mblt-vision-python 0.0.3__tar.gz → 0.0.4__tar.gz

{mblt_vision_python-0.0.3/mblt_vision_python.egg-info → mblt_vision_python-0.0.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mblt-vision-python
-Version: 0.0.3
+Version: 0.0.4
 Summary: Mobilint NPU vision models: classification, detection, segmentation, pose
 Author: Mobilint
 License: BSD-3-Clause

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/__init__.py

@@ -20,7 +20,7 @@ from ._api import list_models as list_models
 from ._api import list_tasks as list_tasks
 from .wrapper import MBLT_Engine as MBLT_Engine
 
-__version__ = "0.0.3"
+__version__ = "0.0.4"
 
 _TASK_MODULES = (
     face_detection,

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/CoAtNet_0_RW_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/CoAtNet_1_RW_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/CoAtNet_2_RW_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/DeiT3_Base_Patch16_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/DeiT3_Large_Patch16_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/DeiT3_Medium_Patch16_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/DeiT3_Small_Patch16_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/DeiT_Base_Patch16_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/DeiT_Small_Patch16_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/DeiT_Tiny_Patch16_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/EfficientFormer_L1.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/EfficientFormer_L3.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/EfficientFormer_L7.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/LeViT_Conv_128.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/LeViT_Conv_128S.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/LeViT_Conv_192.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/LeViT_Conv_256.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/LeViT_Conv_384.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/RepViT_M0_9.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/RepViT_M1.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/RepViT_M1_0.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/RepViT_M1_1.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/RepViT_M1_5.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/RepViT_M2.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/RepViT_M2_3.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/RepViT_M3.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 236
+      size: 235
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/ViT_Base_Patch16_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/ViT_Base_Patch32_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/ViT_Base_Patch8_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/ViT_Large_Patch16_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/ViT_Small_Patch16_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/ViT_Small_Patch32_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/ViT_Tiny_Patch16_224.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/VisFormer_Small.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/models/VisFormer_Tiny.yaml

@@ -7,7 +7,7 @@ DEFAULT:
     Reader:
       style: pil
     Resize:
-      size: 249
+      size: 248
       interpolation: bicubic
     CenterCrop:
       size:

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/utils/evaluation/eval_ade20k.py

@@ -41,6 +41,21 @@ class SemanticSegmentationResult(NamedTuple):
 ADE20KResult = SemanticSegmentationResult
 
 
+def _class_id_mask(values: np.ndarray, nc: int) -> np.ndarray:
+    """Whether each entry is a finite, integral class ID in ``[0, nc)``.
+
+    The range test already rejects NaN and both infinities, so only a
+    floating-point array needs the integral test on top of it and an integer
+    array needs neither. On a full-resolution class map each of those passes
+    costs about as much as the histogram they guard.
+    """
+
+    inside = (values >= 0) & (values < nc)
+    if np.issubdtype(np.asarray(values).dtype, np.integer):
+        return inside
+    return inside & (values == np.floor(values))
+
+
 class SemanticMetricAccumulator:
     """Accumulate an ignore-aware semantic confusion matrix."""
 
@@ -73,12 +88,7 @@ class SemanticMetricAccumulator:
             raise ValueError(
                 f"Semantic prediction and target shapes must match, got {prediction.shape} and {target.shape}."
             )
-        valid_target = (
-            np.isfinite(target)
-            & (target >= 0)
-            & (target < self.nc)
-            & (target == np.floor(target))
-        )
+        valid_target = _class_id_mask(target, self.nc)
         allowed_target = valid_target | (target == self.ignore_label)
         if not bool(allowed_target.all()):
             invalid_values = np.asarray(np.unique(target[~allowed_target]))
@@ -86,23 +96,23 @@ class SemanticMetricAccumulator:
                 f"Semantic targets must be finite class IDs in [0, {self.nc - 1}] "
                 f"or ignore label {self.ignore_label}; got {invalid_values.tolist()}."
             )
-        valid_prediction = (
-            np.isfinite(prediction)
-            & (prediction >= 0)
-            & (prediction < self.nc)
-            & (prediction == np.floor(prediction))
-        )
-        invalid_prediction = valid_target & ~valid_prediction
-        if invalid_prediction.any():
-            invalid_values = np.asarray(np.unique(prediction[invalid_prediction]))
+        # Gather both class maps once: on a full-resolution map the boolean
+        # gather costs about as much as the histogram, and validating
+        # predictions only ever concerned the pixels with a valid target, so
+        # checking the gathered values is the same test over far fewer of them.
+        valid_targets = target[valid_target].astype(np.int64, copy=False)
+        valid_predictions = prediction[valid_target]
+        valid_prediction = _class_id_mask(valid_predictions, self.nc)
+        if not bool(valid_prediction.all()):
+            invalid_values = np.asarray(np.unique(valid_predictions[~valid_prediction]))
             raise ValueError(
                 f"Semantic predictions at valid target pixels must be finite class IDs in [0, {self.nc - 1}], "
                 f"got {invalid_values.tolist()}."
             )
-        if valid_target.any():
+        if valid_targets.size:
             histogram = np.bincount(
-                self.nc * target[valid_target].astype(np.int64)
-                + prediction[valid_target].astype(np.int64),
+                self.nc * valid_targets
+                + valid_predictions.astype(np.int64, copy=False),
                 minlength=self.nc**2,
             )
             self.matrix += histogram.reshape(self.nc, self.nc)

