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

disable autosized prediction entirely (also for _patched)…

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

But CUDA-based tiling also increases GPU memory requirements
a lot. And with the new parallel subprocess predictors, Numpy-
based tiling is not necessarily slower anymore.
This commit is contained in:
Robert Sachunsky 2026-04-10 18:23:10 +02:00
parent ccef63f08b
commit 0d21b62aee
1 changed files with 22 additions and 20 deletions
Download patch file Download diff file Expand all files Collapse all files

42
src/eynollah/eynollah.py
View file

@ -179,9 +179,9 @@ class Eynollah:
] ]
if self.input_binary: if self.input_binary:
loadable.append("binarization") # todo: binarization_patched loadable.append("binarization") # todo: binarization_patched
loadable.append("textline_patched") # textline loadable.append("textline") # textline_patched
loadable.append("region_1_2") loadable.append("region_1_2")
loadable.append("region_1_2_patched") #loadable.append("region_1_2_patched")
if self.full_layout: if self.full_layout:
loadable.append("region_fl_np") loadable.append("region_fl_np")
#loadable.append("region_fl_patched") #loadable.append("region_fl_patched")
@ -914,10 +914,10 @@ class Eynollah:
img = resize_image(img, int(img_height_h * 2500 / float(img_width_h)), 2500).astype(np.uint8) img = resize_image(img, int(img_height_h * 2500 / float(img_width_h)), 2500).astype(np.uint8)
if patches: if patches:
prediction_regions, _ = self.do_prediction_new_concept_autosize( # prediction_regions, _ = self.do_prediction_new_concept_autosize(
img, self.model_zoo.get("region_fl_patched"), # img, self.model_zoo.get("region_fl_patched"),
# prediction_regions, _ = self.do_prediction_new_concept( prediction_regions, _ = self.do_prediction_new_concept(
# True, img, self.model_zoo.get("region_fl"), True, img, self.model_zoo.get("region_fl"),
n_batch_inference=2, n_batch_inference=2,
thresholding_for_heading=True) thresholding_for_heading=True)
else: else:
@ -1075,10 +1075,10 @@ class Eynollah:
thresholding_for_artificial_class=True, thresholding_for_artificial_class=True,
threshold_art_class=self.threshold_art_class_textline) threshold_art_class=self.threshold_art_class_textline)
if use_patches: if use_patches:
prediction_textline, _ = self.do_prediction_new_concept_autosize( # prediction_textline, _ = self.do_prediction_new_concept_autosize(
img, self.model_zoo.get("textline_patched"), **kwargs) # img, self.model_zoo.get("textline_patched"), **kwargs)
# prediction_textline, _ = self.do_prediction_new_concept( prediction_textline, _ = self.do_prediction_new_concept(
# True, img, self.model_zoo.get("textline"), **kwargs) True, img, self.model_zoo.get("textline"), **kwargs)
else: else:
prediction_textline, _ = self.do_prediction_new_concept( prediction_textline, _ = self.do_prediction_new_concept(
False, img, self.model_zoo.get("textline"), **kwargs) False, img, self.model_zoo.get("textline"), **kwargs)
@ -1136,12 +1136,13 @@ class Eynollah:
if img_height_h / img_width_h > 2.5: if img_height_h / img_width_h > 2.5:
self.logger.debug("resized to %dx%d for %d cols", self.logger.debug("resized to %dx%d for %d cols",
img_resized.shape[1], img_resized.shape[0], num_col_classifier) img_resized.shape[1], img_resized.shape[0], num_col_classifier)
prediction_regions_org, confidence_matrix = \ prediction_regions_org, confidence_matrix = (
self.do_prediction_new_concept_autosize( # self.do_prediction_new_concept_autosize(
img_resized, self.model_zoo.get("region_1_2_patched"), # img_resized, self.model_zoo.get("region_1_2_patched"),
# self.do_prediction_new_concept( self.do_prediction_new_concept(
# True, img_resized, self.model_zoo.get("region_1_2"), True, img_resized, self.model_zoo.get("region_1_2"),
**kwargs) **kwargs)
)
else: else:
prediction_regions_org, confidence_matrix = \ prediction_regions_org, confidence_matrix = \
self.do_prediction_new_concept( self.do_prediction_new_concept(
@ -1154,12 +1155,13 @@ class Eynollah:
self.logger.debug("resized to %dx%d (new_w=%d) for %d cols", self.logger.debug("resized to %dx%d (new_w=%d) for %d cols",
img_resized.shape[1], img_resized.shape[0], img_resized.shape[1], img_resized.shape[0],
new_w, num_col_classifier) new_w, num_col_classifier)
prediction_regions_org, confidence_matrix = \ prediction_regions_org, confidence_matrix = (
self.do_prediction_new_concept_autosize( # self.do_prediction_new_concept_autosize(
img_resized, self.model_zoo.get("region_1_2_patched"), # img_resized, self.model_zoo.get("region_1_2_patched"),
# self.do_prediction_new_concept( self.do_prediction_new_concept(
# True, img_resized, self.model_zoo.get("region_1_2"), True, img_resized, self.model_zoo.get("region_1_2"),
**kwargs) **kwargs)
)
prediction_regions_org = resize_image(prediction_regions_org, img_height_h, img_width_h ) prediction_regions_org = resize_image(prediction_regions_org, img_height_h, img_width_h )
confidence_matrix = resize_image(confidence_matrix, img_height_h, img_width_h ) confidence_matrix = resize_image(confidence_matrix, img_height_h, img_width_h )