qurator-spk/eynollah
1
0
Fork 0
mirror of https://github.com/qurator-spk/eynollah.git synced 2026-04-14 19:31:57 +02:00
Code Issues Projects Releases Wiki Activity

autosized prediction is only faster for _patched, not for _resized…

Browse source 
When 338c4a0e wrapped all prediction models for automatic
image size adaptation in CUDA,
- tiling (`_patched`) was indeed faster
- whole  (`_resized`) was actually slower

So this reverts the latter part.
This commit is contained in:
Robert Sachunsky 2026-03-13 18:15:30 +01:00
parent 6d55f297a5
commit 576e120ba6
2 changed files with 17 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

22
src/eynollah/eynollah.py
View file

@ -889,8 +889,8 @@ class Eynollah:
self.logger.debug("enter extract_text_regions")
img_height_h = img.shape[0]
img_width_h = img.shape[1]
#model_name = "region_fl" if patches else "region_fl_np"
model_name = "region_fl_patched" if patches else "region_fl_np_resized"
#model_name = "region_fl_patched" if patches else "region_fl_np_resized"
model_name = "region_fl_patched" if patches else "region_fl_np"
model_region = self.model_zoo.get(model_name)
thresholding_for_heading = True
@ -911,9 +911,15 @@ class Eynollah:
else:
img = resize_image(img, int(img_height_h * 2500 / float(img_width_h)), 2500).astype(np.uint8)
prediction_regions, _ = self.do_prediction_new_concept_autosize(
img, model_region,
thresholding_for_heading=thresholding_for_heading)
if patches:
prediction_regions, _ = self.do_prediction_new_concept_autosize(
img, model_region,
thresholding_for_heading=True)
else:
prediction_regions = self.do_prediction(
False, img, model_region,
n_batch_inference=2,
thresholding_for_heading=False)
prediction_regions = resize_image(prediction_regions, img_height_h, img_width_h)
self.logger.debug("exit extract_text_regions")
return prediction_regions
@ -1125,8 +1131,8 @@ class Eynollah:
prediction_regions_org = np.zeros((img_height_org, img_width_org), dtype=np.uint8)
confidence_matrix = np.zeros((img_height_org, img_width_org))
prediction_regions_page, confidence_matrix_page = \
self.do_prediction_new_concept_autosize(
image['img_page'], self.model_zoo.get("region_1_2_resized"),
self.do_prediction_new_concept(
False, image['img_page'], self.model_zoo.get("region_1_2"),
thresholding_for_artificial_class=True,
threshold_art_class=self.threshold_art_class_layout)
ys = slice(*image['coord_page'][0:2])
@ -1523,7 +1529,7 @@ class Eynollah:
return image_revised_last
def get_tables_from_model(self, img):
prediction_table, _ = self.do_prediction_new_concept_autosize(img, self.model_zoo.get("table_resized"))
prediction_table, _ = self.do_prediction_new_concept(False, img, self.model_zoo.get("table"))
prediction_table = prediction_table.astype(np.uint8)
return prediction_table

5
src/eynollah/model_zoo/model_zoo.py
View file

@ -143,8 +143,9 @@ class EynollahModelZoo:
)
model._name = model_category
self._loaded[model_category] = model
if model_category in ['region_1_2', 'table', 'region_fl_np']:
self._loaded[model_category + '_resized'] = wrap_layout_model_resized(model)
# autosized for full page images is too slow (better resize on CPU in numpy):
# if model_category in ['region_1_2', 'table', 'region_fl_np']:
# self._loaded[model_category + '_resized'] = wrap_layout_model_resized(model)
if model_category in ['region_1_2', 'textline']:
self._loaded[model_category + '_patched'] = wrap_layout_model_patched(model)
return model # type: ignore