{mblt_vision_python-0.0.3 → mblt_vision_python-0.0.4}/mblt_vision/utils/evaluation/eval_dota.py

@@ -447,16 +447,40 @@ def _match_predictions(
 
     correct_class = true_classes[:, None] == pred_classes
     iou_np = (iou * correct_class).cpu().numpy()
-    for iou_index, threshold in enumerate(iouv.cpu().tolist()):
-        matches = np.array(np.nonzero(iou_np >= threshold)).T
-        if matches.shape[0] == 0:
-            continue
-        if matches.shape[0] > 1:
-            matches = matches[iou_np[matches[:, 0], matches[:, 1]].argsort()[::-1]]
-            # Preserve the IoU-ranked order while keeping the first match per prediction and target.
-            matches = matches[np.sort(np.unique(matches[:, 1], return_index=True)[1])]
-            matches = matches[np.sort(np.unique(matches[:, 0], return_index=True)[1])]
-        correct[matches[:, 1].astype(int), iou_index] = True
+    thresholds = np.asarray(iouv.cpu().tolist(), dtype=np.float64)
+    # The candidate set at any threshold is a prefix of the IoU-sorted
+    # candidate set at the lowest one, and "first occurrence per prediction,
+    # then per target" over a prefix keeps exactly the entries of the full-list
+    # result that fall inside it. So one sort and one filter answer all ten
+    # thresholds: record the IoU each surviving match was made at and compare
+    # that against the threshold vector, rather than re-sorting ten times.
+    #
+    # `kind="stable"` is what makes that prefix argument true, and it is load
+    # bearing rather than tidy. With an unstable sort, equal IoUs can be
+    # ordered one way in the full candidate set and another when the
+    # lower-IoU entries are gone, so the prefix stops being the prefix and the
+    # collapse picks different pairs than a per-threshold pass would. Measured
+    # over 4000 tie-heavy matrices: stable agrees with a per-threshold pass on
+    # all of them, and an unstable sort disagrees on 45.
+    #
+    # Stability also removes a defect of the per-threshold form it replaces.
+    # Re-sorting each threshold independently let a prediction come out a true
+    # positive at IoU >= 0.75 while not being one at IoU >= 0.50 — 58
+    # non-monotone prediction rows in that same sweep, where both the stable
+    # per-threshold form and this one produce none. Correctness here is
+    # agreement with the stable ordering, not with a particular unstable run.
+    matches = np.array(np.nonzero(iou_np >= thresholds.min())).T
+    if matches.shape[0] == 0:
+        return correct
+    if matches.shape[0] > 1:
+        values = iou_np[matches[:, 0], matches[:, 1]]
+        matches = matches[values.argsort(kind="stable")[::-1]]
+        # Preserve the IoU-ranked order while keeping the first match per prediction and target.
+        matches = matches[np.sort(np.unique(matches[:, 1], return_index=True)[1])]
+        matches = matches[np.sort(np.unique(matches[:, 0], return_index=True)[1])]
+    predictions = matches[:, 1].astype(int)
+    matched_iou = iou_np[matches[:, 0], matches[:, 1]]
+    correct[predictions] = matched_iou[:, None] >= thresholds
     return correct
 
 
@@ -560,9 +584,12 @@ def _process_image_stats(
     if ignore_classes.numel() and predictions["cls"].numel():
         ignored_iou = batch_probiou(ignore_boxes, predictions["bboxes"])
         same_class = ignore_classes[:, None] == predictions["cls"]
-        ignored_matches = (
-            (ignored_iou[:, :, None] >= iouv[None, None, :]) & same_class[:, :, None]
-        ).any(dim=0)
+        # A prediction is ignored at a threshold exactly when the best
+        # same-class ignored target reaches it, so one reduction over the
+        # targets replaces materialising a targets x predictions x thresholds
+        # boolean cube.
+        best = ignored_iou.masked_fill(~same_class, -1.0).amax(dim=0)
+        ignored_matches = best[:, None] >= iouv[None, :]
         ignore = ignored_matches.cpu().numpy() & ~true_positive
     return {
         "tp": true_positive,