Spaces:
Runtime error
Runtime error
initial commit
Browse files- AUTHORS.rst +13 -0
- README.md +17 -7
- app.py +740 -0
- components/callbacks.py +4 -0
- components/data_module.py +81 -0
- config.py +3 -0
- config/config.yaml +0 -0
- data/.gitkeep +0 -0
- docker-compose.yml +23 -0
- models/__init__.py +0 -0
- models/base_model/classification.py +63 -0
- models/base_model/gan.py +0 -0
- models/base_model/regression.py +55 -0
- models/metrics/classification.py +44 -0
- models/metrics/regression.py +28 -0
- models/model_lit.py +50 -0
- models/modules/sample_torch_module.py +12 -0
- requirements.txt +10 -0
- test_pdf2img.py +16 -0
- tests/test_resource.py +4 -0
- utils/.gitkeep +0 -0
AUTHORS.rst
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
=======
|
| 2 |
+
Credits
|
| 3 |
+
=======
|
| 4 |
+
|
| 5 |
+
Development Lead
|
| 6 |
+
----------------
|
| 7 |
+
|
| 8 |
+
* Full name of the author <Email of the author>
|
| 9 |
+
|
| 10 |
+
Contributors
|
| 11 |
+
------------
|
| 12 |
+
|
| 13 |
+
None yet. Why not be the first?
|
README.md
CHANGED
|
@@ -1,13 +1,23 @@
|
|
| 1 |
---
|
| 2 |
-
title: Table
|
| 3 |
-
emoji:
|
| 4 |
-
colorFrom:
|
| 5 |
-
colorTo:
|
| 6 |
sdk: streamlit
|
| 7 |
-
sdk_version: 1.
|
| 8 |
app_file: app.py
|
| 9 |
pinned: false
|
| 10 |
-
license: mit
|
| 11 |
---
|
|
|
|
| 12 |
|
| 13 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
title: Table Extraction (Table Transformer + PaddleOCR)
|
| 3 |
+
emoji: 🚀
|
| 4 |
+
colorFrom: blue
|
| 5 |
+
colorTo: yellow
|
| 6 |
sdk: streamlit
|
| 7 |
+
sdk_version: 1.21.0
|
| 8 |
app_file: app.py
|
| 9 |
pinned: false
|
|
|
|
| 10 |
---
|
| 11 |
+
# huggingface-space
|
| 12 |
|
| 13 |
+
Imported from https://huggingface.co/spaces/jurgendn/table-extraction with some adjustment.
|
| 14 |
+
|
| 15 |
+
Current pipeline:
|
| 16 |
+
|
| 17 |
+
Table detection: https://huggingface.co/microsoft/table-transformer-detection
|
| 18 |
+
|
| 19 |
+
Table recognition: https://huggingface.co/microsoft/table-transformer-structure-recognition
|
| 20 |
+
|
| 21 |
+
OCR: https://github.com/pbcquoc/vietocr
|
| 22 |
+
|
| 23 |
+
OCR-new: https://github.com/PaddlePaddle/PaddleOCR
|
app.py
ADDED
|
@@ -0,0 +1,740 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import asyncio
|
| 2 |
+
import string
|
| 3 |
+
import random
|
| 4 |
+
from collections import Counter
|
| 5 |
+
from itertools import count, tee
|
| 6 |
+
import base64
|
| 7 |
+
|
| 8 |
+
import cv2
|
| 9 |
+
import matplotlib.pyplot as plt
|
| 10 |
+
import numpy as np
|
| 11 |
+
import pandas as pd
|
| 12 |
+
import streamlit as st
|
| 13 |
+
import torch
|
| 14 |
+
from PIL import Image
|
| 15 |
+
from transformers import DetrImageProcessor, TableTransformerForObjectDetection
|
| 16 |
+
from paddleocr import PaddleOCR
|
| 17 |
+
|
| 18 |
+
ocr = PaddleOCR(use_angle_cls=True, lang="en", use_gpu=False, ocr_version='PP-OCRv3')
|
| 19 |
+
|
| 20 |
+
st.set_option('deprecation.showPyplotGlobalUse', False)
|
| 21 |
+
st.set_page_config(layout='wide')
|
| 22 |
+
st.title("Table Detection and Table Structure Recognition")
|
| 23 |
+
st.write(
|
| 24 |
+
"Implemented by MSFT team: https://github.com/microsoft/table-transformer")
|
| 25 |
+
|
| 26 |
+
table_detection_model = TableTransformerForObjectDetection.from_pretrained(
|
| 27 |
+
"microsoft/table-transformer-detection")
|
| 28 |
+
|
| 29 |
+
table_recognition_model = TableTransformerForObjectDetection.from_pretrained(
|
| 30 |
+
"microsoft/table-transformer-structure-recognition")
|
| 31 |
+
|
| 32 |
+
def reload_ocr(vlang):
|
| 33 |
+
global ocr
|
| 34 |
+
ocr = PaddleOCR(use_angle_cls=True, lang=vlang, use_gpu=False, ocr_version='PP-OCRv4')
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
def PIL_to_cv(pil_img):
|
| 38 |
+
return cv2.cvtColor(np.array(pil_img), cv2.COLOR_RGB2BGR)
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
def cv_to_PIL(cv_img):
|
| 42 |
+
return Image.fromarray(cv2.cvtColor(cv_img, cv2.COLOR_BGR2RGB))
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
async def pytess(cell_pil_img, threshold: float = 0.5):
|
| 46 |
+
cell_pil_img = TableExtractionPipeline.add_padding(pil_img=cell_pil_img, top=20, right=10, bottom=20, left=10, color=(255, 255, 255))
|
| 47 |
+
result = ocr.ocr(np.asarray(cell_pil_img), cls=True)[0]
|
| 48 |
+
|
| 49 |
+
#Debug
|
| 50 |
+
# filename = str(random.random())
|
| 51 |
+
# cell_pil_img.save("dump/" + filename + ".png")
|
| 52 |
+
# print(filename)
|
| 53 |
+
# print(result)
|
| 54 |
+
|
| 55 |
+
text = ""
|
| 56 |
+
if result != None:
|
| 57 |
+
txts = [line[1][0] for line in result]
|
| 58 |
+
text = " ".join(txts)
|
| 59 |
+
return text
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
def sharpen_image(pil_img):
|
| 63 |
+
|
| 64 |
+
img = PIL_to_cv(pil_img)
|
| 65 |
+
sharpen_kernel = np.array([[-1, -1, -1], [-1, 9, -1], [-1, -1, -1]])
|
| 66 |
+
|
| 67 |
+
sharpen = cv2.filter2D(img, -1, sharpen_kernel)
|
| 68 |
+
pil_img = cv_to_PIL(sharpen)
|
| 69 |
+
return pil_img
|
| 70 |
+
|
| 71 |
+
|
| 72 |
+
def uniquify(seq, suffs=count(1)):
|
| 73 |
+
"""Make all the items unique by adding a suffix (1, 2, etc).
|
| 74 |
+
Credit: https://stackoverflow.com/questions/30650474/python-rename-duplicates-in-list-with-progressive-numbers-without-sorting-list
|
| 75 |
+
`seq` is mutable sequence of strings.
|
| 76 |
+
`suffs` is an optional alternative suffix iterable.
