Match YOLOX and DAMO-YOLO preprocessing to upstream

.claude/skills/mblt-vision/SKILL.md

@@ -74,10 +74,13 @@ description: >-
   the shared instance attribute).
 - YOLOX and DAMO-YOLO dispatch on `post_cfg.head` (`yolox` / `damoyolo`) and reuse the
   anchorless filter and NMS; only their decode is new (no half-cell offset; DAMO `reg_max: 16`
-  is 17 bins; DAMO head layout resolved jointly). Both letterbox top-left
-  (`LetterBox.center: false`, padding 114 / 0), take unscaled input (no Normalize), and YOLOX
-  reads BGR (`Reader.color_mode`). Shape-only geometry must follow `letterbox_center(pre_cfg)`.
-  Configs follow mblt-model-ops `jm/temp` `pipeline.yaml`; YAMLs are `local_artifact_only`.
+  is 17 bins; DAMO head layout resolved jointly). Upstream's test transform is the source of
+  truth for `pre_cfg`: both letterbox top-left (`LetterBox.center: false`, padding 114 / 0)
+  with upstream's truncated `int(w * r)` size (`size_rounding: floor`) and take unscaled input
+  (no Normalize); YOLOX reads cv2 BGR (`Reader.color_mode`) and restores boxes by `/ r`;
+  DAMO-YOLO reads PIL RGB (`Reader.style: pil`, which COCO evaluation honours through
+  `CustomCOCODataset(decoder=)`) and restores each axis separately (`per_axis_ratio: true`). Shape-only geometry must follow `letterbox_layout(pre_cfg)`.
+  `post_cfg` follows mblt-model-ops `jm/temp` `pipeline.yaml`; YAMLs are `local_artifact_only`.
 - Rank WiderFace evaluation by Hard-set AP. Expose Medium-set then Easy-set AP
   as secondary metrics, and do not compute mean AP across difficulty splits.
 - Treat face_detection as a single-class WiderFace task. Each YOLO head family gets a thin

AGENTS.md

@@ -107,15 +107,24 @@ The current ownership boundary is deliberate:
   bins -- decoded as the softmax expectation times stride. Neither adds Ultralytics' half-cell
   offset. A 640 input's stride-8 DAMO class map is 80x80x80, so the head set's layout (NCHW
   or NHWC) is resolved jointly from the unambiguous distribution maps and a mixed set fails.
-- Their `pre_cfg` follows mblt-model-ops' `models/<Model>/pipeline.yaml` on branch
-  `jm/temp` (commit `0368367e8`): YOLOX is BGR (`Reader.color_mode: BGR`), top-left
-  letterboxed with 114; DAMO-YOLO is RGB, top-left with zeros (the December 2022 checkpoints'
-  geometry, not upstream HEAD's stretch). Neither declares `Normalize`: both take the
-  unscaled 0-255 image, and the ONNX path casts the byte tensor to the graph's float dtype.
-  `LetterBox.center: false` anchors top-left; every place that derives geometry from shapes
-  alone (`PostBase.ratio_pads_for`, `Results` plotting) must use the model's own
-  `letterbox_center(pre_cfg)` rather than the centered default. No Hub repository exists
-  yet, so their YAMLs keep `file_cfg.local_artifact_only: true`.
+- Their `pre_cfg` is upstream's own test transform, which is the source of truth (their
+  `post_cfg` follows mblt-model-ops' `models/<Model>/pipeline.yaml` on branch `jm/temp`,
+  commit `0368367e8`). YOLOX is `yolox/data/data_augment.py:preproc`: a cv2 BGR image
+  (`Reader.color_mode: BGR`), resized to `int(w * r) x int(h * r)`, top-left on a 114 canvas,
+  boxes restored by `/ r`. DAMO-YOLO is the December 2022 release's `Resize` and
+  `to_image_list` (tinyvision/DAMO-YOLO `55ae14f`; not upstream HEAD's stretch): a PIL RGB
+  decode (`Reader.style: pil`), the same truncated resize, top-left with zeros, and boxes
+  restored per axis by `BoxList.resize`. COCO evaluation decodes with the same library as the
+  model's `Reader` (`CustomCOCODataset(decoder="pil")` for `style: pil`), because the loader
+  decodes before preprocessing. Hence `LetterBox.size_rounding: floor` on both and
+  `per_axis_ratio: true` on DAMO-YOLO; tests compare both pipelines pixel-for-pixel with
+  inlined copies of the upstream code. Neither declares `Normalize`: both take the unscaled
+  0-255 image, and the ONNX path casts the byte tensor to the graph's float dtype. Every place
+  that derives geometry from shapes alone (`PostBase.ratio_pads_for`, `Results` plotting,
+  dense crops, semantic targets) must use the model's own `letterbox_layout(pre_cfg)`
+  rather than Ultralytics' default. Instance segmentation rejects `size_rounding: floor` and
+  `per_axis_ratio`, and OBB rejects `per_axis_ratio`, because their restorations cannot honour
+  them. No Hub repository exists yet, so their YAMLs keep `file_cfg.local_artifact_only: true`.
 - For WiderFace evaluation, rank results by Hard-set AP and retain Medium-set
   then Easy-set AP as secondary metrics. Do not compute a mean across splits.
 - The YOLOv5/YOLOv7 face repositories do not yet publish project-pinned immutable revisions

