htr_demo / src /htr_pipeline /gradio_backend.py
Gabriel's picture
0.0.3 release with Trocr and compare support
c60ebd1
import os
import xml.etree.ElementTree as ET
from difflib import Differ
import cv2
import evaluate
import gradio as gr
import numpy as np
import pandas as pd
from src.htr_pipeline.inferencer import Inferencer, InferencerInterface
from src.htr_pipeline.pipeline import Pipeline, PipelineInterface
from src.htr_pipeline.utils.helper import gradio_info
class SingletonModelLoader:
_instance = None
def __new__(cls, *args, **kwargs):
if not cls._instance:
cls._instance = super(SingletonModelLoader, cls).__new__(cls, *args, **kwargs)
return cls._instance
def __init__(self):
if not hasattr(self, "inferencer"):
self.inferencer = Inferencer(local_run=True)
if not hasattr(self, "pipeline"):
self.pipeline = Pipeline(self.inferencer)
def handling_callback_stop_inferencer():
from src.htr_pipeline.utils import pipeline_inferencer
pipeline_inferencer.terminate = False
# fast track
class FastTrack:
def __init__(self, model_loader):
self.pipeline: PipelineInterface = model_loader.pipeline
def segment_to_xml(self, image, radio_button_choices, htr_tool_transcriber_model_dropdown):
handling_callback_stop_inferencer()
gr.Info("Excuting HTR on image")
xml_xml = "page_xml.xml"
xml_txt = "page_txt.txt"
if os.path.exists(f"./{xml_xml}"):
os.remove(f"./{xml_xml}")
htr_tool_transcriber_model_dropdown
rendered_xml = self.pipeline.running_htr_pipeline(image, htr_tool_transcriber_model_dropdown)
with open(xml_xml, "w") as f:
f.write(rendered_xml)
if os.path.exists(f"./{xml_txt}"):
os.remove(f"./{xml_txt}")
self.pipeline.parse_xml_to_txt()
returned_file_extension = self.file_extenstion_to_return(radio_button_choices, xml_xml, xml_txt)
return returned_file_extension, gr.update(visible=True)
def visualize_image_viewer(self, image):
xml_img, text_polygon_dict = self.pipeline.visualize_xml(image)
return xml_img, text_polygon_dict
def file_extenstion_to_return(self, radio_button_choices, xml_xml, xml_txt):
if len(radio_button_choices) < 2:
if radio_button_choices[0] == "Txt":
returned_file_extension = xml_txt
else:
returned_file_extension = xml_xml
else:
returned_file_extension = [xml_txt, xml_xml]
return returned_file_extension
def get_text_from_coords(self, text_polygon_dict, evt: gr.SelectData):
x, y = evt.index[0], evt.index[1]
for text, polygon_coords in text_polygon_dict.items():
if (
cv2.pointPolygonTest(np.array(polygon_coords), (x, y), False) >= 0
): # >= 0 means on the polygon or inside
return text
def segment_to_xml_api(self, image):
rendered_xml = self.pipeline.running_htr_pipeline(image)
return rendered_xml
# Custom track
class CustomTrack:
def __init__(self, model_loader):
self.inferencer: InferencerInterface = model_loader.inferencer
@gradio_info("Running Segment Region")
def region_segment(self, image, pred_score_threshold, containments_treshold):
predicted_regions, regions_cropped_ordered, _, _ = self.inferencer.predict_regions(
image, pred_score_threshold, containments_treshold
)
return predicted_regions, regions_cropped_ordered, gr.update(visible=False), gr.update(visible=True)
@gradio_info("Running Segment Line")
def line_segment(self, image, pred_score_threshold, containments_threshold):
predicted_lines, lines_cropped_ordered, _ = self.inferencer.predict_lines(
image, pred_score_threshold, containments_threshold
)
return (
predicted_lines,
image,
lines_cropped_ordered,
lines_cropped_ordered, #
lines_cropped_ordered, # temp_gallery
gr.update(visible=True),
gr.update(visible=True),
gr.update(visible=False),
gr.update(visible=True),
)
def transcribe_text(self, images):
gr.Info("Running Transcribe Lines")
transcription_temp_list_with_score = []
mapping_dict = {}
total_images = len(images)
current_index = 0
bool_to_show_placeholder = gr.update(visible=True)
bool_to_show_control_results_transcribe = gr.update(visible=False)
for image in images:
current_index += 1
if current_index == total_images:
bool_to_show_control_results_transcribe = gr.update(visible=True)
bool_to_show_placeholder = gr.update(visible=False)
transcribed_text, prediction_score_from_htr = self.inferencer.transcribe(image)
transcription_temp_list_with_score.append((transcribed_text, prediction_score_from_htr))
df_trans_explore = pd.DataFrame(
transcription_temp_list_with_score, columns=["Transcribed text", "Pred score"]
)
joined_transcription_temp_list = "\n".join([tup[0] for tup in transcription_temp_list_with_score])
mapping_dict[transcribed_text] = image
yield joined_transcription_temp_list, df_trans_explore, mapping_dict, bool_to_show_control_results_transcribe, bool_to_show_placeholder
def get_select_index_image(self, images_from_gallery, evt: gr.SelectData):
return images_from_gallery[evt.index]["name"]
def get_select_index_df(self, transcribed_text_df_finish, mapping_dict, evt: gr.SelectData):
df_list = transcribed_text_df_finish["Transcribed text"].tolist()
key_text = df_list[evt.index[0]]
sorted_image = mapping_dict[key_text]
new_first = [sorted_image]
new_list = [img for txt, img in mapping_dict.items() if txt != key_text]
new_first.extend(new_list)
return new_first, key_text
def download_df_to_txt(self, transcribed_df):
text_in_list = transcribed_df["Transcribed text"].tolist()
file_name = "./transcribed_text.txt"
text_file = open(file_name, "w")
for text in text_in_list:
text_file.write(text + "\n")
text_file.close()
return file_name, gr.update(visible=True)
# Temporary structured here...
def upload_file(files):
return files.name, gr.update(visible=True)
def diff_texts(text1, text2):
d = Differ()
return [(token[2:], token[0] if token[0] != " " else None) for token in d.compare(text1, text2)]
def compute_cer_a_and_b_with_gt(run_a, run_b, run_gt):
text_run_a, text_run_b, text_run_gt = reading_xml_files_string(run_a, run_b, run_gt)
cer_metric = evaluate.load("cer")
if text_run_a == text_run_gt:
return "No Ground Truth was provided."
elif text_run_a == text_run_b:
return f"A & B -> GT: {round(cer_metric.compute(predictions=[text_run_a], references=[text_run_gt]), 4)}"
else:
return f"A -> GT: {round(cer_metric.compute(predictions=[text_run_a], references=[text_run_gt]), 4)}, B -> GT {round(cer_metric.compute(predictions=[text_run_b], references=[text_run_gt]), 4)}"
def temporary_xml_parser(page_xml):
tree = ET.parse(page_xml, parser=ET.XMLParser(encoding="utf-8"))
root = tree.getroot()
namespace = "{http://schema.primaresearch.org/PAGE/gts/pagecontent/2013-07-15}"
text_list = []
for textregion in root.findall(f".//{namespace}TextRegion"):
for textline in textregion.findall(f".//{namespace}TextLine"):
text = textline.find(f"{namespace}TextEquiv").find(f"{namespace}Unicode").text
text_list.append(text)
return " ".join(text_list)
def compare_diff_runs_highlight(run_a, run_b, run_gt):
text_run_a, text_run_b, text_run_gt = reading_xml_files_string(run_a, run_b, run_gt)
diff_runs = diff_texts(text_run_a, text_run_b)
diff_gt = diff_texts(text_run_a, text_run_gt)
return diff_runs, diff_gt
def reading_xml_files_string(run_a, run_b, run_gt):
if run_a is None:
return
if run_gt is None:
gr.Warning("No GT was provided, setting GT to A")
run_gt = run_a
if run_b is None:
gr.Warning("No B was provided, setting B to A")
run_b = run_a
text_run_a = temporary_xml_parser(run_a.name)
text_run_b = temporary_xml_parser(run_b.name)
text_run_gt = temporary_xml_parser(run_gt.name)
return text_run_a, text_run_b, text_run_gt
def update_selected_tab_output_and_setting():
return gr.update(visible=True), gr.update(visible=False), gr.update(visible=False)
def update_selected_tab_image_viewer():
return gr.update(visible=False), gr.update(visible=True), gr.update(visible=False)
def update_selected_tab_model_compare():
return gr.update(visible=False), gr.update(visible=False), gr.update(visible=True)
if __name__ == "__main__":
pass