Skip to content

DigiNsure Inc. is an innovative insurance company focused on enhancing the efficiency of processing claims and customer service interactions. Their newest initiative is digitizing all historical insurance claim documents, which includes improving the labeling of some IDs scanned from paper documents and identifying them as primary or secondary IDs.

To help them in their effort, you'll be using multi-modal learning to train an Optical Character Recognition (OCR) model. To improve the classification, the model will use images of the scanned documents as input and their insurance type (home, life, auto, health, or other). Integrating different data modalities (such as image and text) enables the model to perform better in complex scenarios, helping to capture more nuanced information. The labels that the model will be trained to identify are of two types: a primary and a secondary ID, for each image-insurance type pair.

! pip install torchvision
Hidden output 
# Importing the necessary libraries
import matplotlib.pyplot as plt
import numpy as np
from project_utils import ProjectDataset
import pickle 
import torch
import torch.nn as nn
from torch.utils.data import DataLoader

# Loading the data using pickle
dataset = pickle.load(open('ocr_insurance_dataset.pkl', 'rb'))

#Defining the function 'show_dataset_images' to visualize codes with their corresponding types and labels 
def show_dataset_images(dataset, num_images=5):
    fig, axes = plt.subplots(1, min(num_images, len(dataset)), figsize=(20, 4))
    for ax, idx in zip(axes, np.random.choice(len(dataset), min(num_images, len(dataset)), False)):
        img, lbl = dataset[idx]
        ax.imshow((img[0].numpy() * 255).astype(np.uint8).reshape(64,64), cmap='gray'), ax.axis('off')
        ax.set_title(f"Type: {list(dataset.type_mapping.keys())[img[1].tolist().index(1)]}\nLabel: {list(dataset.label_mapping.keys())[list(dataset.label_mapping.values()).index(lbl)]}")
    plt.show()

#Inspecting 5 images from the dataset
show_dataset_images(dataset)
# Importing Pytorch
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader

# Creating the class 'OCRModel' to define the model's structure
class OCRModel(nn.Module):
    def __init__(self):
        super(OCRModel, self).__init__()
        self.image_layer = nn.Sequential(
            nn.Conv2d(1, 16, kernel_size=3, padding=1),
            nn.MaxPool2d(kernel_size=2),
            nn.ReLU(),
            nn.Flatten(),
            nn.Linear(16*32*32, 128))
        
        self.type_layer = nn.Sequential(
            nn.Linear(5, 10),
            nn.ReLU())
        
        self.classifier = nn.Sequential(
            nn.Linear(128 + 10, 64), 
            nn.ReLU(),
            nn.Linear(64, 2))
        
    def forward(self, x_image, x_type):
        x_image = self.image_layer(x_image)
        x_type = self.type_layer(x_type)
        x = torch.cat((x_image, x_type), dim=1)  
        return self.classifier(x)

# Loading the data in batches
train_dataloader = DataLoader(dataset, batch_size=10, shuffle=True)

# Time to call the model, and it will answer lol
model = OCRModel()

# Defining the optimizer and loss function
optimizer = optim.Adam(model.parameters(), lr=0.001)
criterion = nn.CrossEntropyLoss()

# Training for ten epochs
for epoch in range(10):  
    for (images, types), labels in train_dataloader:
        optimizer.zero_grad()
        outputs = model(images, types)
        loss = criterion(outputs, labels) 
        loss.backward()
        optimizer.step()

    print(f"Epoch {epoch+1}, Loss: {loss.item()}")

After each epoch, we can see that the loss is decreasing, this implies that the model is learning.