enhancement: optimize outputs to objects metbo2xp (#443)

From codeflash: Optimizes table model pipeline by keeping the data as a
tensor for longer, enabling vector operations and reducing conversions.

---------

Co-authored-by: codeflash-ai[bot] <148906541+codeflash-ai[bot]@users.noreply.github.com>
Co-authored-by: aseembits93 <[email protected]>

Author: 3 people

CHANGELOG.md

@@ -1,3 +1,7 @@
+## 1.0.8-dev0
+
+* Enhancement: Optimized `outputs_to_objects` for an 88% speedup in some cases (codeflash)
+
 ## 1.0.7
 
 * Fix a hardcoded file extension causing confusion in the logs

unstructured_inference/__version__.py

@@ -1 +1 @@
-__version__ = "1.0.7"  # pragma: no cover
+__version__ = "1.0.8-dev0"  # pragma: no cover

unstructured_inference/models/tables.py

@@ -213,16 +213,18 @@ def outputs_to_objects(
 ):
     """Output table element types."""
     m = outputs["logits"].softmax(-1).max(-1)
-    pred_labels = list(m.indices.detach().cpu().numpy())[0]
-    pred_scores = list(m.values.detach().cpu().numpy())[0]
+    pred_labels = m.indices.detach().cpu().numpy()[0]
+    pred_scores = m.values.detach().cpu().numpy()[0]
     pred_bboxes = outputs["pred_boxes"].detach().cpu()[0]
 
     pad = outputs.get("pad_for_structure_detection", 0)
     scale_size = (img_size[0] + pad * 2, img_size[1] + pad * 2)
-    pred_bboxes = [elem.tolist() for elem in rescale_bboxes(pred_bboxes, scale_size)]
+    rescaled = rescale_bboxes(pred_bboxes, scale_size)
     # unshift the padding; padding effectively shifted the bounding boxes of structures in the
     # original image with half of the total pad
-    shift_size = pad
+    if pad != 0:
+        rescaled = rescaled - pad
+    pred_bboxes = rescaled.tolist()
 
     objects = []
     for label, score, bbox in zip(pred_labels, pred_scores, pred_bboxes):
@@ -232,7 +234,7 @@ def outputs_to_objects(
                 {
                     "label": class_label,
                     "score": float(score),
-                    "bbox": [float(elem) - shift_size for elem in bbox],
+                    "bbox": bbox,
                 },
             )
 
@@ -279,7 +281,7 @@ def rescale_bboxes(out_bbox, size):
     """Rescale relative bounding box to box of size given by size."""
     img_w, img_h = size
     b = box_cxcywh_to_xyxy(out_bbox)
-    b = b * torch.tensor([img_w, img_h, img_w, img_h], dtype=torch.float32)
+    b = b * torch.tensor([img_w, img_h, img_w, img_h], dtype=torch.float32, device=out_bbox.device)
     return b