|
| 77 |
+
"""
|
| 78 |
+
not_unique = [k for k, v in Counter(seq).items() if v > 1]
|
| 79 |
+
|
| 80 |
+
suff_gens = dict(zip(not_unique, tee(suffs, len(not_unique))))
|
| 81 |
+
for idx, s in enumerate(seq):
|
| 82 |
+
try:
|
| 83 |
+
suffix = str(next(suff_gens[s]))
|
| 84 |
+
except KeyError:
|
| 85 |
+
continue
|
| 86 |
+
else:
|
| 87 |
+
seq[idx] += suffix
|
| 88 |
+
|
| 89 |
+
return seq
|
| 90 |
+
|
| 91 |
+
|
| 92 |
+
def binarizeBlur_image(pil_img):
|
| 93 |
+
image = PIL_to_cv(pil_img)
|
| 94 |
+
thresh = cv2.threshold(image, 150, 255, cv2.THRESH_BINARY_INV)[1]
|
| 95 |
+
|
| 96 |
+
result = cv2.GaussianBlur(thresh, (5, 5), 0)
|
| 97 |
+
result = 255 - result
|
| 98 |
+
return cv_to_PIL(result)
|
| 99 |
+
|
| 100 |
+
|
| 101 |
+
def td_postprocess(pil_img):
|
| 102 |
+
'''
|
| 103 |
+
Removes gray background from tables
|
| 104 |
+
'''
|
| 105 |
+
img = PIL_to_cv(pil_img)
|
| 106 |
+
|
| 107 |
+
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
|
| 108 |
+
mask = cv2.inRange(hsv, (0, 0, 100),
|
| 109 |
+
(255, 5, 255)) # (0, 0, 100), (255, 5, 255)
|
| 110 |
+
nzmask = cv2.inRange(hsv, (0, 0, 5),
|
| 111 |
+
(255, 255, 255)) # (0, 0, 5), (255, 255, 255))
|
| 112 |
+
nzmask = cv2.erode(nzmask, np.ones((3, 3))) # (3,3)
|
| 113 |
+
mask = mask & nzmask
|
| 114 |
+
|
| 115 |
+
new_img = img.copy()
|
| 116 |
+
new_img[np.where(mask)] = 255
|
| 117 |
+
|
| 118 |
+
return cv_to_PIL(new_img)
|
| 119 |
+
|
| 120 |
+
|
| 121 |
+
# def super_res(pil_img):
|
| 122 |
+
# # requires opencv-contrib-python installed without the opencv-python
|
| 123 |
+
# sr = dnn_superres.DnnSuperResImpl_create()
|
| 124 |
+
# image = PIL_to_cv(pil_img)
|
| 125 |
+
# model_path = "./LapSRN_x8.pb"
|
| 126 |
+
# model_name = model_path.split('/')[1].split('_')[0].lower()
|
| 127 |
+
# model_scale = int(model_path.split('/')[1].split('_')[1].split('.')[0][1])
|
| 128 |
+
|
| 129 |
+
# sr.readModel(model_path)
|
| 130 |
+
# sr.setModel(model_name, model_scale)
|
| 131 |
+
# final_img = sr.upsample(image)
|
| 132 |
+
# final_img = cv_to_PIL(final_img)
|
| 133 |
+
|
| 134 |
+
# return final_img
|
| 135 |
+
|
| 136 |
+
|
| 137 |
+
def table_detector(image, THRESHOLD_PROBA):
|
| 138 |
+
'''
|
| 139 |
+
Table detection using DEtect-object TRansformer pre-trained on 1 million tables
|
| 140 |
+
|
| 141 |
+
'''
|
| 142 |
+
|
| 143 |
+
feature_extractor = DetrImageProcessor(do_resize=True,
|
| 144 |
+
size=800,
|
| 145 |
+
max_size=800)
|
| 146 |
+
encoding = feature_extractor(image, return_tensors="pt")
|
| 147 |
+
|
| 148 |
+
with torch.no_grad():
|
| 149 |
+
outputs = table_detection_model(**encoding)
|
| 150 |
+
|
| 151 |
+
probas = outputs.logits.softmax(-1)[0, :, :-1]
|
| 152 |
+
keep = probas.max(-1).values > THRESHOLD_PROBA
|
| 153 |
+
|
| 154 |
+
target_sizes = torch.tensor(image.size[::-1]).unsqueeze(0)
|
| 155 |
+
postprocessed_outputs = feature_extractor.post_process(
|
| 156 |
+
outputs, target_sizes)
|
| 157 |
+
bboxes_scaled = postprocessed_outputs[0]['boxes'][keep]
|
| 158 |
+
|
| 159 |
+
return (probas[keep], bboxes_scaled)
|
| 160 |
+
|
| 161 |
+
|
| 162 |
+
def table_struct_recog(image, THRESHOLD_PROBA):
|
| 163 |
+
'''
|
| 164 |
+
Table structure recognition using DEtect-object TRansformer pre-trained on 1 million tables
|
| 165 |
+
'''
|
| 166 |
+
|
| 167 |
+
feature_extractor = DetrImageProcessor(do_resize=True,
|
| 168 |
+
size=1000,
|
| 169 |
+
max_size=1000)
|
| 170 |
+
encoding = feature_extractor(image, return_tensors="pt")
|
| 171 |
+
|
| 172 |
+
with torch.no_grad():
|
| 173 |
+
outputs = table_recognition_model(**encoding)
|
| 174 |
+
|
| 175 |
+
probas = outputs.logits.softmax(-1)[0, :, :-1]
|
| 176 |
+
keep = probas.max(-1).values > THRESHOLD_PROBA
|
| 177 |
+
|
| 178 |
+
target_sizes = torch.tensor(image.size[::-1]).unsqueeze(0)
|
| 179 |
+
postprocessed_outputs = feature_extractor.post_process(
|
| 180 |
+
outputs, target_sizes)
|
| 181 |
+
bboxes_scaled = postprocessed_outputs[0]['boxes'][keep]
|
| 182 |
+
|
| 183 |
+
return (probas[keep], bboxes_scaled)
|
| 184 |
+
|
| 185 |
+
|
| 186 |
+
class TableExtractionPipeline():
|
| 187 |
+
|
| 188 |
+
colors = ["red", "blue", "green", "yellow", "orange", "violet"]
|
| 189 |
+
|
| 190 |
+
# colors = ["red", "blue", "green", "red", "red", "red"]
|
| 191 |
+
|
| 192 |
+
@staticmethod
|
| 193 |
+
def add_padding(pil_img,
|
| 194 |
+
top,
|
| 195 |
+
right,
|
| 196 |
+
bottom,
|
| 197 |
+
left,
|
| 198 |
+
color=(255, 255, 255)):
|
| 199 |
+
'''
|
| 200 |
+
Image padding as part of TSR pre-processing to prevent missing table edges
|
| 201 |
+
'''
|
| 202 |
+
width, height = pil_img.size
|
| 203 |
+
new_width = width + right + left
|
| 204 |
+
new_height = height + top + bottom
|
| 205 |
+
result = Image.new(pil_img.mode, (new_width, new_height), color)
|
| 206 |
+
result.paste(pil_img, (left, top))
|
| 207 |
+
return result
|
| 208 |
+
|
| 209 |
+
@staticmethod
|
| 210 |
+
def dynamic_delta(xmin, ymin, xmax, ymax, delta_xmin, delta_ymin, delta_xmax, delta_ymax, pil_img):
|
| 211 |
+
offset_x = (xmax - xmin) * 0.05
|
| 212 |
+
offset_y = (ymax - ymin) * 0.05
|
| 213 |
+
|
| 214 |
+
w_img, h_img = pil_img.size
|
| 215 |
+
|
| 216 |
+
doxmin = xmin - (delta_xmin + offset_x)
|
| 217 |
+
if (doxmin < 0):
|
| 218 |
+
doxmin = 0
|
| 219 |
+
|
| 220 |
+
doymin = ymin - (delta_ymin + offset_y)
|
| 221 |
+
if (doymin < 0):
|
| 222 |
+
doymin = 0
|
| 223 |
+
|
| 224 |
+
doxmax = xmax + (delta_xmax + offset_x)
|
| 225 |
+
if (doxmax > w_img):
|
| 226 |
+
doxmax = w_img
|
| 227 |
+
|
| 228 |
+
doymax = ymax + (delta_ymax + offset_y)
|
| 229 |
+
if (doymax > h_img):
|
| 230 |
+
doymax = h_img
|
| 231 |
+
|
| 232 |
+
|
| 233 |
+
return doxmin, doymin, doxmax, doymax
|
| 234 |
+
|
| 235 |
+
@staticmethod
|
| 236 |
+
def get_cxy(pil_img, xmin, ymin, xmax, ymax, offset):
|
| 237 |
+
'''
|
| 238 |
+
get the possible position of table caption
|
| 239 |
+
'''
|
| 240 |
+
w_img, h_img = pil_img.size
|
| 241 |
+
c_xmin = xmin
|
| 242 |
+
c_xmax = xmax
|
| 243 |
+
|
| 244 |
+
