Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Update demo_validator and make a better starting point for tutorial #4

Merged
merged 1 commit into from
Dec 12, 2023
Merged

Update demo_validator and make a better starting point for tutorial #4

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 3 additions & 16 deletions ocr_subnet/validator/reward.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,22 +40,9 @@ def loss(label: dict, pred: dict, alpha_p=1.0, alpha_f=1.0, alpha_t=1.0):
Returns:
- float: The score for the section. Bounded between 0 and 1.
"""
# position loss is IOU of the bounding boxes
if pred.get('position'):
position_loss = get_iou(label['position'], pred['position'])
else:
# otherwise set to max loss
position_loss = 1.0

if pred.get('font'):
font_loss = get_font_distance(label['font'], pred['font'])
else:
font_loss = 1.0

if pred.get('text'):
text_loss = get_edit_distance(label['text'], pred['text'])
else:
text_loss = 1.0
position_loss = get_iou(label['position'], pred.get('position'))
font_loss = get_font_distance(label['font'], pred.get('font'))
text_loss = get_edit_distance(label['text'], pred.get('text'))

total_loss = (alpha_p * position_loss + alpha_f * font_loss + alpha_t * text_loss) / (alpha_p + alpha_f + alpha_t)

Expand Down
15 changes: 12 additions & 3 deletions ocr_subnet/validator/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
import editdistance


def get_iou(bb1, bb2):
def get_iou(bb1, bb2 = None):
"""
Calculate the Intersection over Union (IoU) of two bounding boxes.
NOTE: Thanks to this guy! https://stackoverflow.com/questions/25349178/calculating-percentage-of-bounding-box-overlap-for-image-detector-evaluation
Expand All @@ -21,6 +21,10 @@ def get_iou(bb1, bb2):
-------
float: Normalized between 0 and 1.
"""

if not bb2:
return 1.0

assert bb1['x1'] < bb1['x2']
assert bb1['y1'] < bb1['y2']
assert bb2['x1'] < bb2['x2']
Expand Down Expand Up @@ -52,7 +56,7 @@ def get_iou(bb1, bb2):
return iou


def get_edit_distance(text1: str, text2: str):
def get_edit_distance(text1: str, text2: str = None):
"""Calculate the edit distance between two strings.

Parameters
Expand All @@ -67,10 +71,12 @@ def get_edit_distance(text1: str, text2: str):
float
The edit distance between the two strings, normalized to be between 0 and 1.
"""
if not text2:
return 1.0

return editdistance.eval(text1, text2) / max(len(text1), len(text2))

def get_font_distance(font1: dict, font2: dict):
def get_font_distance(font1: dict, font2: dict = None):
"""Calculate the distance between two fonts.

Parameters
Expand All @@ -85,6 +91,9 @@ def get_font_distance(font1: dict, font2: dict):
float
The distance between the two fonts. Normalized to be between 0 and 1.
"""
if not font2:
return 1.0

font_size_loss = abs(font1['size'] - font2['size']) / max(font1['size'], font2['size'])
font_family_loss = 0.0 if font1['family'] == font2['family'] else 1.0
return (font_size_loss + font_family_loss) / 2
165 changes: 155 additions & 10 deletions scripts/demo_validator.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,16 +23,22 @@
"outputs": [],
"source": [
"import os\n",
"import datetime\n",
"import pdf2image\n",
"import math\n",
"import torch\n",
"import random\n",
"import datetime\n",
"import bittensor as bt\n",
"from IPython.display import display\n",
"\n",
"import pdf2image\n",
"import editdistance\n",
"\n",
"from faker import Faker\n",
"from typing import List\n",
"from PIL import Image, ImageFilter, ImageDraw\n",
"from reportlab.lib.pagesizes import letter\n",
"from reportlab.pdfgen import canvas\n",
"from reportlab.lib.units import inch\n",
"from faker import Faker\n"
"from reportlab.lib.units import inch\n"
]
},
{
Expand Down Expand Up @@ -173,7 +179,7 @@
"\n",
"# Use the function and pass the data and the filename you want to save as\n",
"data = create_invoice(invoice_info, \"sample_invoice.pdf\")\n",
"data"
"data[:3]"
]
},
{
Expand All @@ -184,7 +190,7 @@
"source": [
"\n",
"\n",
"def corrupt_image(input_pdf_path, output_pdf_path, theta=1, border=50, noise=0.1, scale=0.95, blur=1, spot=(100,100)):\n",
"def corrupt_image(input_pdf_path, output_pdf_path, theta=0.2, border=50, noise=0.1, scale=0.95, blur=1, spot=(100,100)):\n",
" # Convert PDF to images\n",
" images = pdf2image.convert_from_path(input_pdf_path)\n",
"\n",
Expand Down Expand Up @@ -362,7 +368,18 @@
" return image\n",
"\n",
"# scale data so that it matches the image size\n",
"def scale_data(data, w, h):\n",
"def scale_data(data: List[dict], w: int, h: int):\n",
" \"\"\"\n",
" Rescales the position data so that it matches the image size\n",
"\n",
" Args:\n",
" data (List[dict]): List of dictionaries containing the position, font and text of each section\n",
" w (int): Width of the image\n",
" h (int): Height of the image\n",
"\n",
" Returns:\n",
" List[dict]: List of dictionaries containing the position, font and text of each section\n",
" \"\"\"\n",
" scaled_data = []\n",
" for section in data:\n",
" entry = section.copy()\n",
Expand All @@ -379,13 +396,12 @@
"outputs": [],
"source": [
"# Path to the image you want to extract text from\n",
"path = os.path.join(os.getcwd(), 'sample_invoice.pdf')\n",
"path = os.path.join(os.getcwd(), 'noisy_invoice.pdf')\n",
"\n",
"# Read the pdf into memory\n",
"image = pdf2image.convert_from_path(path)[0]\n",
"# convert to PIL image\n",
"image = image.convert('RGB')\n",
"# image\n",
"\n",
"\n",
"scaled_data = scale_data(data, *image.size)\n",
Expand All @@ -397,7 +413,136 @@
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
"source": [
"\n",
"def get_position_error(boxA: List[float], boxB: List[float] = None):\n",
" \"\"\"\n",
" Calculate the intersection over union (IoU) of two bounding boxes.\n",
"\n",
" Args:\n",
" - boxA (list): Bounding box coordinates of box A in the format [x1, y1, x2, y2].\n",
" - boxB (list): Bounding box coordinates of box B in the format [x1, y1, x2, y2].\n",
"\n",
" Returns:\n",
" - float: The IoU value, ranging from 0 to 1.\n",
" \"\"\"\n",
" if not boxB:\n",
" return 1.0\n",
"\n",
" xA = max(boxA[0], boxB[0])\n",
" yA = max(boxA[1], boxB[1])\n",
" xB = min(boxA[2], boxB[2])\n",
" yB = min(boxA[3], boxB[3])\n",
"\n",
" intersection_area = max(0, xB - xA + 1) * max(0, yB - yA + 1)\n",
"\n",
" boxA_area = (boxA[2] - boxA[0] + 1) * (boxA[3] - boxA[1] + 1)\n",
" boxB_area = (boxB[2] - boxB[0] + 1) * (boxB[3] - boxB[1] + 1)\n",
"\n",
" iou = intersection_area / float(boxA_area + boxB_area - intersection_area)\n",
" \n",
" return 1 - iou\n",
"\n",
"def get_edit_distance(text1: str, text2: str = None):\n",
" \"\"\"\n",
" Calculate the edit distance between two strings.\n",
"\n",
" Args:\n",
" - text1 (str): The first string.\n",
" - text2 (str): The second string.\n",
"\n",
" Returns:\n",
" - float: The edit distance between the two strings. Normalized to be between 0 and 1.\n",
" \"\"\"\n",
" if not text2:\n",
" return 1.0\n",
"\n",
" return editdistance.eval(text1, text2) / max(len(text1), len(text2))\n",
"\n",
"def get_font_distance(font1: dict, font2: dict = None):\n",
" \"\"\"\n",
" Calculate the distance between two fonts, based on the font size and font family.\n",
"\n",
" Args:\n",
" - font1 (dict): The first font.\n",
" - font2 (dict): The second font.\n",
"\n",
" Returns:\n",
" - float: The distance between the two fonts. Normalized to be between 0 and 1.\n",
" \"\"\"\n",
" if not font2:\n",
" return 1.0\n",
"\n",
" font_size_loss = abs(font1['size'] - font2['size']) / max(font1['size'], font2['size'])\n",
" font_family_loss = 0.0 if font1['family'] == font2['family'] else 1.0\n",
" return (font_size_loss + font_family_loss) / 2\n",
"\n",
"def loss(label: dict, pred: dict, alpha_p=1.0, alpha_f=1.0, alpha_t=1.0):\n",
" \"\"\"\n",
" Score a section of the image based on the section's correctness.\n",
" Correctness is defined as:\n",
" - the intersection over union of the bounding boxes,\n",
" - the delta between the predicted font and the ground truth font,\n",
" - and the edit distance between the predicted text and the ground truth text.\n",
"\n",
" Args:\n",
" - label (dict): The ground truth data for the section.\n",
" - pred (dict): The predicted data for the section.\n",
"\n",
" Returns:\n",
" - float: The score for the section. Bounded between 0 and 1.\n",
" \"\"\"\n",
" position_loss = get_position_error(label['position'], pred.get('position'))\n",
" font_loss = get_font_distance(label['font'], pred.get('font'))\n",
" text_loss = get_edit_distance(label['text'], pred.get('text'))\n",
"\n",
" total_loss = (alpha_p * position_loss + alpha_f * font_loss + alpha_t * text_loss) / (alpha_p + alpha_f + alpha_t)\n",
"\n",
" bt.logging.info(f\"position_loss: {position_loss:.3f}, font_loss: {font_loss:.3f}, text_loss: {text_loss:.3f}, total_loss: {total_loss:.3f}\")\n",
"\n",
" return total_loss\n",
"\n",
"\n",
"def reward(image_data: List[dict], predictions: List[dict]) -> float:\n",
" \"\"\"\n",
" Reward the miner response to the OCR request.\n",
"\n",
" Args:\n",
" - image_data (list): The ground truth data for the image.\n",
" - predictions (list): The predicted data for the image.\n",
"\n",
" Returns:\n",
" - float: The reward for the miner response. Bounded between 0 and 1.\n",
" \"\"\"\n",
"\n",
" if predictions is None:\n",
" return 0.0\n",
"\n",
" # Take mean score over all sections in document\n",
" predictions_loss = torch.mean(torch.FloatTensor([loss(label, pred) for label, pred in zip(image_data, predictions)]))\n",
"\n",
" # convert loss to reward (invert and scale)\n",
" raw_reward = 1.0 / (predictions_loss + 0.1)\n",
" # NOTE: Tanh will saturate quickly and so two losses of 0.1 and 0.01 would produce raw_rewards of 10 and 100 which would both have tanh values of effectively 1.0.\n",
" normalized_reward = torch.tanh(raw_reward)\n",
"\n",
" bt.logging.info(f\"predictions_loss: {predictions_loss:.3f}, raw_reward: {raw_reward:.3f}, normalized_reward: {normalized_reward:.3f}\")\n",
" return normalized_reward"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# Test the reward function with slightly modified image data\n",
"w, h = image.size\n",
"scaled_data2 = scale_data(data, w+100, h+100)\n",
"\n",
"\n",
"reward(scaled_data, scaled_data2)"
]
},
{
"cell_type": "code",
Expand Down
Binary file modified scripts/noisy_invoice.pdf
View file
Open in desktop
Binary file not shown.
Binary file modified scripts/sample_invoice.pdf
View file
Open in desktop
Binary file not shown.