app.py

import streamlit as st
import torch
from PIL import Image
import matplotlib.pyplot as plt
from safetensors.torch import load_model
from transformers import pipeline
import torch
from torch import nn
from torch.nn import functional as func_nn
from einops import rearrange
from huggingface_hub import PyTorchModelHubMixin
from torchvision import models


# main model network
class SiameseNetwork(nn.Module, PyTorchModelHubMixin):
    def __init__(self):
        super().__init__()

        # convolutional layer/block
        # self.convnet = MobileNet()
        self.convnet = models.mobilenet_v2(pretrained=True) # pretrained backbone   
        num_ftrs = self.convnet.classifier[1].in_features # get the first deimnesion of model head

        self.convnet.classifier[1] = nn.Linear(num_ftrs, 512) # change/switch backbone linear head  

        # fully connected layer for classification
        self.fc_linear = nn.Sequential(
            nn.Linear(512, 128),
            nn.ReLU(inplace=True), # actvation layer
            nn.Linear(128, 2)
        )

    def single_pass(self, x) -> torch.Tensor: 
        # sinlge Forward pass for each image
        x = rearrange(x, 'b h w c -> b c h w') # rearrange to (batch, channels, height, width)  to match model input
        output = self.convnet(x)
        output = self.fc_linear(output)

        return output

    def forward(self, input_1: torch.Tensor, input_2: torch.Tensor) -> torch.Tensor:

        # forward pass of first image
        output_1 = self.single_pass(input_1)

        # forward pass of second contrast image
        output_2 = self.single_pass(input_2)

        return output_1, output_2


# pretrained model file
model_file = 'best_signature_mobilenet.safetensors' #config.safetensor_file


# Function to compute similarity
def compute_similarity(output1, output2):
    return torch.nn.functional.cosine_similarity(output1, output2).item()

# Function to visualize feature heatmaps
def visualize_heatmap(model, image):
    model.eval()
    x = image.unsqueeze(0) # remove batch dimension
    features = model.convnet(x) # feature heatmap learnt by model
    heatmap = torch.mean(features, dim=1).squeeze().detach().numpy() # normalize heatmap to ndarray
    plt.imshow(heatmap, cmap="hot") # display heatmap as plot
    plt.axis("off")
    
    return plt

# Load the pre-trained model from safeetesor file
def load_pipeline():
    model_file = 'best_signature_mobilenet.safetensors' #config.safetensor_file
        # model_id = 'tensorkelechi/signature_mobilenet'
    model = SiameseNetwork() # model class/skeleton
    
    # model.load_state_dict(torch.load(model_file))
    model = load_model(model, model_file)
    # model = pipeline('image-classification', model=model_id, device='cpu') 
    model.eval()
    
    return model.to('cpu')

# Streamlit app UI template
st.title("Signature Forgery Detection")
st.write('Application to run/test signature forgery detecton model')

st.subheader('Compare signatures')
# File uploaders for the two images
original_image = st.file_uploader(
    "Upload the original signature", type=["png", "jpg", "jpeg"]
)
comparison_image = st.file_uploader(
    "Upload the signature to compare", type=["png", "jpg", "jpeg"]
)

def run_model_pipeline(model, original_image, comparison_image, threshold=0.5):
    if original_image is not None and comparison_image is not None: # ensure both images are uploaded

        # Preprocess images
        img1 = Image.open(original_image).convert("RGB") # load images from file paths to PIL Image
        img2 = Image.open(comparison_image).convert("RGB")

        # read/reshape and normalize as numpy array
        img1 = read_image(img1)
        img2 = read_image(img2)

        # convert to tensors and add batch dimensions to match model input shape
        img1_tensor = torch.unsqueeze(torch.as_tensor(img1), 0)
        img2_tensor = torch.unsqueeze(torch.as_tensor(img2), 0)

        # Get model embeddings/probabilites
        output1, output2 = model(img1_tensor, img2_tensor)
        st.success('outputs extracted')

        # Compute similarity
        similarity = compute_similarity(output1, output2)

        # Determine if it's a forgery based on determined threshold
        is_forgery = similarity < threshold

        # Display results
        st.subheader("Results")
        st.write(f"Similarity: {similarity:.2f}")
        st.write(f"Classification: {'Forgery' if is_forgery else 'Genuine'}")

        # Display images
        col1, col2 = st.columns(2) # GUI columns

        with col1:
            st.image(img1, caption="Original Signature", use_column_width=True)
        with col2:
            st.image(img2, caption="Comparison Signature", use_column_width=True)

        # Visualize heatmaps from extracted model features
        st.subheader("Feature Heatmaps")
        col3, col4 = st.columns(2)
        with col3:
            fig1 = visualize_heatmap(model, img1_tensor)
            st.pyplot(fig1)
        with col4:
            fig2 = visualize_heatmap(model, img2_tensor)
            st.pyplot(fig2)

    else:
        st.write("Please upload both the original and comparison signatures.")

# Run the model pipeline if a button is clicked
if st.button("Run Model Pipeline"):
    model = load_pipeline()
    
    # button click to process images
    if st.button("Process Images"):
        run_model_pipeline(model, original_image, comparison_image)