delta_x = xmax - xmin
|
| 245 |
+
|
| 246 |
+
if delta_x / w_img > 0.5: # full page
|
| 247 |
+
c_xmin = 0
|
| 248 |
+
c_xmax = w_img
|
| 249 |
+
else:
|
| 250 |
+
cx_dist = c_xmax-c_xmin
|
| 251 |
+
delta_dist = w_img * 0.4 #0.4 is from 0.5 assumed that paper has padding around 0.1 of total width. In assumption that the paper is 2 column
|
| 252 |
+
print("cx_dist: " + str(cx_dist))
|
| 253 |
+
print("delta_dist: " + str(delta_dist))
|
| 254 |
+
if cx_dist < delta_dist:
|
| 255 |
+
d_off = int((delta_dist - cx_dist) / 2)
|
| 256 |
+
print("d_off: " + str(d_off))
|
| 257 |
+
c_xmin = c_xmin - d_off
|
| 258 |
+
if c_xmin < 0:
|
| 259 |
+
c_xmin = 0
|
| 260 |
+
c_xmax = c_xmax + d_off
|
| 261 |
+
if c_xmax > w_img:
|
| 262 |
+
c_xmax = w_img
|
| 263 |
+
|
| 264 |
+
|
| 265 |
+
if offset < 0:
|
| 266 |
+
c_ymin = ymin + offset
|
| 267 |
+
c_ymax = ymin
|
| 268 |
+
if c_ymin < 0:
|
| 269 |
+
c_ymin = 0
|
| 270 |
+
else:
|
| 271 |
+
c_ymin = ymax
|
| 272 |
+
c_ymax = ymax + offset
|
| 273 |
+
if c_ymax > h_img:
|
| 274 |
+
c_ymax = h_img
|
| 275 |
+
|
| 276 |
+
return c_xmin, c_ymin, c_xmax, c_ymax
|
| 277 |
+
|
| 278 |
+
def plot_results_detection(self, c1, model, pil_img, prob, boxes,
|
| 279 |
+
delta_xmin, delta_ymin, delta_xmax, delta_ymax):
|
| 280 |
+
'''
|
| 281 |
+
crop_tables and plot_results_detection must have same co-ord shifts because 1 only plots the other one updates co-ordinates
|
| 282 |
+
'''
|
| 283 |
+
# st.write('img_obj')
|
| 284 |
+
# st.write(pil_img)
|
| 285 |
+
plt.imshow(pil_img)
|
| 286 |
+
ax = plt.gca()
|
| 287 |
+
|
| 288 |
+
for p, (xmin, ymin, xmax, ymax) in zip(prob, boxes.tolist()):
|
| 289 |
+
cl = p.argmax()
|
| 290 |
+
xmin, ymin, xmax, ymax = self.dynamic_delta(xmin, ymin, xmax, ymax, delta_xmin, delta_ymin, delta_xmax, delta_ymax, pil_img)
|
| 291 |
+
ax.add_patch(
|
| 292 |
+
plt.Rectangle((xmin, ymin),
|
| 293 |
+
xmax - xmin,
|
| 294 |
+
ymax - ymin,
|
| 295 |
+
fill=False,
|
| 296 |
+
color='red',
|
| 297 |
+
linewidth=3))
|
| 298 |
+
text = f'{model.config.id2label[cl.item()]}: {p[cl]:0.2f}'
|
| 299 |
+
ax.text(xmin - 20,
|
| 300 |
+
ymin - 50,
|
| 301 |
+
text,
|
| 302 |
+
fontsize=10,
|
| 303 |
+
bbox=dict(facecolor='yellow', alpha=0.5))
|
| 304 |
+
|
| 305 |
+
# Caption possibility (bottom)
|
| 306 |
+
offset = 200
|
| 307 |
+
c_xmin, c_ymin, c_xmax, c_ymax = self.get_cxy(pil_img, xmin, ymin, xmax, ymax, offset)
|
| 308 |
+
|
| 309 |
+
ax.add_patch(
|
| 310 |
+
plt.Rectangle((c_xmin, c_ymin),
|
| 311 |
+
c_xmax - c_xmin,
|
| 312 |
+
c_ymax - c_ymin,
|
| 313 |
+
fill=False,
|
| 314 |
+
color='blue',
|
| 315 |
+
linewidth=1))
|
| 316 |
+
|
| 317 |
+
# Caption possibility (top)
|
| 318 |
+
offset = -200
|
| 319 |
+
c_xmin, c_ymin, c_xmax, c_ymax = self.get_cxy(pil_img, xmin, ymin, xmax, ymax, offset)
|
| 320 |
+
|
| 321 |
+
ax.add_patch(
|
| 322 |
+
plt.Rectangle((c_xmin, c_ymin),
|
| 323 |
+
c_xmax - c_xmin,
|
| 324 |
+
c_ymax - c_ymin,
|
| 325 |
+
fill=False,
|
| 326 |
+
color='green',
|
| 327 |
+
linewidth=1))
|
| 328 |
+
|
| 329 |
+
plt.axis('off')
|
| 330 |
+
c1.pyplot()
|
| 331 |
+
|
| 332 |
+
def crop_tables(self, pil_img, prob, boxes, delta_xmin, delta_ymin,
|
| 333 |
+
delta_xmax, delta_ymax):
|
| 334 |
+
'''
|
| 335 |
+
crop_tables and plot_results_detection must have same co-ord shifts because 1 only plots the other one updates co-ordinates
|
| 336 |
+
'''
|
| 337 |
+
cropped_img_list = []
|
| 338 |
+
|
| 339 |
+
for p, (xmin, ymin, xmax, ymax) in zip(prob, boxes.tolist()):
|
| 340 |
+
xmin, ymin, xmax, ymax = self.dynamic_delta(xmin, ymin, xmax, ymax, delta_xmin, delta_ymin, delta_xmax, delta_ymax, pil_img)
|
| 341 |
+
cropped_img = pil_img.crop((xmin, ymin, xmax, ymax))
|
| 342 |
+
cropped_img_list.append(cropped_img)
|
| 343 |
+
|
| 344 |
+
return cropped_img_list
|
| 345 |
+
|
| 346 |
+
def crop_caption(self, pil_img, prob, boxes, delta_xmin, delta_ymin,
|
| 347 |
+
delta_xmax, delta_ymax):
|
| 348 |
+
'''
|
| 349 |
+
crop_tables and plot_results_detection must have same co-ord shifts because 1 only plots the other one updates co-ordinates
|
| 350 |
+
'''
|
| 351 |
+
|
| 352 |
+
cropped_caption_list = []
|
| 353 |
+
|
| 354 |
+
for p, (xmin, ymin, xmax, ymax) in zip(prob, boxes.tolist()):
|
| 355 |
+
xmin, ymin, xmax, ymax = self.dynamic_delta(xmin, ymin, xmax, ymax, delta_xmin, delta_ymin, delta_xmax, delta_ymax, pil_img)
|
| 356 |
+
|
| 357 |
+
# Caption possibility (top)
|
| 358 |
+
offset = -200
|
| 359 |
+
c_xmin, c_ymin, c_xmax, c_ymax = self.get_cxy(pil_img, xmin, ymin, xmax, ymax, offset)
|
| 360 |
+
cropped_caption = pil_img.crop((c_xmin, c_ymin, c_xmax, c_ymax))
|
| 361 |
+
cropped_caption_list.append(cropped_caption)
|
| 362 |
+
|
| 363 |
+
# Caption possibility (bottom)
|
| 364 |
+
offset = 200
|
| 365 |
+
c_xmin, c_ymin, c_xmax, c_ymax = self.get_cxy(pil_img, xmin, ymin, xmax, ymax, offset)
|
| 366 |
+
cropped_caption = pil_img.crop((c_xmin, c_ymin, c_xmax, c_ymax))
|
| 367 |
+
cropped_caption_list.append(cropped_caption)
|
| 368 |
+
|
| 369 |
+
return cropped_caption_list
|
| 370 |
+
|
| 371 |
+
def generate_structure(self, c2, model, pil_img, prob, boxes,
|
| 372 |
+
expand_rowcol_bbox_top, expand_rowcol_bbox_bottom):
|
| 373 |
+
'''
|
| 374 |
+
Co-ordinates are adjusted here by 3 'pixels'
|
| 375 |
+
To plot table pillow image and the TSR bounding boxes on the table
|
| 376 |
+
'''
|
| 377 |
+
# st.write('img_obj')
|
| 378 |
+
# st.write(pil_img)
|
| 379 |
+
plt.figure(figsize=(32, 20))
|
| 380 |
+
plt.imshow(pil_img)
|
| 381 |
+
ax = plt.gca()
|
| 382 |
+
rows = {}
|
| 383 |
+
cols = {}
|
| 384 |
+
idx = 0
|
| 385 |
+
|
| 386 |
+
for p, (xmin, ymin, xmax, ymax) in zip(prob, boxes.tolist()):
|
| 387 |
+
|
| 388 |
+
xmin, ymin, xmax, ymax = xmin, ymin, xmax, ymax
|
| 389 |
+
cl = p.argmax()
|
| 390 |
+
class_text = model.config.id2label[cl.item()]
|
| 391 |
+
text = f'{class_text}: {p[cl]:0.2f}'
|
| 392 |
+
# or (class_text == 'table column')
|
| 393 |
+
if (class_text
|
| 394 |
+
== 'table row') or (class_text
|
| 395 |
+
== 'table projected row header') or (
|
| 396 |
+
class_text == 'table column'):
|
| 397 |
+
ax.add_patch(
|
| 398 |
+
plt.Rectangle((xmin, ymin),
|
| 399 |
+
xmax - xmin,
|
| 400 |
+
ymax - ymin,
|
| 401 |
+
fill=False,
|
| 402 |
+
color=self.colors[cl.item()],
|
| 403 |
+
linewidth=2))
|
| 404 |
+
ax.text(xmin - 10,
|
| 405 |
+
ymin - 10,
|
| 406 |
+
text,
|
| 407 |
+
fontsize=5,
|
| 408 |
+
bbox=dict(facecolor='yellow', alpha=0.5))
|
| 409 |
+
|
| 410 |
+
if class_text == 'table row':
|
| 411 |
+
rows['table row.' +
|
| 412 |
+
str(idx)] = (xmin, ymin - expand_rowcol_bbox_top, xmax,
|
| 413 |
+
ymax + expand_rowcol_bbox_bottom)
|
| 414 |
+
if class_text == 'table column':
|
| 415 |
+
cols['table column.' +
|
| 416 |
+
str(idx)] = (xmin, ymin - expand_rowcol_bbox_top, xmax,
|
| 417 |
+
ymax + expand_rowcol_bbox_bottom)
|
| 418 |
+
|
| 419 |
+
idx += 1
|
| 420 |
+
|
| 421 |
+
plt.axis('on')
|
| 422 |
+
c2.pyplot()
|
| 423 |
+
return rows, cols
|
| 424 |
+
|
| 425 |
+
def sort_table_featuresv2(self, rows: dict, cols: dict):
|
| 426 |
+
# Sometimes the header and first row overlap, and we need the header bbox not to have first row's bbox inside the headers bbox
|
| 427 |
+
rows_ = {
|
| 428 |
+
table_feature: (xmin, ymin, xmax, ymax)
|
| 429 |
+
for table_feature, (
|
| 430 |
+
xmin, ymin, xmax,
|
| 431 |
+
ymax) in sorted(rows.items(), key=lambda tup: tup[1][1])
|
| 432 |
+
}
|
| 433 |
+
cols_ = {
|
| 434 |
+
table_feature: (xmin, ymin, xmax, ymax)
|
| 435 |
+
for table_feature, (
|
| 436 |
+
xmin, ymin, xmax,
|
| 437 |
+
ymax) in sorted(cols.items(), key=lambda tup: tup[1][0])
|
| 438 |
+
}
|
| 439 |
+
|
| 440 |
+
return rows_, cols_
|
| 441 |
+
|
| 442 |
+
def individual_table_featuresv2(self, pil_img, rows: dict, cols: dict):
|
| 443 |
+
|
| 444 |
+
for k, v in rows.items():
|
| 445 |
+
xmin, ymin, xmax, ymax = v
|
| 446 |
+
cropped_img = pil_img.crop((xmin, ymin, xmax, ymax))
|
| 447 |
+
rows[k] = xmin, ymin, xmax, ymax, cropped_img
|
| 448 |
+
|
| 449 |
+
for k, v in cols.items():
|
| 450 |
+
xmin, ymin, xmax, ymax = v
|
| 451 |
+
cropped_img = pil_img.crop((xmin, ymin, xmax, ymax))
|
| 452 |
+
cols[k] = xmin, ymin, xmax, ymax, cropped_img
|
| 453 |
+
|
| 454 |
+
return rows, cols
|
| 455 |
+
|
| 456 |
+
def object_to_cellsv2(self, master_row: dict, cols: dict,
|
| 457 |
+
expand_rowcol_bbox_top, expand_rowcol_bbox_bottom,
|
| 458 |
+
padd_left):
|
| 459 |
+
'''Removes redundant bbox for rows&columns and divides each row into cells from columns
|
| 460 |
+
Args:
|
| 461 |
+
|
| 462 |
+
Returns:
|
| 463 |
+
|
| 464 |
+
|
| 465 |
+
'''
|
| 466 |
+
cells_img = {}
|
| 467 |
+
header_idx = 0
|
| 468 |
+
row_idx = 0
|
| 469 |
+
previous_xmax_col = 0
|
| 470 |
+
new_cols = {}
|
| 471 |
+
new_master_row = {}
|
| 472 |
+
previous_ymin_row = 0
|
| 473 |
+
new_cols = cols
|
| 474 |
+
new_master_row = master_row
|
| 475 |
+
## Below 2 for loops remove redundant bounding boxes ###
|
| 476 |
+
# for k_col, v_col in cols.items():
|
| 477 |
+
# xmin_col, _, xmax_col, _, col_img = v_col
|
| 478 |
+
# if (np.isclose(previous_xmax_col, xmax_col, atol=5)) or (xmin_col >= xmax_col):
|
| 479 |
+
# print('Found a column with double bbox')
|
| 480 |
+
# continue
|
| 481 |
+
# previous_xmax_col = xmax_col
|
| 482 |
+
# new_cols[k_col] = v_col
|
| 483 |
+
|
| 484 |
+
# for k_row, v_row in master_row.items():
|
| 485 |
+
# _, ymin_row, _, ymax_row, row_img = v_row
|
| 486 |
+
# if (np.isclose(previous_ymin_row, ymin_row, atol=5)) or (ymin_row >= ymax_row):
|
| 487 |
+
# print('Found a row with double bbox')
|
| 488 |
+
# continue
|
| 489 |
+
# previous_ymin_row = ymin_row
|
| 490 |
+
# new_master_row[k_row] = v_row
|
| 491 |
+
######################################################
|
| 492 |
+
for k_row, v_row in new_master_row.items():
|
| 493 |
+
|
| 494 |
+
_, _, _, _, row_img = v_row
|
| 495 |
+
xmax, ymax = row_img.size
|
| 496 |
+
xa, ya, xb, yb = 0, 0, 0, ymax
|
| 497 |
+
row_img_list = []
|
| 498 |
+
# plt.imshow(row_img)
|
| 499 |
+
# st.pyplot()
|
| 500 |
+
for idx, kv in enumerate(new_cols.items()):
|
| 501 |
+
k_col, v_col = kv
|
| 502 |
+
xmin_col, _, xmax_col, _, col_img = v_col
|
| 503 |
+
xmin_col, xmax_col = xmin_col - padd_left - 10, xmax_col - padd_left
|
| 504 |
+
xa = xmin_col
|
| 505 |
+
xb = xmax_col
|
| 506 |
+
if idx == 0:
|
| 507 |
+
xa = 0
|
| 508 |
+
if idx == len(new_cols) - 1:
|
| 509 |
+
xb = xmax
|
| 510 |
+
xa, ya, xb, yb = xa, ya, xb, yb
|
| 511 |
+
|
| 512 |
+
row_img_cropped = row_img.crop((xa, ya, xb, yb))
|
| 513 |
+
row_img_list.append(row_img_cropped)
|
| 514 |
+
|
| 515 |
+
cells_img[k_row + '.' + str(row_idx)] = row_img_list
|
| 516 |
+
row_idx += 1
|
| 517 |
+
|
| 518 |
+
return cells_img, len(new_cols), len(new_master_row) - 1
|
| 519 |
+
|
| 520 |
+
def clean_dataframe(self, df):
|
| 521 |
+
'''
|
| 522 |
+
Remove irrelevant symbols that appear with tesseractOCR
|
| 523 |
+
'''
|
| 524 |
+
# df.columns = [col.replace('|', '') for col in df.columns]
|
| 525 |
+
|
| 526 |
+
for col in df.columns:
|
| 527 |
+
|
| 528 |
+
df[col] = df[col].str.replace("'", '', regex=True)
|
| 529 |
+
df[col] = df[col].str.replace('"', '', regex=True)
|
| 530 |
+
df[col] = df[col].str.replace(']', '', regex=True)
|
| 531 |
+
df[col] = df[col].str.replace('[', '', regex=True)
|
| 532 |
+
df[col] = df[col].str.replace('{', '', regex=True)
|
| 533 |
+
df[col] = df[col].str.replace('}', '', regex=True)
|
| 534 |
+
return df
|
| 535 |
+
|
| 536 |
+
@st.cache
|
| 537 |
+
def convert_df(self, df):
|
| 538 |
+
csv = df.to_csv(index=False, encoding='utf-8-sig') # utf-8-sig to handle BOM for Excel
|
| 539 |
+
return csv.encode('utf-8')
|
| 540 |
+
|
| 541 |
+
def create_dataframe(self, c3, cell_ocr_res: list, max_cols: int,
|
| 542 |
+
max_rows: int):
|
| 543 |
+
'''Create dataframe using list of cell values of the table, also checks for valid header of dataframe
|
| 544 |
+
Args:
|
| 545 |
+
cell_ocr_res: list of strings, each element representing a cell in a table
|
| 546 |
+
max_cols, max_rows: number of columns and rows
|
| 547 |
+
Returns:
|
| 548 |
+
dataframe : final dataframe after all pre-processing
|
| 549 |
+
'''
|
| 550 |
+
|
| 551 |
+
headers = cell_ocr_res[:max_cols]
|
| 552 |
+
new_headers = uniquify(headers,
|
| 553 |
+
(f' {x!s}' for x in string.ascii_lowercase))
|
| 554 |
+
counter = 0
|
| 555 |
+
|
| 556 |
+
cells_list = cell_ocr_res[max_cols:]
|
| 557 |
+
df = pd.DataFrame("", index=range(0, max_rows), columns=new_headers)
|
| 558 |
+
|
| 559 |
+
cell_idx = 0
|
| 560 |
+
for nrows in range(max_rows):
|
| 561 |
+
for ncols in range(max_cols):
|
| 562 |
+
df.iat[nrows, ncols] = str(cells_list[cell_idx])
|
| 563 |
+
cell_idx += 1
|
| 564 |
+
|
| 565 |
+
## To check if there are duplicate headers if result of uniquify+col == col
|
| 566 |
+
## This check removes headers when all headers are empty or if median of header word count is less than 6
|
| 567 |
+
for x, col in zip(string.ascii_lowercase, new_headers):
|
| 568 |
+
if f' {x!s}' == col:
|
| 569 |
+
counter += 1
|
| 570 |
+
header_char_count = [len(col) for col in new_headers]
|
| 571 |
+
|
| 572 |
+
# if (counter == len(new_headers)) or (statistics.median(header_char_count) < 6):
|
| 573 |
+
# st.write('woooot')
|
| 574 |
+
# df.columns = uniquify(df.iloc[0], (f' {x!s}' for x in string.ascii_lowercase))
|
| 575 |
+
# df = df.iloc[1:,:]
|
| 576 |
+
|
| 577 |
+
df = self.clean_dataframe(df)
|
| 578 |
+
|
| 579 |
+
c3.dataframe(df)
|
| 580 |
+
csv = self.convert_df(df)
|
| 581 |
+
|
| 582 |
+
try:
|
| 583 |
+
numkey = str(df.iloc[0, 0])
|
| 584 |
+
except IndexError:
|
| 585 |
+
numkey = str(0)
|
| 586 |
+
|
| 587 |
+
# Create a download link with filename and extension
|
| 588 |
+
filename = f"table_{numkey}.csv" # Adjust the filename as needed
|
| 589 |
+
b64_csv = base64.b64encode(csv).decode() # Encode CSV data to base64
|
| 590 |
+
href = f'<a href="data:file/csv;base64,{b64_csv}" download="{filename}">Download {filename}</a>'
|
| 591 |
+
c3.markdown(href, unsafe_allow_html=True)
|
| 592 |
+
|
| 593 |
+
return df
|
| 594 |
+
|
| 595 |
+
async def start_process(self, image_path: str, TD_THRESHOLD, TSR_THRESHOLD,
|
| 596 |
+
OCR_THRESHOLD, padd_top, padd_left, padd_bottom,
|
| 597 |
+
padd_right, delta_xmin, delta_ymin, delta_xmax,
|
| 598 |
+
delta_ymax, expand_rowcol_bbox_top,
|
| 599 |
+
expand_rowcol_bbox_bottom):
|
| 600 |
+
'''
|
| 601 |
+
Initiates process of generating pandas dataframes from raw pdf-page images
|
| 602 |
+
|
| 603 |
+
'''
|
| 604 |
+
image = Image.open(image_path).convert("RGB")
|
| 605 |
+
probas, bboxes_scaled = table_detector(image,
|
| 606 |
+
THRESHOLD_PROBA=TD_THRESHOLD)
|
| 607 |
+
|
| 608 |
+
if bboxes_scaled.nelement() == 0:
|
| 609 |
+
st.write('No table found in the pdf-page image')
|
| 610 |
+
return ''
|
| 611 |
+
|
| 612 |
+
# try:
|
| 613 |
+
# st.write('Document: '+image_path.split('/')[-1])
|
| 614 |
+
c1, c2, c3 = st.columns((1, 1, 1))
|
| 615 |
+
|
| 616 |
+
self.plot_results_detection(c1, table_detection_model, image, probas,
|
| 617 |
+
bboxes_scaled, delta_xmin, delta_ymin,
|
| 618 |
+
delta_xmax, delta_ymax)
|
| 619 |
+
cropped_img_list = self.crop_tables(image, probas, bboxes_scaled,
|
| 620 |
+
delta_xmin, delta_ymin, delta_xmax,
|
| 621 |
+
delta_ymax)
|
| 622 |
+
|
| 623 |
+
cropped_caption_list = self.crop_caption(image, probas, bboxes_scaled,
|
| 624 |
+
delta_xmin, delta_ymin, delta_xmax,
|
| 625 |
+
delta_ymax)
|
| 626 |
+
|
| 627 |
+
# for p, (xmin, ymin, xmax, ymax) in zip(probas, bboxes_scaled.tolist()):
|
| 628 |
+
# print(p.argmax())
|
| 629 |
+
# print(xmin, ymin, xmax, ymax)
|
| 630 |
+
|
| 631 |
+
sequential_caption_img_list = []
|
| 632 |
+
for idx, caption_img in enumerate(cropped_caption_list):
|
| 633 |
+
if idx%2 == 0: # top
|
| 634 |
+
print("top")
|
| 635 |
+
else: # bottom
|
| 636 |
+
print("bottom")
|
| 637 |
+
plt.imshow(caption_img)
|
| 638 |
+
c2.pyplot()
|
| 639 |
+
sequential_caption_img_list.append(pytess(cell_pil_img=caption_img, threshold=OCR_THRESHOLD))
|
| 640 |
+
|
| 641 |
+
caption_ocr_res = await asyncio.gather(*sequential_caption_img_list)
|
| 642 |
+
flag_caption_pos = 0 # 0=top, 1=bottom
|
| 643 |
+
for idx, caption_text in enumerate(caption_ocr_res):
|
| 644 |
+
if caption_text == "" or "table" not in caption_text.lower():
|
| 645 |
+
if idx%2==0:
|
| 646 |
+
flag_caption_pos=1
|
| 647 |
+
break
|
| 648 |
+
|
| 649 |
+
for idx, caption_text in enumerate(caption_ocr_res):
|
| 650 |
+
if idx%2==flag_caption_pos:
|
| 651 |
+
c3.text(str(idx) + "_" + caption_text)
|
| 652 |
+
|
| 653 |
+
|
| 654 |
+
# for idx, unpadded_table in enumerate(cropped_img_list):
|
| 655 |
+
|
| 656 |
+
# table = self.add_padding(unpadded_table, padd_top, padd_right,
|
| 657 |
+
# padd_bottom, padd_left)
|
| 658 |
+
# # table = super_res(table)
|
| 659 |
+
# # table = binarizeBlur_image(table)
|
| 660 |
+
# # table = sharpen_image(table) # Test sharpen image next
|
| 661 |
+
# # table = td_postprocess(table)
|
| 662 |
+
|
| 663 |
+
# # table.save("result"+str(idx)+".png")
|
| 664 |
+
|
| 665 |
+
# probas, bboxes_scaled = table_struct_recog(
|
| 666 |
+
# table, THRESHOLD_PROBA=TSR_THRESHOLD)
|
| 667 |
+
# rows, cols = self.generate_structure(c2, table_recognition_model,
|
| 668 |
+
# table, probas, bboxes_scaled,
|
| 669 |
+
# expand_rowcol_bbox_top,
|
| 670 |
+
# expand_rowcol_bbox_bottom)
|
| 671 |
+
# # st.write(len(rows), len(cols))
|
| 672 |
+
# rows, cols = self.sort_table_featuresv2(rows, cols)
|
| 673 |
+
# master_row, cols = self.individual_table_featuresv2(
|
| 674 |
+
# table, rows, cols)
|
| 675 |
+
|
| 676 |
+
# cells_img, max_cols, max_rows = self.object_to_cellsv2(
|
| 677 |
+
# master_row, cols, expand_rowcol_bbox_top,
|
| 678 |
+
# expand_rowcol_bbox_bottom, padd_left)
|
| 679 |
+
|
| 680 |
+
# sequential_cell_img_list = []
|
| 681 |
+
# for k, img_list in cells_img.items():
|
| 682 |
+
# for img in img_list:
|
| 683 |
+
# # img = super_res(img)
|
| 684 |
+
# # img = sharpen_image(img) # Test sharpen image next
|
| 685 |
+
# # img = binarizeBlur_image(img)
|
| 686 |
+
# # img = self.add_padding(img, 10,10,10,10)
|
| 687 |
+
# # plt.imshow(img)
|
| 688 |
+
# # c3.pyplot()
|
| 689 |
+
# sequential_cell_img_list.append(
|
| 690 |
+
# pytess(cell_pil_img=img, threshold=OCR_THRESHOLD))
|
| 691 |
+
|
| 692 |
+
# cell_ocr_res = await asyncio.gather(*sequential_cell_img_list)
|
| 693 |
+
|
| 694 |
+
# self.create_dataframe(c3, cell_ocr_res, max_cols, max_rows)
|
| 695 |
+
# st.write(
|
| 696 |
+
# 'Errors in OCR is due to either quality of the image or performance of the OCR'
|
| 697 |
+
# )
|
| 698 |
+
# except:
|
| 699 |
+
# st.write('Either incorrectly identified table or no table, to debug remove try/except')
|
| 700 |
+
# break
|
| 701 |
+
# break
|
| 702 |
+
|
| 703 |
+
|
| 704 |
+
if __name__ == "__main__":
|
| 705 |
+
|
| 706 |
+
st_up, st_lang = st.columns((1, 1))
|
| 707 |
+
img_name = st_up.file_uploader("Upload an image with table(s)")
|
| 708 |
+
lang = st_lang.selectbox('Language', ('en', 'japan'))
|
| 709 |
+
reload_ocr(lang)
|
| 710 |
+
|
| 711 |
+
st1, st2, st3 = st.columns((1, 1, 1))
|
| 712 |
+
TD_th = st1.slider('Table detection threshold', 0.0, 1.0, 0.8)
|
| 713 |
+
TSR_th = st2.slider('Table structure recognition threshold', 0.0, 1.0, 0.7)
|
| 714 |
+
OCR_th = st3.slider("Text Probs Threshold", 0.0, 1.0, 0.5)
|
| 715 |
+
|
| 716 |
+
st1, st2, st3, st4 = st.columns((1, 1, 1, 1))
|
| 717 |
+
|
| 718 |
+
padd_top = st1.slider('Padding top', 0, 200, 90)
|
| 719 |
+
padd_left = st2.slider('Padding left', 0, 200, 40)
|
| 720 |
+
padd_right = st3.slider('Padding right', 0, 200, 40)
|
| 721 |
+
padd_bottom = st4.slider('Padding bottom', 0, 200, 90)
|
| 722 |
+
|
| 723 |
+
te = TableExtractionPipeline()
|
| 724 |
+
# for img in image_list:
|
| 725 |
+
if img_name is not None:
|
| 726 |
+
asyncio.run(
|
| 727 |
+
te.start_process(img_name,
|
| 728 |
+
TD_THRESHOLD=TD_th,
|
| 729 |
+
TSR_THRESHOLD=TSR_th,
|
| 730 |
+
OCR_THRESHOLD=OCR_th,
|
| 731 |
+
padd_top=padd_top,
|
| 732 |
+
padd_left=padd_left,
|
| 733 |
+
padd_bottom=padd_bottom,
|
| 734 |
+
padd_right=padd_right,
|
| 735 |
+
delta_xmin=10, # add offset to the left of the table
|
| 736 |
+
delta_ymin=3, # add offset to the bottom of the table
|
| 737 |
+
delta_xmax=10, # add offset to the right of the table
|
| 738 |
+
delta_ymax=3, # add offset to the top of the table
|
| 739 |
+
expand_rowcol_bbox_top=0,
|
| 740 |
+
expand_rowcol_bbox_bottom=0))
|
components/callbacks.py
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Define callbacks here
|
| 2 |
+
from pytorch_lightning.callbacks import EarlyStopping
|
| 3 |
+
|
| 4 |
+
early_stopping = EarlyStopping(monitor="loss", min_delta=0, patience=3)
|
components/data_module.py
ADDED
|
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Callable, List, Optional, Union
|
| 2 |
+
|
| 3 |
+
import torch
|
| 4 |
+
from pytorch_lightning import LightningDataModule
|
| 5 |
+
from torch.utils.data import DataLoader, Dataset
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
class SampleDataset(Dataset):
|
| 9 |
+
|
| 10 |
+
def __init__(self,
|
| 11 |
+
x: Union[List, torch.Tensor],
|
| 12 |
+
y: Union[List, torch.Tensor],
|
| 13 |
+
transforms: Optional[Callable] = None) -> None:
|
| 14 |
+
super(SampleDataset, self).__init__()
|
| 15 |
+
self.x = x
|
| 16 |
+
self.y = y
|
| 17 |
+
|
| 18 |
+
if transforms is None:
|
| 19 |
+
# Replace None with some default transforms
|
| 20 |
+
# If image, could be an Resize and ToTensor
|
| 21 |
+
self.transforms = lambda x: x
|
| 22 |
+
else:
|
| 23 |
+
self.transforms = transforms
|
| 24 |
+
|
| 25 |
+
def __len__(self):
|
| 26 |
+
return len(self.x)
|
| 27 |
+
|
| 28 |
+
def __getitem__(self, index: int):
|
| 29 |
+
x = self.x[index]
|
| 30 |
+
y = self.y[index]
|
| 31 |
+
|
| 32 |
+
x = self.transforms(x)
|
| 33 |
+
return x, y
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
class SampleDataModule(LightningDataModule):
|
| 37 |
+
|
| 38 |
+
def __init__(self,
|
| 39 |
+
x: Union[List, torch.Tensor],
|
| 40 |
+
y: Union[List, torch.Tensor],
|
| 41 |
+
transforms: Optional[Callable] = None,
|
| 42 |
+
val_ratio: float = 0,
|
| 43 |
+
batch_size: int = 32) -> None:
|
| 44 |
+
super(SampleDataModule, self).__init__()
|
| 45 |
+
assert 0 <= val_ratio < 1
|
| 46 |
+
assert isinstance(batch_size, int)
|
| 47 |
+
self.x = x
|
| 48 |
+
self.y = y
|
| 49 |
+
|
| 50 |
+
self.transforms = transforms
|
| 51 |
+
self.val_ratio = val_ratio
|
| 52 |
+
self.batch_size = batch_size
|
| 53 |
+
|
| 54 |
+
self.setup()
|
| 55 |
+
self.prepare_data()
|
| 56 |
+
|
| 57 |
+
def setup(self, stage: Optional[str] = None) -> None:
|
| 58 |
+
pass
|
| 59 |
+
|
| 60 |
+
def prepare_data(self) -> None:
|
| 61 |
+
n_samples: int = len(self.x)
|
| 62 |
+
train_size: int = n_samples - int(n_samples * self.val_ratio)
|
| 63 |
+
|
| 64 |
+
self.train_dataset = SampleDataset(x=self.x[:train_size],
|
| 65 |
+
y=self.y[:train_size],
|
| 66 |
+
transforms=self.transforms)
|
| 67 |
+
if train_size < n_samples:
|
| 68 |
+
self.val_dataset = SampleDataset(x=self.x[train_size:],
|
| 69 |
+
y=self.y[train_size:],
|
| 70 |
+
transforms=self.transforms)
|
| 71 |
+
else:
|
| 72 |
+
self.val_dataset = SampleDataset(x=self.x[-self.batch_size:],
|
| 73 |
+
y=self.y[-self.batch_size:],
|
| 74 |
+
transforms=self.transforms)
|
| 75 |
+
|
| 76 |
+
def train_dataloader(self) -> DataLoader:
|
| 77 |
+
return DataLoader(dataset=self.train_dataset,
|
| 78 |
+
batch_size=self.batch_size)
|
| 79 |
+
|
| 80 |
+
def val_dataloader(self) -> DataLoader:
|
| 81 |
+
return DataLoader(dataset=self.val_dataset, batch_size=self.batch_size)
|
config.py
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from dynaconf import Dynaconf
|
| 2 |
+
|
| 3 |
+
CFG = Dynaconf(envvar_prefix="DYNACONF", settings_files=["config/config.yaml"])
|
config/config.yaml
ADDED
|
File without changes
|
data/.gitkeep
ADDED
|
File without changes
|
docker-compose.yml
ADDED
|
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version: "3.7"
|
| 2 |
+
|
| 3 |
+
services:
|
| 4 |
+
model_name:
|
| 5 |
+
build:
|
| 6 |
+
context: .
|
| 7 |
+
dockerfile: .docker/Dockerfile
|
| 8 |
+
container_name: model_name
|
| 9 |
+
ports:
|
| 10 |
+
- "8996:8996"
|
| 11 |
+
env_file:
|
| 12 |
+
- ./.env
|
| 13 |
+
volumes:
|
| 14 |
+
- ./data:/home/working/data:ro
|
| 15 |
+
|
| 16 |
+
# This part is used to enable GPU support
|
| 17 |
+
deploy:
|
| 18 |
+
resources:
|
| 19 |
+
reservations:
|
| 20 |
+
devices:
|
| 21 |
+
- driver: nvidia
|
| 22 |
+
count: 1
|
| 23 |
+
capabilities: [ gpu ]
|
models/__init__.py
ADDED
|
File without changes
|
models/base_model/classification.py
ADDED
|
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from abc import abstractmethod
|
| 2 |
+
from typing import Any, Dict, List
|
| 3 |
+
|
| 4 |
+
import torch
|
| 5 |
+
from pytorch_lightning import LightningModule
|
| 6 |
+
from torch import Tensor
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
class LightningClassification(LightningModule):
|
| 10 |
+
|
| 11 |
+
@abstractmethod
|
| 12 |
+
def __init__(self, *args, **kwargs) -> None:
|
| 13 |
+
super(LightningClassification, self).__init__(*args, **kwargs)
|
| 14 |
+
self.train_batch_output: List[Dict] = []
|
| 15 |
+
self.validation_batch_output: List[Dict] = []
|
| 16 |
+
self.log_value_list: List[str] = ['loss', 'f1', 'precision', 'recall']
|
| 17 |
+
|
| 18 |
+
@abstractmethod
|
| 19 |
+
def forward(self, *args, **kwargs) -> Any:
|
| 20 |
+
pass
|
| 21 |
+
|
| 22 |
+
@abstractmethod
|
| 23 |
+
def configure_optimizers(self):
|
| 24 |
+
pass
|
| 25 |
+
|
| 26 |
+
@abstractmethod
|
| 27 |
+
def loss(self, input: Tensor, target: Tensor, **kwargs) -> Tensor:
|
| 28 |
+
pass
|
| 29 |
+
|
| 30 |
+
@abstractmethod
|
| 31 |
+
def training_step(self, batch, batch_idx):
|
| 32 |
+
pass
|
| 33 |
+
|
| 34 |
+
def __average(self, key: str, outputs: List[Dict]) -> Tensor:
|
| 35 |
+
target_arr = torch.Tensor([val[key] for val in outputs]).float()
|
| 36 |
+
return target_arr.mean()
|
| 37 |
+
|
| 38 |
+
@torch.no_grad()
|
| 39 |
+
def on_train_epoch_start(self) -> None:
|
| 40 |
+
self.train_batch_output = []
|
| 41 |
+
|
| 42 |
+
@torch.no_grad()
|
| 43 |
+
def on_train_epoch_end(self) -> None:
|
| 44 |
+
for key in self.log_value_list:
|
| 45 |
+
val = self.__average(key=key, outputs=self.train_batch_output)
|
| 46 |
+
log_name = f"training/{key}"
|
| 47 |
+
self.log(name=log_name, value=val)
|
| 48 |
+
|
| 49 |
+
@abstractmethod
|
| 50 |
+
@torch.no_grad()
|
| 51 |
+
def validation_step(self, batch, batch_idx):
|
| 52 |
+
pass
|
| 53 |
+
|
| 54 |
+
@torch.no_grad()
|
| 55 |
+
def on_validation_epoch_start(self) -> None:
|
| 56 |
+
self.validation_batch_output = []
|
| 57 |
+
|
| 58 |
+
@torch.no_grad()
|
| 59 |
+
def on_validation_epoch_end(self) -> None:
|
| 60 |
+
for key in self.log_value_list:
|
| 61 |
+
val = self.__average(key=key, outputs=self.validation_batch_output)
|
| 62 |
+
log_name = f"val/{key}"
|
| 63 |
+
self.log(name=log_name, value=val)
|
models/base_model/gan.py
ADDED
|
File without changes
|
models/base_model/regression.py
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from abc import abstractmethod
|
| 2 |
+
from typing import Any, Dict, List
|
| 3 |
+
|
| 4 |
+
import torch
|
| 5 |
+
from pytorch_lightning import LightningModule
|
| 6 |
+
from torch import Tensor
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
class LightningRegression(LightningModule):
|
| 10 |
+
|
| 11 |
+
@abstractmethod
|
| 12 |
+
def __init__(self, *args, **kwargs) -> None:
|
| 13 |
+
super(LightningRegression, self).__init__(*args, **kwargs)
|
| 14 |
+
self.train_step_output: List[Dict] = []
|
| 15 |
+
self.validation_step_output: List[Dict] = []
|
| 16 |
+
self.log_value_list: List[str] = ['loss', 'mse', 'mape']
|
| 17 |
+
|
| 18 |
+
@abstractmethod
|
| 19 |
+
def forward(self, *args, **kwargs) -> Any:
|
| 20 |
+
pass
|
| 21 |
+
|
| 22 |
+
@abstractmethod
|
| 23 |
+
def configure_optimizers(self):
|
| 24 |
+
pass
|
| 25 |
+
|
| 26 |
+
@abstractmethod
|
| 27 |
+
def loss(self, input: Tensor, output: Tensor, **kwargs):
|
| 28 |
+
return 0
|
| 29 |
+
|
| 30 |
+
@abstractmethod
|
| 31 |
+
def training_step(self, batch, batch_idx):
|
| 32 |
+
pass
|
| 33 |
+
|
| 34 |
+
def __average(self, key: str, outputs: List[Dict]) -> Tensor:
|
| 35 |
+
target_arr = torch.Tensor([val[key] for val in outputs]).float()
|
| 36 |
+
return target_arr.mean()
|
| 37 |
+
|
| 38 |
+
@torch.no_grad()
|
| 39 |
+
def on_train_epoch_end(self) -> None:
|
| 40 |
+
for key in self.log_value_list:
|
| 41 |
+
val = self.__average(key=key, outputs=self.train_step_output)
|
| 42 |
+
log_name = f"training/{key}"
|
| 43 |
+
self.log(name=log_name, value=val)
|
| 44 |
+
|
| 45 |
+
@torch.no_grad()
|
| 46 |
+
@abstractmethod
|
| 47 |
+
def validation_step(self, batch, batch_idx):
|
| 48 |
+
pass
|
| 49 |
+
|
| 50 |
+
@torch.no_grad()
|
| 51 |
+
def validation_epoch_end(self, outputs):
|
| 52 |
+
for key in self.log_value_list:
|
| 53 |
+
val = self.__average(key=key, outputs=self.validation_step_output)
|
| 54 |
+
log_name = f"training/{key}"
|
| 55 |
+
self.log(name=log_name, value=val)
|
models/metrics/classification.py
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Dict
|
| 2 |
+
|
| 3 |
+
import torch
|
| 4 |
+
from torchmetrics import functional as FM
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
def classification_metrics(
|
| 8 |
+
preds: torch.Tensor,
|
| 9 |
+
target: torch.Tensor,
|
| 10 |
+
num_classes: int,
|
| 11 |
+
average: str = 'macro',
|
| 12 |
+
task: str = 'multiclass') -> Dict[str, torch.Tensor]:
|
| 13 |
+
"""
|
| 14 |
+
get_classification_metrics
|
| 15 |
+
Return some metrics evaluation the classification task
|
| 16 |
+
|
| 17 |
+
Parameters
|
| 18 |
+
----------
|
| 19 |
+
preds : torch.Tensor
|
| 20 |
+
logits, probs
|
| 21 |
+
target : torch.Tensor
|
| 22 |
+
targets label
|
| 23 |
+
|
| 24 |
+
Returns
|
| 25 |
+
-------
|
| 26 |
+
Dict[str, torch.Tensor]
|
| 27 |
+
_description_
|
| 28 |
+
"""
|
| 29 |
+
f1 = FM.f1_score(preds=preds,
|
| 30 |
+
target=target,
|
| 31 |
+
num_classes=num_classes,
|
| 32 |
+
task=task,
|
| 33 |
+
average=average)
|
| 34 |
+
recall = FM.recall(preds=preds,
|
| 35 |
+
target=target,
|
| 36 |
+
num_classes=num_classes,
|
| 37 |
+
task=task,
|
| 38 |
+
average=average)
|
| 39 |
+
precision = FM.precision(preds=preds,
|
| 40 |
+
target=target,
|
| 41 |
+
num_classes=num_classes,
|
| 42 |
+
task=task,
|
| 43 |
+
average=average)
|
| 44 |
+
return dict(f1=f1, precision=precision, recall=recall)
|
models/metrics/regression.py
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Dict
|
| 2 |
+
|
| 3 |
+
import torch
|
| 4 |
+
from torchmetrics import functional as FM
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
def regression_metrics(preds: torch.Tensor,
|
| 8 |
+
target: torch.Tensor) -> Dict[str, torch.Tensor]:
|
| 9 |
+
"""
|
| 10 |
+
get_classification_metrics
|
| 11 |
+
Return some metrics evaluation the classification task
|
| 12 |
+
|
| 13 |
+
Parameters
|
| 14 |
+
----------
|
| 15 |
+
preds : torch.Tensor
|
| 16 |
+
logits, probs
|
| 17 |
+
target : torch.Tensor
|
| 18 |
+
targets label
|
| 19 |
+
|
| 20 |
+
Returns
|
| 21 |
+
-------
|
| 22 |
+
Dict[str, torch.Tensor]
|
| 23 |
+
_description_
|
| 24 |
+
"""
|
| 25 |
+
mse: torch.Tensor = FM.mean_squared_error(preds=preds, target=target)
|
| 26 |
+
mape: torch.Tensor = FM.mean_absolute_percentage_error(preds=preds,
|
| 27 |
+
target=target)
|
| 28 |
+
return dict(mse=mse, mape=mape)
|
models/model_lit.py
ADDED
|
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from torch import Tensor, nn, optim
|
| 2 |
+
from torch.nn import functional as F
|
| 3 |
+
|
| 4 |
+
from .base_model.classification import LightningClassification
|
| 5 |
+
from .metrics.classification import classification_metrics
|
| 6 |
+
from .modules.sample_torch_module import UselessLayer
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
class UselessClassification(LightningClassification):
|
| 10 |
+
|
| 11 |
+
def __init__(self, n_classes: int, lr: float, **kwargs) -> None:
|
| 12 |
+
super(UselessClassification).__init__()
|
| 13 |
+
self.save_hyperparameters()
|
| 14 |
+
self.n_classes = n_classes
|
| 15 |
+
self.lr = lr
|
| 16 |
+
self.main = nn.Sequential(UselessLayer(), nn.GELU())
|
| 17 |
+
|
| 18 |
+
def forward(self, x: Tensor) -> Tensor:
|
| 19 |
+
return self.main(x)
|
| 20 |
+
|
| 21 |
+
def loss(self, input: Tensor, target: Tensor) -> Tensor:
|
| 22 |
+
return F.mse_loss(input=input, target=target)
|
| 23 |
+
|
| 24 |
+
def configure_optimizers(self):
|
| 25 |
+
optimizer = optim.Adam(params=self.parameters(), lr=self.lr)
|
| 26 |
+
return optimizer
|
| 27 |
+
|
| 28 |
+
def training_step(self, batch, batch_idx):
|
| 29 |
+
x, y = batch
|
| 30 |
+
|
| 31 |
+
logits = self.forward(x)
|
| 32 |
+
loss = self.loss(input=x, target=y)
|
| 33 |
+
metrics = classification_metrics(preds=logits,
|
| 34 |
+
target=y,
|
| 35 |
+
num_classes=self.n_classes)
|
| 36 |
+
|
| 37 |
+
self.train_batch_output.append({'loss': loss, **metrics})
|
| 38 |
+
return loss
|
| 39 |
+
|
| 40 |
+
def validation_step(self, batch, batch_idx):
|
| 41 |
+
x, y = batch
|
| 42 |
+
|
| 43 |
+
logits = self.forward(x)
|
| 44 |
+
loss = self.loss(input=x, target=y)
|
| 45 |
+
metrics = classification_metrics(preds=logits,
|
| 46 |
+
target=y,
|
| 47 |
+
num_classes=self.n_classes)
|
| 48 |
+
|
| 49 |
+
self.validation_batch_output.append({'loss': loss, **metrics})
|
| 50 |
+
return loss
|
models/modules/sample_torch_module.py
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from torch import Tensor, nn
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
class UselessLayer(nn.Module):
|
| 5 |
+
|
| 6 |
+
def __init__(self) -> None:
|
| 7 |
+
super(UselessLayer, self).__init__()
|
| 8 |
+
self.seq = nn.Identity()
|
| 9 |
+
|
| 10 |
+
def forward(self, x: Tensor) -> Tensor:
|
| 11 |
+
x = self.seq(x)
|
| 12 |
+
return x
|
requirements.txt
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
timm==0.9.2
|
| 2 |
+
torch --index-url https://download.pytorch.org/whl/cpu
|
| 3 |
+
torchvision --index-url https://download.pytorch.org/whl/cpu
|
| 4 |
+
torchaudio --index-url https://download.pytorch.org/whl/cpu
|
| 5 |
+
streamlit==1.21.0
|
| 6 |
+
pandas
|
| 7 |
+
transformers==4.29.1
|
| 8 |
+
Pillow==10.0.1
|
| 9 |
+
paddlepaddle
|
| 10 |
+
paddleocr
|
test_pdf2img.py
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
from pdf2image import convert_from_path
|
| 3 |
+
|
| 4 |
+
# Set the PDF file path
|
| 5 |
+
pdf_path = 'test.pdf'
|
| 6 |
+
|
| 7 |
+
# Convert the first page of the PDF to a JPEG image
|
| 8 |
+
first = 14
|
| 9 |
+
last = 14
|
| 10 |
+
images = convert_from_path(pdf_path, dpi=300, first_page=first, last_page=last, poppler_path=r"C:\poppler-23.07.0\Library\bin")
|
| 11 |
+
|
| 12 |
+
# Save the image file
|
| 13 |
+
image_path = os.path.splitext(pdf_path)[0]
|
| 14 |
+
|
| 15 |
+
for index, image in enumerate(images):
|
| 16 |
+
image.save(image_path + "p" + str(index+first) + '.jpg', 'JPEG')
|
tests/test_resource.py
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
def test_cuda():
|
| 2 |
+
from torch.cuda import is_available
|
| 3 |
+
assert is_available()
|
| 4 |
+
|
utils/.gitkeep
ADDED
|
File without changes
|