mblt_vision/README.md

@@ -369,8 +369,12 @@ validation split, following Ultralytics' depth-validation convention.
 - YOLOX and DAMO-YOLO have no Hub artifact yet, so pass a local `model_path`. Their GPU
   scores are the ONNX exports of the upstream checkpoints, measured through this package
   (upstream publishes 40.5 for YOLOX-s and 41.8 for DAMO-YOLO-T). Neither preprocesses like
-  Ultralytics: both letterbox top-left and feed the unscaled 0-255 image, and YOLOX reads
-  BGR. DAMO-YOLO follows its December 2022 checkpoints, the only ones still downloadable.
+  Ultralytics: both reproduce upstream's own test transform, which letterboxes top-left with
+  a truncated `int(w * r)` size (`LetterBox.size_rounding: floor`) and feeds the unscaled
+  0-255 image. YOLOX reads cv2 BGR and restores boxes by the one resize ratio. DAMO-YOLO
+  follows its December 2022 checkpoints, the only ones still downloadable: it decodes with
+  PIL (`Reader.style: pil`) and restores each axis by its own ratio
+  (`LetterBox.per_axis_ratio: true`).
 
 </details>
 

mblt_vision/__init__.py

@@ -19,7 +19,7 @@
 from ._api import list_tasks as list_tasks
 from .wrapper import MBLT_Engine as MBLT_Engine
 
-__version__ = "0.0.5"
+__version__ = "0.0.6"
 
 _TASK_MODULES = (
     face_detection,

mblt_vision/models/DAMO-YOLO-M-distill.yaml

@@ -1,9 +1,13 @@
 DEFAULT:
-  # pre_cfg/post_cfg follow mblt-model-ops' DAMO-YOLO-M-distill/pipeline.yaml (branch jm/temp, 0368367e8):
-  # the downloadable checkpoints are the December 2022 release, whose test
-  # transform letterboxes RGB with zeros anchored top-left (41.79 mAP50-95 on
-  # COCO val2017, against 39.72 through upstream HEAD's stretch) and feeds the
-  # unscaled image -- hence no Normalize. reg_max 16 means 17 distance bins.
+  # pre_cfg is the test transform of upstream DAMO-YOLO's December 2022 release
+  # (tinyvision/DAMO-YOLO 55ae14f), the only checkpoints still downloadable: a
+  # PIL RGB decode, cv2 resized to int(w * r) x int(h * r) -- truncated, hence
+  # size_rounding: floor -- zero-padded into the top-left corner, and left
+  # unscaled, hence no Normalize. BoxList.resize restores each axis by its own
+  # original-over-resized ratio, hence per_axis_ratio. (41.79 mAP50-95 on COCO
+  # val2017, against 39.72 through upstream HEAD's stretch.) post_cfg follows
+  # mblt-model-ops' DAMO-YOLO-M-distill/pipeline.yaml (branch jm/temp, 0368367e8);
+  # reg_max 16 means 17 distance bins.
   # No Hub artifact is published yet: pass a local model_path.
   file_cfg:
     repo_id: mobilint/DAMO-YOLO-M-distill
@@ -12,14 +16,15 @@ DEFAULT:
     local_artifact_only: true
   pre_cfg:
     Reader:
-      style: numpy
-      color_mode: RGB
+      style: pil
     LetterBox:
       img_size:
       - 640
       - 640
       center: false
       padding_value: 0
+      size_rounding: floor
+      per_axis_ratio: true
     SetOrder:
       shape: HWC
   post_cfg:

mblt_vision/models/DAMO-YOLO-M.yaml

@@ -1,9 +1,13 @@
 DEFAULT:
-  # pre_cfg/post_cfg follow mblt-model-ops' DAMO-YOLO-M/pipeline.yaml (branch jm/temp, 0368367e8):
-  # the downloadable checkpoints are the December 2022 release, whose test
-  # transform letterboxes RGB with zeros anchored top-left (41.79 mAP50-95 on
-  # COCO val2017, against 39.72 through upstream HEAD's stretch) and feeds the
-  # unscaled image -- hence no Normalize. reg_max 16 means 17 distance bins.
+  # pre_cfg is the test transform of upstream DAMO-YOLO's December 2022 release
+  # (tinyvision/DAMO-YOLO 55ae14f), the only checkpoints still downloadable: a
+  # PIL RGB decode, cv2 resized to int(w * r) x int(h * r) -- truncated, hence
+  # size_rounding: floor -- zero-padded into the top-left corner, and left
+  # unscaled, hence no Normalize. BoxList.resize restores each axis by its own
+  # original-over-resized ratio, hence per_axis_ratio. (41.79 mAP50-95 on COCO
+  # val2017, against 39.72 through upstream HEAD's stretch.) post_cfg follows
+  # mblt-model-ops' DAMO-YOLO-M/pipeline.yaml (branch jm/temp, 0368367e8);
+  # reg_max 16 means 17 distance bins.
   # No Hub artifact is published yet: pass a local model_path.
   file_cfg:
     repo_id: mobilint/DAMO-YOLO-M
@@ -12,14 +16,15 @@ DEFAULT:
     local_artifact_only: true
   pre_cfg:
     Reader:
-      style: numpy
-      color_mode: RGB
+      style: pil
     LetterBox:
       img_size:
       - 640
       - 640
       center: false
       padding_value: 0
+      size_rounding: floor
+      per_axis_ratio: true
     SetOrder:
       shape: HWC
   post_cfg:

mblt_vision/models/DAMO-YOLO-S-distill.yaml

@@ -1,9 +1,13 @@
 DEFAULT:
-  # pre_cfg/post_cfg follow mblt-model-ops' DAMO-YOLO-S-distill/pipeline.yaml (branch jm/temp, 0368367e8):
-  # the downloadable checkpoints are the December 2022 release, whose test
-  # transform letterboxes RGB with zeros anchored top-left (41.79 mAP50-95 on
-  # COCO val2017, against 39.72 through upstream HEAD's stretch) and feeds the
-  # unscaled image -- hence no Normalize. reg_max 16 means 17 distance bins.
+  # pre_cfg is the test transform of upstream DAMO-YOLO's December 2022 release
+  # (tinyvision/DAMO-YOLO 55ae14f), the only checkpoints still downloadable: a
+  # PIL RGB decode, cv2 resized to int(w * r) x int(h * r) -- truncated, hence
+  # size_rounding: floor -- zero-padded into the top-left corner, and left
+  # unscaled, hence no Normalize. BoxList.resize restores each axis by its own
+  # original-over-resized ratio, hence per_axis_ratio. (41.79 mAP50-95 on COCO
+  # val2017, against 39.72 through upstream HEAD's stretch.) post_cfg follows
+  # mblt-model-ops' DAMO-YOLO-S-distill/pipeline.yaml (branch jm/temp, 0368367e8);
+  # reg_max 16 means 17 distance bins.
   # No Hub artifact is published yet: pass a local model_path.
   file_cfg:
     repo_id: mobilint/DAMO-YOLO-S-distill
@@ -12,14 +16,15 @@ DEFAULT:
     local_artifact_only: true
   pre_cfg:
     Reader:
-      style: numpy
-      color_mode: RGB
+      style: pil
     LetterBox:
       img_size:
       - 640
       - 640
       center: false
       padding_value: 0
+      size_rounding: floor
+      per_axis_ratio: true
     SetOrder:
       shape: HWC
   post_cfg:

mblt_vision/models/DAMO-YOLO-S.yaml

@@ -1,9 +1,13 @@
 DEFAULT:
-  # pre_cfg/post_cfg follow mblt-model-ops' DAMO-YOLO-S/pipeline.yaml (branch jm/temp, 0368367e8):
-  # the downloadable checkpoints are the December 2022 release, whose test
-  # transform letterboxes RGB with zeros anchored top-left (41.79 mAP50-95 on
-  # COCO val2017, against 39.72 through upstream HEAD's stretch) and feeds the
-  # unscaled image -- hence no Normalize. reg_max 16 means 17 distance bins.
+  # pre_cfg is the test transform of upstream DAMO-YOLO's December 2022 release
+  # (tinyvision/DAMO-YOLO 55ae14f), the only checkpoints still downloadable: a
+  # PIL RGB decode, cv2 resized to int(w * r) x int(h * r) -- truncated, hence
+  # size_rounding: floor -- zero-padded into the top-left corner, and left
+  # unscaled, hence no Normalize. BoxList.resize restores each axis by its own
+  # original-over-resized ratio, hence per_axis_ratio. (41.79 mAP50-95 on COCO
+  # val2017, against 39.72 through upstream HEAD's stretch.) post_cfg follows
+  # mblt-model-ops' DAMO-YOLO-S/pipeline.yaml (branch jm/temp, 0368367e8);
+  # reg_max 16 means 17 distance bins.
   # No Hub artifact is published yet: pass a local model_path.
   file_cfg:
     repo_id: mobilint/DAMO-YOLO-S
@@ -12,14 +16,15 @@ DEFAULT:
     local_artifact_only: true
   pre_cfg:
     Reader:
-      style: numpy
-      color_mode: RGB
+      style: pil
     LetterBox:
       img_size:
       - 640
       - 640
       center: false
       padding_value: 0
+      size_rounding: floor
+      per_axis_ratio: true
     SetOrder:
       shape: HWC
   post_cfg:

mblt_vision/models/DAMO-YOLO-T-distill.yaml

@@ -1,9 +1,13 @@
 DEFAULT:
-  # pre_cfg/post_cfg follow mblt-model-ops' DAMO-YOLO-T-distill/pipeline.yaml (branch jm/temp, 0368367e8):
-  # the downloadable checkpoints are the December 2022 release, whose test
-  # transform letterboxes RGB with zeros anchored top-left (41.79 mAP50-95 on
-  # COCO val2017, against 39.72 through upstream HEAD's stretch) and feeds the
-  # unscaled image -- hence no Normalize. reg_max 16 means 17 distance bins.
+  # pre_cfg is the test transform of upstream DAMO-YOLO's December 2022 release
+  # (tinyvision/DAMO-YOLO 55ae14f), the only checkpoints still downloadable: a
+  # PIL RGB decode, cv2 resized to int(w * r) x int(h * r) -- truncated, hence
+  # size_rounding: floor -- zero-padded into the top-left corner, and left
+  # unscaled, hence no Normalize. BoxList.resize restores each axis by its own
+  # original-over-resized ratio, hence per_axis_ratio. (41.79 mAP50-95 on COCO
+  # val2017, against 39.72 through upstream HEAD's stretch.) post_cfg follows
+  # mblt-model-ops' DAMO-YOLO-T-distill/pipeline.yaml (branch jm/temp, 0368367e8);
+  # reg_max 16 means 17 distance bins.
   # No Hub artifact is published yet: pass a local model_path.
   file_cfg:
     repo_id: mobilint/DAMO-YOLO-T-distill
@@ -12,14 +16,15 @@ DEFAULT:
     local_artifact_only: true
   pre_cfg:
     Reader:
-      style: numpy
-      color_mode: RGB
+      style: pil
     LetterBox:
       img_size:
       - 640
       - 640
       center: false
       padding_value: 0
+      size_rounding: floor
+      per_axis_ratio: true
     SetOrder:
       shape: HWC
   post_cfg:

mblt_vision/models/DAMO-YOLO-T.yaml

@@ -1,9 +1,13 @@
 DEFAULT:
-  # pre_cfg/post_cfg follow mblt-model-ops' DAMO-YOLO-T/pipeline.yaml (branch jm/temp, 0368367e8):
-  # the downloadable checkpoints are the December 2022 release, whose test
-  # transform letterboxes RGB with zeros anchored top-left (41.79 mAP50-95 on
-  # COCO val2017, against 39.72 through upstream HEAD's stretch) and feeds the
-  # unscaled image -- hence no Normalize. reg_max 16 means 17 distance bins.
+  # pre_cfg is the test transform of upstream DAMO-YOLO's December 2022 release
+  # (tinyvision/DAMO-YOLO 55ae14f), the only checkpoints still downloadable: a
+  # PIL RGB decode, cv2 resized to int(w * r) x int(h * r) -- truncated, hence
+  # size_rounding: floor -- zero-padded into the top-left corner, and left
+  # unscaled, hence no Normalize. BoxList.resize restores each axis by its own
+  # original-over-resized ratio, hence per_axis_ratio. (41.79 mAP50-95 on COCO
+  # val2017, against 39.72 through upstream HEAD's stretch.) post_cfg follows
+  # mblt-model-ops' DAMO-YOLO-T/pipeline.yaml (branch jm/temp, 0368367e8);
+  # reg_max 16 means 17 distance bins.
   # No Hub artifact is published yet: pass a local model_path.
   file_cfg:
     repo_id: mobilint/DAMO-YOLO-T
@@ -12,14 +16,15 @@ DEFAULT:
     local_artifact_only: true
   pre_cfg:
     Reader:
-      style: numpy
-      color_mode: RGB
+      style: pil
     LetterBox:
       img_size:
       - 640
       - 640
       center: false
       padding_value: 0
+      size_rounding: floor
+      per_axis_ratio: true
     SetOrder:
       shape: HWC
   post_cfg:

mblt_vision/models/YOLOX-Darknet53.yaml

@@ -1,7 +1,10 @@
 DEFAULT:
-  # pre_cfg/post_cfg follow mblt-model-ops' YOLOX-Darknet53/pipeline.yaml (branch jm/temp, 0368367e8):
-  # upstream's ValTransform(legacy=False) feeds a cv2 BGR image, resized onto a
-  # 114-filled canvas anchored top-left, and left unscaled -- hence no Normalize.
+  # pre_cfg is upstream YOLOX's own test transform (yolox/data/data_augment.py
+  # preproc via ValTransform(legacy=False)): a cv2 BGR image resized to
+  # int(w * r) x int(h * r) -- truncated, hence size_rounding: floor -- onto a
+  # 114-filled canvas anchored top-left, and left unscaled, hence no Normalize.
+  # Boxes are restored by dividing by r. post_cfg follows mblt-model-ops'
+  # YOLOX-Darknet53/pipeline.yaml (branch jm/temp, 0368367e8).
   # No Hub artifact is published yet: pass a local model_path.
   file_cfg:
     repo_id: mobilint/YOLOX-Darknet53
@@ -18,6 +21,7 @@ DEFAULT:
       - 640
       center: false
       padding_value: 114
+      size_rounding: floor
     SetOrder:
       shape: HWC
   post_cfg:

mblt_vision/models/YOLOX-Nano.yaml

@@ -1,7 +1,10 @@
 DEFAULT:
-  # pre_cfg/post_cfg follow mblt-model-ops' YOLOX-Nano/pipeline.yaml (branch jm/temp, 0368367e8):
-  # upstream's ValTransform(legacy=False) feeds a cv2 BGR image, resized onto a
-  # 114-filled canvas anchored top-left, and left unscaled -- hence no Normalize.
+  # pre_cfg is upstream YOLOX's own test transform (yolox/data/data_augment.py
+  # preproc via ValTransform(legacy=False)): a cv2 BGR image resized to
+  # int(w * r) x int(h * r) -- truncated, hence size_rounding: floor -- onto a
+  # 114-filled canvas anchored top-left, and left unscaled, hence no Normalize.
+  # Boxes are restored by dividing by r. post_cfg follows mblt-model-ops'
+  # YOLOX-Nano/pipeline.yaml (branch jm/temp, 0368367e8).
   # No Hub artifact is published yet: pass a local model_path.
   file_cfg:
     repo_id: mobilint/YOLOX-Nano
@@ -18,6 +21,7 @@ DEFAULT:
       - 416
       center: false
       padding_value: 114
+      size_rounding: floor
     SetOrder:
       shape: HWC
   post_cfg:

mblt_vision/models/YOLOX-Tiny.yaml

@@ -1,7 +1,10 @@
 DEFAULT:
-  # pre_cfg/post_cfg follow mblt-model-ops' YOLOX-Tiny/pipeline.yaml (branch jm/temp, 0368367e8):
-  # upstream's ValTransform(legacy=False) feeds a cv2 BGR image, resized onto a
-  # 114-filled canvas anchored top-left, and left unscaled -- hence no Normalize.
+  # pre_cfg is upstream YOLOX's own test transform (yolox/data/data_augment.py
+  # preproc via ValTransform(legacy=False)): a cv2 BGR image resized to
+  # int(w * r) x int(h * r) -- truncated, hence size_rounding: floor -- onto a
+  # 114-filled canvas anchored top-left, and left unscaled, hence no Normalize.
+  # Boxes are restored by dividing by r. post_cfg follows mblt-model-ops'
+  # YOLOX-Tiny/pipeline.yaml (branch jm/temp, 0368367e8).
   # No Hub artifact is published yet: pass a local model_path.
   file_cfg:
     repo_id: mobilint/YOLOX-Tiny
@@ -18,6 +21,7 @@ DEFAULT:
       - 416
       center: false
       padding_value: 114
+      size_rounding: floor
     SetOrder:
       shape: HWC
   post_cfg:

mblt_vision/models/YOLOX-l.yaml

@@ -1,7 +1,10 @@
 DEFAULT:
-  # pre_cfg/post_cfg follow mblt-model-ops' YOLOX-l/pipeline.yaml (branch jm/temp, 0368367e8):
-  # upstream's ValTransform(legacy=False) feeds a cv2 BGR image, resized onto a
-  # 114-filled canvas anchored top-left, and left unscaled -- hence no Normalize.
+  # pre_cfg is upstream YOLOX's own test transform (yolox/data/data_augment.py
+  # preproc via ValTransform(legacy=False)): a cv2 BGR image resized to
+  # int(w * r) x int(h * r) -- truncated, hence size_rounding: floor -- onto a
+  # 114-filled canvas anchored top-left, and left unscaled, hence no Normalize.
+  # Boxes are restored by dividing by r. post_cfg follows mblt-model-ops'
+  # YOLOX-l/pipeline.yaml (branch jm/temp, 0368367e8).
   # No Hub artifact is published yet: pass a local model_path.
   file_cfg:
     repo_id: mobilint/YOLOX-l
@@ -18,6 +21,7 @@ DEFAULT:
       - 640
       center: false
       padding_value: 114
+      size_rounding: floor
     SetOrder:
       shape: HWC
   post_cfg: