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 
# Import the necessary libraries
import matplotlib.pyplot as plt
import numpy as np
from project_utils import ProjectDataset
import pickle 
import torch
import torch.optim as optim
import torch.nn as nn
from torch.utils.data import DataLoader

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

# Define a function 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()

# Inspect 5 codes images from the dataset
show_dataset_images(dataset)
# Define DataLoader for the full dataset
data_loader = DataLoader(dataset, batch_size=32, shuffle=True)

# Define the OCR Model
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.ReLU(),
            nn.MaxPool2d(kernel_size=2),
            nn.Conv2d(16, 32, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2),
            nn.Flatten()
        )
        self.fc = nn.Sequential(
            nn.Linear(32 * 16 * 16, 128),
            nn.ReLU(),
            nn.Linear(128, 10)
        )

    def forward(self, x):
        # Ensure input is properly formatted
        if isinstance(x, list):
            x = torch.stack([item for item in x if isinstance(item, torch.Tensor)])  # Extract image tensors
        
        # Add batch dimension if necessary
        if len(x.shape) == 3:
            x = x.unsqueeze(0)
        
        # Ensure proper shape (batch_size, channels, height, width)
        if len(x.shape) == 4 and x.shape[1] != 1:
            x = x.permute(0, 3, 1, 2)
        
        x = self.image_layer(x)
        x = self.fc(x)
        return x

# Initialize the model, loss function, and optimizer
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = OCRModel().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)

# Training loop
epochs = 10
for epoch in range(epochs):
    model.train()
    running_loss = 0.0
    for batch in data_loader:
        images, labels = batch
        
        # Ensure images are in the correct format
        if isinstance(images, list):
            images = [img for img in images if isinstance(img, torch.Tensor) and img.dim() == 3]
            if len(images) == 0:
                continue  # Skip this batch if no valid images are found
            images = torch.stack(images)
        
        # Move to device and ensure proper shape
        images = images.to(device)
        if len(images.shape) == 3:
            images = images.unsqueeze(1)  # Add channel dimension
        labels = labels.to(device)
        
        optimizer.zero_grad()
        outputs = model(images)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()
        running_loss += loss.item()
    
    print(f"Epoch [{epoch+1}/{epochs}], Loss: {running_loss / len(data_loader):.4f}")

print("Training complete!")