: true
+
+backbone:
+  - [Conv, [3, 64, 3, 2]]
+  - [Conv, [64, 128, 3, 2]]
+  - [C2f, [128, 128, 3, True]]
+  - [Conv, [128, 256, 3, 2]]
+  - [C2f, [256, 256, 6, True]]
+  - [Conv, [256, 512, 3, 2]]
+  - [C2f, [512, 512, 9, True]]
+  - [Conv, [512, 1024, 3, 2]]
+  - [C2f, [1024, 1024, 3, True]]
+
+neck:
+  width_multiple: 1.0
+  depth_multiple: 1.0
+
+head:
+  width_multiple: 1.0

fotonet/config/models/fotonetm-p2.yaml

@@ -0,0 +1,29 @@
+# FOTO-NET Medium P2 - compact mid-size O2O target with P2 small-object head
+nc: 80
+depth_multiple: 0.67
+width_multiple: 0.38
+p2_head: true
+reg_max: 1
+p5_gate_blocks: 1
+neck_fusion: weighted
+p2_context_blocks: 1
+p3_context_blocks: 1
+quality_head: true
+
+backbone:
+  - [Conv, [3, 64, 3, 2]]
+  - [Conv, [64, 128, 3, 2]]
+  - [C2f, [128, 128, 6, True]]
+  - [Conv, [128, 256, 3, 2]]
+  - [C2f, [256, 256, 9, True]]
+  - [Conv, [256, 512, 3, 2]]
+  - [C2f, [512, 512, 9, True]]
+  - [Conv, [512, 1024, 3, 2]]
+  - [C2f, [1024, 1024, 3, True]]
+
+neck:
+  width_multiple: 0.75
+  depth_multiple: 0.67
+
+head:
+  width_multiple: 0.75

fotonet/config/models/fotonetm.yaml

@@ -0,0 +1,32 @@
+# FOTO-NET Medium - P2-less compact mid-size O2O target
+nc: 80
+depth_multiple: 0.67
+width_multiple: 0.38
+p2_head: false
+reg_max: 1
+p3_extra_blocks: 2
+p4_extra_blocks: 1
+p5_extra_blocks: 1
+p5_gate_blocks: 1
+neck_fusion: weighted
+p2_context_blocks: 0
+p3_context_blocks: 1
+quality_head: true
+
+backbone:
+  - [Conv, [3, 64, 3, 2]]
+  - [Conv, [64, 128, 3, 2]]
+  - [C2f, [128, 128, 6, True]]
+  - [Conv, [128, 256, 3, 2]]
+  - [C2f, [256, 256, 9, True]]
+  - [Conv, [256, 512, 3, 2]]
+  - [C2f, [512, 512, 9, True]]
+  - [Conv, [512, 1024, 3, 2]]
+  - [C2f, [1024, 1024, 3, True]]
+
+neck:
+  width_multiple: 0.75
+  depth_multiple: 0.67
+
+head:
+  width_multiple: 0.75

fotonet/config/models/fotonetn-p2.yaml

@@ -0,0 +1,30 @@
+# FOTO-NET Nano - O2O deployment target, P2 small-object head
+nc: 80
+depth_multiple: 0.33
+width_multiple: 0.20
+p2_head: true
+reg_max: 1
+p5_extra_blocks: 1
+p5_gate_blocks: 2
+neck_fusion: weighted
+p2_context_blocks: 1
+p3_context_blocks: 1
+quality_head: true
+
+backbone:
+  - [Conv, [3, 64, 3, 2]]
+  - [Conv, [64, 128, 3, 2]]
+  - [C2f, [128, 128, 2, True]]
+  - [Conv, [128, 256, 3, 2]]
+  - [C2f, [256, 256, 4, True]]
+  - [Conv, [256, 512, 3, 2]]
+  - [C2f, [512, 512, 4, True]]
+  - [Conv, [512, 1024, 3, 2]]
+  - [C2f, [1024, 1024, 2, True]]
+
+neck:
+  width_multiple: 0.25
+  depth_multiple: 0.33
+
+head:
+  width_multiple: 0.25

fotonet/config/models/fotonetn.yaml

@@ -0,0 +1,30 @@
+# FOTO-NET Nano-0 - P2-less edge/general deployment target
+nc: 80
+depth_multiple: 0.42
+width_multiple: 0.205
+p2_head: false
+reg_max: 1
+p5_extra_blocks: 1
+p5_gate_blocks: 2
+neck_fusion: weighted
+p2_context_blocks: 0
+p3_context_blocks: 1
+quality_head: false
+
+backbone:
+  - [Conv, [3, 64, 3, 2]]
+  - [Conv, [64, 128, 3, 2]]
+  - [C2f, [128, 128, 2, true]]
+  - [Conv, [128, 256, 3, 2]]
+  - [C2f, [256, 256, 5, true]]
+  - [Conv, [256, 512, 3, 2]]
+  - [C2f, [512, 512, 5, true]]
+  - [Conv, [512, 1024, 3, 2]]
+  - [C2f, [1024, 1024, 2, true]]
+
+neck:
+  width_multiple: 0.205
+  depth_multiple: 0.42
+
+head:
+  width_multiple: 0.205

fotonet/config/models/fotonets-p2.yaml

@@ -0,0 +1,29 @@
+# FOTO-NET Small P2 - compact O2O target with P2 small-object head
+nc: 80
+depth_multiple: 0.33
+width_multiple: 0.30
+p2_head: true
+reg_max: 1
+p5_gate_blocks: 1
+neck_fusion: weighted
+p2_context_blocks: 1
+p3_context_blocks: 1
+quality_head: true
+
+backbone:
+  - [Conv, [3, 64, 3, 2]]
+  - [Conv, [64, 128, 3, 2]]
+  - [C2f, [128, 128, 3, True]]
+  - [Conv, [128, 256, 3, 2]]
+  - [C2f, [256, 256, 6, True]]
+  - [Conv, [256, 512, 3, 2]]
+  - [C2f, [512, 512, 6, True]]
+  - [Conv, [512, 1024, 3, 2]]
+  - [C2f, [1024, 1024, 3, True]]
+
+neck:
+  width_multiple: 0.50
+  depth_multiple: 0.33
+
+head:
+  width_multiple: 0.50

fotonet/config/models/fotonets.yaml

@@ -0,0 +1,32 @@
+# FOTO-NET Small - P2-less compact O2O deployment target
+nc: 80
+depth_multiple: 0.33
+width_multiple: 0.30
+p2_head: false
+reg_max: 1
+p3_extra_blocks: 2
+p4_extra_blocks: 1
+p5_extra_blocks: 1
+p5_gate_blocks: 1
+neck_fusion: weighted
+p2_context_blocks: 0
+p3_context_blocks: 1
+quality_head: true
+
+backbone:
+  - [Conv, [3, 64, 3, 2]]
+  - [Conv, [64, 128, 3, 2]]
+  - [C2f, [128, 128, 3, True]]
+  - [Conv, [128, 256, 3, 2]]
+  - [C2f, [256, 256, 6, True]]
+  - [Conv, [256, 512, 3, 2]]
+  - [C2f, [512, 512, 6, True]]
+  - [Conv, [512, 1024, 3, 2]]
+  - [C2f, [1024, 1024, 3, True]]
+
+neck:
+  width_multiple: 0.50
+  depth_multiple: 0.33
+
+head:
+  width_multiple: 0.50

fotonet/config/models/fotonetx-p2.yaml

@@ -0,0 +1,29 @@
+# FOTO-NET Extra-Large P2 - max-capacity O2O target with P2 small-object head
+nc: 80
+depth_multiple: 0.50
+width_multiple: 0.70
+p2_head: true
+reg_max: 1
+p5_gate_blocks: 1
+neck_fusion: weighted
+p2_context_blocks: 1
+p3_context_blocks: 1
+quality_head: true
+
+backbone:
+  - [Conv, [3, 64, 3, 2]]
+  - [Conv, [64, 128, 3, 2]]
+  - [C2f, [128, 128, 3, True]]
+  - [Conv, [128, 256, 3, 2]]
+  - [C2f, [256, 256, 6, True]]
+  - [Conv, [256, 512, 3, 2]]
+  - [C2f, [512, 512, 9, True]]
+  - [Conv, [512, 1024, 3, 2]]
+  - [C2f, [1024, 1024, 3, True]]
+
+neck:
+  width_multiple: 1.25
+  depth_multiple: 1.0
+
+head:
+  width_multiple: 1.25

fotonet/config/models/fotonetx.yaml

@@ -0,0 +1,32 @@
+# FOTO-NET Extra-Large - P2-less max-capacity O2O target
+nc: 80
+depth_multiple: 0.50
+width_multiple: 0.70
+p2_head: false
+reg_max: 1
+p3_extra_blocks: 3
+p4_extra_blocks: 2
+p5_extra_blocks: 1
+p5_gate_blocks: 1
+neck_fusion: weighted
+p2_context_blocks: 0
+p3_context_blocks: 1
+quality_head: true
+
+backbone:
+  - [Conv, [3, 64, 3, 2]]
+  - [Conv, [64, 128, 3, 2]]
+  - [C2f, [128, 128, 3, True]]
+  - [Conv, [128, 256, 3, 2]]
+  - [C2f, [256, 256, 6, True]]
+  - [Conv, [256, 512, 3, 2]]
+  - [C2f, [512, 512, 9, True]]
+  - [Conv, [512, 1024, 3, 2]]
+  - [C2f, [1024, 1024, 3, True]]
+
+neck:
+  width_multiple: 1.25
+  depth_multiple: 1.0
+
+head:
+  width_multiple: 1.25

fotonet/config/recipes/fotonetn_scratch.yaml

@@ -0,0 +1,129 @@
+# FOTO-NET-N scratch recipe for COCO-style object detection.
+# Designed for full training, not smoke tests.
+
+model: fotonetn
+pretrained: false
+resume: false
+distill_teacher: null
+distill_weight: 0.0
+
+train:
+  epochs: 500
+  imgsz: 640
+  imgsz_schedule:
+    - [0.35, 640]
+  batch: 16
+  val_batch: 8
+  nbs: 128
+  workers: 4
+  pin_memory: true
+  cache_to_ram: true
+  ram_cache_images: 1024
+  cache_labels: true
+  disk_cache_images: false
+  amp: true
+  val_amp: true
+  cos_lr: true
+  compile_model: false
+  cuda_graphs: false
+  val_period: 5
+  val_subset_size: 0
+  full_val_after: 0.90
+  save_period: -1
+  slim_best: true
+  best_metric: mAP50_95
+
+optimizer:
+  # Current trainer implements SGD. Keep MuSGD as future recipe target once optimizer lands.
+  name: sgd
+  lr0: 0.0054
+  lrf: 0.015
+  momentum: 0.947
+  weight_decay: 0.00064
+  warmup_epochs: 3
+
+augment:
+  mosaic: 0.909
+  mixup: 0.012
+  copy_paste: 0.075
+  copy_paste_mode: flip
+  close_mosaic: 15
+  scale: 0.562
+  fliplr: 0.606
+  flipud: 0.0
+  degrees: 1.11
+  shear: 1.46
+  translate: 0.071
+  hsv_h: 0.014
+  hsv_s: 0.645
+  hsv_v: 0.566
+  bgr: 0.106
+  mixup_alpha: 32.0
+
+loss:
+  box: 7.5
+  cls: 1.0
+  dfl: 1.5
+  dfl_mean_over_coords: true
+  o2o_label_smoothing: 0.01
+  o2m_label_smoothing: 0.05
+  focal_gamma_pos: 2.0
+  focal_gamma_neg: 2.0
+  quality_targets: true
+  quality_power: 1.0
+  quality_floor: 0.05
+  quality_start_epoch: 25
+  quality_warmup_epochs: 50
+  qfl_beta: 2.0
+  hard_negative_topk: 256
+  hard_negative_weight: 0.10
+  hard_negative_min_score: 0.15
+  hard_negative_start_epoch: 8
+  hard_negative_warmup_epochs: 32
+  o2o_weight: 1.0
+  o2m_weight: 1.0
+  o2m_full_epochs: 250
+  o2m_sparse_epochs: 400
+  o2m_every_n_steps: 4
+  o2m_disable_epoch: 400
+  consistency_weight: 0.35
+  consistency_start_epoch: 10
+  consistency_warmup_epochs: 40
+  consistency_cls_weight: 0.25
+  consistency_box_weight: 0.75
+  prog_loss: true
+  prog_start_epoch: 1
+  prog_warmup_epochs: 60
+  cls_start_factor: 1.50
+  box_start_factor: 1.30
+  dfl_start_factor: 0.85
+  quality: 1.0
+  quality_loss_start_epoch: 1
+  quality_loss_warmup_epochs: 25
+
+matcher:
+  topk: 8
+  alpha: 1.0
+  beta: 2.0
+  cost_class: 2.0
+  cost_bbox: 2.0
+  cost_giou: 1.0
+  exact_o2o: true
+  exact_o2o_warmup_epochs: 10
+  exact_o2o_period: 8
+  exact_o2o_period_end_epoch: 250
+  min_small_assign: 4
+  small_obj_px: 8.0
+
+smoke_limits:
+  max_epochs_2500_images: 15
+  max_epochs_5000_images: 8
+  recommended_quick_check:
+    epochs: 2
+    max_images: 256
+    imgsz: 128
+    val_subset_size: 64
+    full_val_after: 1.01  # keep validation subset on final smoke epoch
+    mosaic: 0.0
+    mixup: 0.0
+    copy_paste: 0.0

fotonet/data/augmentations/__init__.py

@@ -0,0 +1,40 @@
+"""Public augmentation API."""
+from fotonet.data.augmentations.boxes import (
+    bbox_ioa,
+    clip_boxes_xyxy,
+    filter_boxes_xywhn,
+    sanitize_boxes_xywhn,
+    xywhn_to_xyxy,
+    xyxy_to_xywhn,
+)
+from fotonet.data.augmentations.compose import copy_paste, finalize_sample, load_mosaic, mixup
+from fotonet.data.augmentations.geometric import random_affine
+from fotonet.data.augmentations.hyp import DEFAULT_AUGMENT_HYP, build_augment_hyp
+from fotonet.data.augmentations.image import (
+    augment_hsv_safe,
+    letterbox_resize,
+    random_channel_swap,
+    random_flip,
+    to_tensor,
+)
+
+__all__ = [
+    "DEFAULT_AUGMENT_HYP",
+    "augment_hsv_safe",
+    "bbox_ioa",
+    "build_augment_hyp",
+    "clip_boxes_xyxy",
+    "copy_paste",
+    "filter_boxes_xywhn",
+    "finalize_sample",
+    "letterbox_resize",
+    "load_mosaic",
+    "mixup",
+    "random_affine",
+    "random_channel_swap",
+    "random_flip",
+    "sanitize_boxes_xywhn",
+    "to_tensor",
+    "xywhn_to_xyxy",
+    "xyxy_to_xywhn",
+]